Data-driven permanent magnet direct drive wind power plant equivalent modeling method

Abstract

The invention relates to a method for calculating equivalent parameters of a data-driven permanent magnet direct drive wind power plant, which belongs to the technical field of analog computation of a power system and comprises the following steps of: identifying a wind turbine generator group with similar dynamic behaviors by carrying out clustering analysis on a transient current time sequence of an outlet of a wind turbine generator; a neural network model is constructed, key electrical parameters such as operation condition parameters and voltage drop serve as input, a clustering result serves as output, and automatic classification and dynamic response feature extraction of the wind turbine generator are achieved by training a neural network; calculating unit equivalent parameters and current collection line parameters to obtain an equivalent model; and multi-operation scene simulation verification of the large-scale permanent magnet direct-driven wind power plant is passed. According to the method, equivalent modeling of the large-scale permanent-magnet direct-drive wind power plant is realized, and the modeling efficiency and the adaptability to complex operation scenes are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of power system simulation and calculation technology, and in particular, it is a data-driven method for calculating equivalent parameters of permanent magnet direct-drive wind farms. Background Technology

[0002] Wind power generation, with its advantages of high efficiency, cleanliness, renewable energy, and short infrastructure construction period, has ushered in unprecedented development opportunities. However, due to the relatively small capacity of a single wind turbine, permanent magnet direct-drive wind farms typically consist of dozens, hundreds, or even thousands of wind turbines. Their large scale and complex dynamic responses place a significant computational burden on electromagnetic transient simulations by requiring detailed modeling of each unit. To simplify wind farm models and improve simulation efficiency, equivalent modeling methods for wind farms have been widely researched and applied.

[0003] Existing equivalent methods for permanent magnet direct-drive wind farms include single-unit equivalent methods and multi-unit equivalent methods. Single-unit equivalent models typically employ a capacity-weighted approach, representing the entire wind farm as a single representative wind turbine, simplifying the modeling process and improving computational efficiency. However, this model has limited modeling accuracy and the ability to reflect local dynamics when dealing with wind farms with diverse control strategies and complex dynamic responses. Multi-unit equivalent models select clustering indices and methods to group wind turbines within large-scale wind farms, with each type of turbine represented by an equivalent unit. This reduces model size and computational burden while preserving key dynamic characteristics of the system. Common clustering indices include mechanical characteristic parameters such as wind speed and power curves, and electrical variables such as voltage and current. This equivalent method selects characteristic quantities that characterize the operating state of the wind turbines as clustering indices, grouping turbines with similar response characteristics into a single equivalent group for simplification. However, this type of mechanistic method requires re-evaluation based on the clustering indices under different operating scenarios, and there are difficulties in solving for the clustering indices and collector line parameters when analyzing anticipated faults. In addition, while using steady-state quantities before the fault as the basis for clustering makes the data easy to obtain, the error in cluster equivalence is relatively large; while using the time series of the fault transient process as the basis for clustering can achieve wind turbine clustering based on dynamic characteristics, there is a problem of difficulty in obtaining the data.

[0004] After selecting the clustering index for the simulation scenario, an appropriate clustering method needs to be adopted to divide the permanent magnet direct-drive wind farm cluster. Currently, the main clustering strategies include three types: clustering algorithm-driven, operation state characteristic-based, and multi-stage combined. The first two strategies divide the units based on mathematical models or operating conditions, while the multi-stage method integrates multi-dimensional information and improves the clustering accuracy through preliminary screening and subsequent optimization. This type of clustering method can be used for rapid transient simulation of wind farms under certain scenarios, but it is not universally applicable to different types of wind turbines, and the applicable scope of the model simulation scenario is usually relatively limited.

[0005] In summary, the adaptability of existing equivalent modeling methods to complex scenarios needs further improvement. For wind farms with specific system topologies and control and protection logic, some equivalent modeling methods face challenges in obtaining cluster partitioning indicators. When analyzing anticipated faults, there are difficulties in solving for clustering indicators and new energy equipment model parameters. Traditional dynamic time warping algorithms and widely used clustering methods rely on distance metrics and have limitations such as sensitivity to noise time series, affecting clustering results. Therefore, how to establish an efficient and accurate equivalent model for permanent magnet direct-drive wind farms to meet the needs of rapid electromagnetic transient simulation of large-scale wind farms has become a key technical problem that needs to be solved. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and propose a data-driven method for calculating equivalent parameters of permanent magnet direct-drive wind farms. This method can maintain the stable operation of the power grid, quickly and accurately provide the optimal load transfer scheme, greatly improve the work efficiency of dispatchers, and ensure the safe and stable operation of the power grid.

[0007] The technical problem solved by this invention is achieved through the following technical solution: A data-driven equivalent modeling method for permanent magnet direct-drive wind farms includes the following steps: Step 1: Collect actual measurement data and simulation data of a single unit grid-connected system to obtain the original transient current time series at the unit outlet; Step 2: Clean and normalize the original transient current time series at the unit outlet to form a standardized transient current time series, and calculate the correlation coefficient. Step 3: Based on the k-shape algorithm clustering, construct a multidimensional distance measure and iteratively update the centroid sequence according to the standardized multidimensional transient current time series in Step 2 to achieve machine grouping; Step 4: Construct a neural network to calculate the voltage drop of the units in the field through power flow and short circuit, while combining environmental parameters as input feature parameters and using the clustering results in Step 3 as output parameters. Step 5: Debug and determine the number of hidden layers and neurons, divide the dataset, and optimize the neural network using the backpropagation algorithm; Step 6: Calculate the equivalent wind turbine parameters and collector network parameters using an optimized neural network.

[0008] Furthermore, step 2 includes the following steps: Step 2.1: Clean the original transient current time series data by using linear interpolation to process missing values ​​and remove outliers; Step 2.2: Pad the end of the short sequence with zeros to obtain two standard sequences of length n. Shift these two sequences forward and backward, and calculate the correlation coefficient for the current shift. This forms a cross-correlation sequence of length w. ;

[0009] in, n To complete the length of the standard transient current time series, The length of the cross-correlation sequence, l The first time series l item, Standard time series The first in item, Standard time series The first in item, For the length of the sequence shift, Normalization yields , representing the time series of transient currents of wind turbine generators and The degree of similarity.

[0010] Furthermore, step 3 includes the following steps: Step 3.1: Construct a multidimensional distance measure for the standardized multidimensional transient current time series. And perform dimensionality reduction operations;

[0011] Where V is the multidimensional distance measure Dimensions For the first Distance measure for 3D time series To find the centroid sequence that minimizes the sum of squares of the distance metrics between the distance metrics of all transient current time series of the wind turbine, we use the reduced-dimensional distance metric. To optimize the objective, the distance measure between the centroid transient current sequence and other transient current time sequences is calculated iteratively.

[0012] in, For the k-th centroid sequence, For the k-th type of wind turbine cluster, the distance measure of each transient current time series is compared with the current centroid series. Based on the comparison results, the wind turbine cluster is divided to update the class members and obtain the dynamic behavior category label of the wind turbine.

[0013] Furthermore, step 4 includes the following steps: Step 4.1: Select the operating condition parameters of the wind turbine and the pre-calculated voltage drop parameters as input features to construct an input feature vector that is both accessible and universal. Step 4.2: Use a backpropagation neural network for modeling. The output of the neural network model is the dynamic behavior category label of the wind turbine obtained in step 3.

[0014] Moreover, the backpropagation neural network model contains two hidden layers. The neurons in the hidden layers use the Sigmoid activation function and have an output range of [0,1], which is used to enhance the smoothness and robustness of the classification results. The training process of a backpropagation neural network model is based on the forward propagation formula:

[0015]

[0016]

[0017]

[0018] In the formula, Indicates the first The weight matrix of the layer, Indicates the first Weighted sum of layers, Indicates the first Layer bias terms, This represents the activation function. Indicates the first The layer's activation output.

[0019] Furthermore, step 5 includes the following steps: Step 5.1: During the training sample generation process, the wind speed is set to be distributed in steps of 0.1 m / s within the range from the cut-in to the rated wind speed, and short-circuit faults of the same type and different severity are simulated at different nodes to obtain diverse transient response data. Step 5.2: Divide the clustering labels and the input feature data of wind speed and voltage drop into training set, validation set and test set according to the proportion, and use them for model training, overfitting control and performance evaluation, respectively. Step 5.3: Train the model using the backpropagation algorithm, with the mean squared error (MSE) as the loss function.

[0020] in, For the output results Dimension The actual value; During backpropagation, the weights and biases of each layer are updated according to the gradient descent rule. The gradient calculation formula is as follows:

[0021]

[0022] in, Let i be the loss function value of the i-th layer. The activation output of the i-th layer, and These are the gradients of the weights and the bias term, respectively. Adjust the values ​​of the repetition values ​​and bias terms of each neuron according to the gradient to reduce the error of the neural network surface model; Step 5.4: After each round of training, construct an equivalent model and compare it with the original unit simulation response. If the error exceeds the set threshold, it will be fed back to the training phase.

[0023] Furthermore, step 6 includes the following steps: Step 6.1: The capacity, active power, reactive power, etc. of the equivalent wind turbine are obtained by directly summing the corresponding quantities of each wind turbine. The internal parameters of the equivalent wind turbine are obtained by weighted summation based on the capacity of each turbine. The internal converter and its control parameters of the equivalent turbine are consistent with those of the original unit. Step 6.2: For the generator terminal transformers of the units in the station, they are equivalent to an expansion transformer located at the generator terminal of the equivalent unit. Its capacity is the sum of the capacities of the generator terminal transformers of each unit, and its impedance is the parallel equivalent of the impedances of the transformers of each unit.

[0024] in, The equivalent impedance of the collector network; , These are the terminal voltage and output current of fan k, respectively. ; This refers to the bus voltage at the fan port. is the ratio of the rated capacity of wind turbine k to the rated capacity of wind turbine, and N is the number of wind turbine units in the wind farm; Step 6.3: The equivalent parameters of the collection line are calculated based on the principle of power loss consistency, and connected to the equivalent unit outlet in the form of series impedance to realize the equivalent modeling of the collection network.

[0025] The advantages and positive effects of this invention are: This invention identifies wind turbine groups with similar dynamic behaviors by clustering the transient current time series at the wind turbine outlet; it constructs a neural network model, using key electrical parameters such as operating condition parameters and voltage drop as inputs and the clustering results as outputs, and trains the neural network to achieve automatic classification and dynamic response feature extraction of wind turbines; it then calculates equivalent parameters and collector line parameters to obtain an equivalent model; and finally verifies the model through simulations of large-scale permanent magnet direct-drive wind farms under multiple operating scenarios. This invention achieves equivalent modeling of large-scale permanent magnet direct-drive wind farms, significantly improving modeling efficiency and adaptability to complex operating scenarios. Attached Figure Description

[0026] Figure 1 A data-driven equivalence model framework diagram provided for embodiments of the present invention; Figure 2 This is a schematic diagram of a standard calculation example of a permanent magnet direct-drive wind power plant provided in an embodiment of the present invention; Figure 3 This is a diagram showing the wind speed distribution of the fans in scenarios one and two, provided in an embodiment of the present invention. Figure 4 A comparison diagram of phase A voltage of the busbar in scenario one provided by an embodiment of the present invention; Figure 5 This is a comparison diagram of phase A current of the busbar in scenario one provided by an embodiment of the present invention; Figure 6 This is a comparison diagram of active power response in scenario one provided by an embodiment of the present invention; Figure 7 This is a comparison diagram of reactive power response in scenario one provided by an embodiment of the present invention; Figure 8 This is a comparison diagram of active power response in scenario two provided by an embodiment of the present invention; Figure 9 This is a comparison diagram of reactive power response in scenario two provided by an embodiment of the present invention; Figure 10 This is a diagram showing the wind speed distribution of three fans in a scenario provided by an embodiment of the present invention; Figure 11 This is a comparison diagram of active power response in scenario three provided by an embodiment of the present invention; Figure 12 This is a comparison diagram of reactive power response in scenario three provided by an embodiment of the present invention. Detailed Implementation

[0027] The present invention will be further described in detail below with reference to the accompanying drawings.

[0028] A data-driven method for calculating equivalent parameters of permanent magnet direct-drive wind farms, such as... Figure 1 As shown, it includes the following steps: Step 1: Collect actual measurement data sets and simulation data sets of a single unit grid-connected system to obtain the original transient current time series at the unit outlet.

[0029] Step 2: Perform data cleaning and normalization on the original transient current time series at the unit outlet to form a standardized transient current time series, and calculate the correlation coefficient.

[0030] Step 2.1: Construct a grid-connected model of a single permanent magnet synchronous wind turbine and simulate transient processes such as single-phase, two-phase, and three-phase faults and wind speed fluctuations under different wind speeds. Obtain the time series dataset of the wind turbine outlet output current, clean the original current time series data, process missing values ​​through linear interpolation, remove outliers, and then perform alignment processing.

[0031] Step 2.2: For any two unit outlet current time series data, pad the end of the shorter series with zeros to make it equal in length to the longer series, resulting in two series of length . The current time series. Shift two series forward and backward, and calculate the correlation coefficient between the two series under the current shift condition. This leads to a cross-correlation sequence of length w. ,right Normalization yields , representing the time series of wind turbine current and Similarity:

[0032] in, n To complete the length of the standard transient current time series, The length of the cross-correlation sequence, l The first time series l item, Standard time series The first in item, Standard time series The first in item, For the length of the sequence shift, Normalization yields , representing the time series of transient currents of wind turbine generators and The degree of similarity.

[0033] Step 3: Based on the k-shape algorithm clustering, construct a multidimensional distance measure and iteratively update the centroid sequence according to the standardized multidimensional transient current time series in Step 2 to achieve machine grouping.

[0034] Step 3.1: Select the k-shape algorithm to mine the similarity of transient characteristics of each wind turbine in the cluster based on current time series indicators. For the standardized multidimensional transient current time series, construct a multidimensional distance measure. And perform dimensionality reduction operations;

[0035] Where V is the multidimensional distance measure Dimensions For the first Distance measure for 3D time series To find the centroid sequence that minimizes the sum of squares of the distance metrics between the distance metrics of all transient current time series of the wind turbine, we use the reduced-dimensional distance metric. To optimize the objective, the distance metric between the centroid transient current sequence and other transient current time sequences is calculated iteratively:

[0036] in, For the k-th centroid sequence, For the k-th type of wind turbine cluster, during the iteration process, the wind turbine clustering... In The optimal sequence shift length is determined based on the current highest centroid sequence identified in the previous step. After k single-step iterations, the distance metric between each current time series and the current centroid sequence is compared. Based on the comparison results, wind turbine clusters are defined, and the cluster members are updated. When the cluster members no longer change or the set maximum number of iterations is reached, the clustering results for wind turbine operating conditions and transient scenarios are obtained.

[0037] Step 4: Construct a neural network to calculate the voltage drop of the units in the field through power flow and short circuits, while combining environmental parameters as input feature parameters and using the clustering results from Step 3 as output parameters.

[0038] Step 4.1: Select the operating condition parameters of the wind turbine and the pre-calculated voltage drop parameters as input features to construct an input feature vector that is both accessible and universal.

[0039] Step 4.2: Use a backpropagation neural network for modeling. The output of the neural network model is the dynamic behavior category label of the wind turbine obtained in step 3.

[0040] The neural network model consists of an input layer, hidden layers, and an output layer. The number of nodes in the input layer is determined by the selected operating condition parameters and the dimensions of the transient response features, including wind speed and voltage drop depth. These parameters effectively reflect the electrical behavior of the wind farm under different operating states and disturbance conditions. The number of nodes in the output layer is consistent with the number of clusters of units based on time-current sequences, and is used to output the dynamic behavior category of each unit under the current input conditions.

[0041] The backpropagation neural network model contains two hidden layers, each with four neurons. The neurons in the hidden layers use the Sigmoid activation function, with an output range of [0,1], which is used to enhance the smoothness and robustness of the classification results.

[0042] The training process of a backpropagation neural network model is based on the forward propagation formula:

[0043]

[0044]

[0045]

[0046] in, Indicates the first The weight matrix of the layer, Indicates the first Weighted sum of layers, Indicates the first Layer bias terms, This represents the activation function. Indicates the first The layer's activation output.

[0047] Step 5: Debug and determine the number of hidden layers and neurons, divide the dataset, and optimize the neural network using the backpropagation algorithm.

[0048] Step 5.1: During the training sample generation process, the wind speed is set to be distributed in steps of 0.1 m / s within the range from the cut-in wind speed to the rated wind speed, and short-circuit faults of the same type and different severity are simulated at different nodes to obtain diverse transient response data.

[0049] Step 5.2: Divide the clustering labels and input feature data such as wind speed and voltage drop into training set, validation set and test set in a ratio of 7:2:1, and use them for model training, overfitting control and performance evaluation, respectively.

[0050] Step 5.3: Train the model using the backpropagation algorithm, with the mean squared error (MSE) as the loss function.

[0051] in, For the output results Dimension This is the actual value.

[0052] During training, the loss value on the validation set was continuously monitored. Training was terminated when the validation loss did not decrease significantly within five consecutive training epochs. To enhance the model's adaptability to small sample and low-dimensional data scenarios, an adaptive learning rate adjustment mechanism was adopted, dynamically adjusting the learning rate based on changes in training error to avoid model oscillations or convergence stalls during training. The initial learning rate was set to 0.01 and gradually adjusted through experimental debugging to ensure the model had a relatively fast convergence speed in the early stages of training, while maintaining small error fluctuations in the later stages.

[0053] During backpropagation, the weights and biases of each layer are updated according to the gradient descent rule. The gradient calculation formula is as follows:

[0054]

[0055] in, Let i be the loss function value of the i-th layer. The activation output of the i-th layer, and These are the gradients of the weights and the bias term, respectively.

[0056] Step 5.4: Regarding parameter optimization, the backpropagation process follows the standard gradient descent update rule. The repetition values ​​and bias terms of each neuron are adjusted according to the gradient to reduce the error of the BPNN. After each training round, an equivalent model is constructed and compared with the original unit simulation response. If the error exceeds a set threshold, it is fed back to the training phase.

[0057] Step 6: Calculate the equivalent wind turbine parameters and collector network parameters using an optimized neural network.

[0058] Step 6.1: The capacity, active power, and reactive power of the equivalent wind turbine are obtained by directly summing the corresponding quantities of each wind turbine. The internal parameters of the equivalent wind turbine are obtained by weighted summation based on the capacity of each turbine. The internal converter and its control parameters of the equivalent turbine are consistent with those of the original unit.

[0059] Step 6.2: For the generator terminal transformers within the power station, they are equivalent to an expansion transformer located at the equivalent generator terminal, with its capacity being the sum of the capacities of all generator terminal transformers and its impedance being the parallel equivalent of the impedances of all generator transformers.

[0060] in, The equivalent impedance of the collector network; , These are the terminal voltage and output current of fan k, respectively. ; This refers to the bus voltage at the fan port. is the ratio of the rated capacity of wind turbine k to the rated capacity of wind turbine, and N is the number of wind turbine units in the wind farm.

[0061] Step 6.3: The equivalent parameters of the collection line are calculated based on the principle of power loss consistency, and connected to the equivalent unit outlet in the form of series impedance to realize the equivalent modeling of the collection network.

[0062] Based on the aforementioned data-driven equivalent modeling method for permanent magnet direct-drive wind farms, an example is provided to illustrate... Figure 2 The present invention is demonstrated by modeling and simulating a permanent magnet direct-drive wind farm.

[0063] This embodiment establishes a standard case study of a permanent magnet direct-drive wind farm in the offline transient simulation software EMTP to verify the effectiveness of the equivalent method. The wind farm in the case study consists of 45 direct-drive permanent magnet synchronous wind turbines, a collection line, and a main transformer. Each turbine is connected to a 35kV busbar via a 0.575 / 35kV box-type transformer and a collection line, and then transmitted to the main transformer via a 35 / 120kV main transformer.

[0064] Figure 3 The examples show the wind speed distribution of wind turbines in scenarios one and two. In scenario one, the wind farm experiences a wind speed fluctuation of 2 m / s at 15 seconds, and the initial wind speed value is restored after 2 seconds. The simulation step size is 50 μs, and the simulation time is 30 seconds. Using the method of this invention, the turbines in the wind farm are divided into two groups to construct a dual-turbine equivalent model. The grouping results are shown in Table 1.

[0065] Table 1

[0066] Scenario 2: A phase-A ground fault occurs on transmission line #1 at the wind farm outlet at 20 seconds, causing the bus voltage to drop to 0.5 pu. The fault is cleared at 20.2 seconds. The simulation step size is 50 μs, and the simulation time is 30 seconds. Using the method of this invention, the turbine units within the wind farm are divided into three groups, and an equivalent model of the three turbines is constructed. The grouping results are shown in Table 2.

[0067] Table 2

[0068] Figure 4 and Figure 5This is a comparison diagram of the voltage and current waveforms of phase A of the wind farm collection bus in scenario one of the embodiments of the present invention. The simulation results of the dual-machine equivalent model obtained by the method of the present invention are compared with the simulation results of the single-machine equivalent model, the four-machine equivalent model and the detailed model.

[0069] Figure 6 and Figure 7 Table 3 shows a comparison of the active and reactive power output curves of the wind farm in Scenario 1 of this invention. The active and reactive power response errors and simulation time of the dual-machine equivalent model, single-machine equivalent model and four-machine equivalent model obtained by the method of this invention are shown in Table 3.

[0070] The maximum error represents the maximum difference between all results from each equivalent model and the detailed model. Due to the large scale of the examples, although the simulation speed of the single-machine equivalent model is fast, its accuracy is low and cannot meet the needs of transient simulation research. Compared with the four-machine equivalent model, the equivalent model of this invention has a higher speedup ratio while ensuring sufficient simulation accuracy to meet the needs of rapid simulation research. The method of this invention has a greater advantage when facing transient simulation of larger-scale wind power clusters.

[0071] Table 3

[0072] Figure 8 and Figure 9 Table 4 shows a comparison of the active and reactive power output curves of the wind farm in Scenario 2 of this invention. The active and reactive power response errors and simulation times of the three-machine equivalent model, single-machine equivalent model, and four-machine equivalent model obtained by the method proposed in this invention are shown in Table 4. Compared to the four-machine equivalent model, the three-machine equivalent model constructed by the method of this invention has a higher speedup ratio while maintaining accuracy, which can meet the needs of transient simulation research.

[0073] Table 4

[0074] Figure 10 The simulation scenario three describes the equivalent wind speeds of each wind turbine in a wind farm. In scenario three, a three-phase ground fault occurs on the wind farm's outlet transmission line #1 at 20 seconds, causing the bus voltage to drop to 0.5 pu. The fault is cleared at 20.2 seconds. The simulation step size is set to 50 μs, and the simulation time is 30 seconds. Using the method of this invention, the turbines in the wind farm are divided into three groups, and an equivalent model of the three turbines is constructed. The grouping results are shown in Figure 5.

[0075] Table 5

[0076] Figure 11 and Figure 12This is a comparison chart of the active and reactive power output curves of the wind farm in Scenario 3 of this invention. Figure 6 shows the active and reactive power response errors and simulation time of the three-machine equivalent model, single-machine equivalent model, and four-machine equivalent model constructed using the method of this invention in Scenario 3. Compared to the other two equivalent models, the equivalent model constructed using the method of this invention has significant advantages in both accuracy and speed, and can basically meet the requirements for rapid simulation.

[0077] Table 6

[0078] To further verify the applicability of the equivalent method of this invention in large-scale wind farms, based on the same control strategy and topology of the wind farms in the above embodiments, a simulation model containing 100 wind turbines was constructed. Its environmental parameters and voltage drop values ​​under different transient scenarios were used as input parameters for a neural network to automatically identify the wind turbine categories, achieving equivalent representation of large-scale wind farms. The clustering results are shown in Table 7, indicating that the method of this invention can still maintain good clustering performance in larger-scale unit scenarios, verifying the robustness and scalability of the equivalent method when handling wind farms of different sizes.

[0079] Table 7

[0080] It should be emphasized that the embodiments described in this invention are illustrative rather than limiting. Therefore, this invention includes, but is not limited to, the embodiments described in the specific implementation. Any other implementations derived by those skilled in the art based on the technical solutions of this invention are also within the scope of protection of this invention.

Claims

1. A data-driven method for calculating equivalent parameters of a permanent magnet direct-drive wind farm, characterized in that: Includes the following steps: Step 1: Collect actual measurement data and simulation data of a single unit grid-connected system to obtain the original transient current time series at the unit outlet; Step 2: Clean and normalize the original transient current time series at the unit outlet to form a standardized transient current time series, and calculate the correlation coefficient. Step 3: Based on the k-shape algorithm clustering, construct a multidimensional distance measure and iteratively update the centroid sequence according to the standardized multidimensional transient current time series in Step 2 to achieve machine grouping; Step 4: Construct a neural network to calculate the voltage drop of the units in the field through power flow and short circuit, while combining environmental parameters as input feature parameters and using the clustering results in Step 3 as output parameters. Step 5: Debug and determine the number of hidden layers and neurons, divide the dataset, and optimize the neural network using the backpropagation algorithm; Step 6: Calculate the equivalent wind turbine parameters and collector network parameters using an optimized neural network.

2. The data-driven equivalent modeling method for permanent magnet direct-drive wind farms according to claim 1, characterized in that: Step 2 includes the following steps: Step 2.1: Clean the original transient current time series data by using linear interpolation to process missing values ​​and remove outliers; Step 2.2: Pad the end of the short sequence with zeros to obtain two standard sequences of length n. Shift these two sequences forward and backward, and calculate the correlation coefficient for the current shift. This forms a cross-correlation sequence of length w. ; ； ； in, n To complete the length of the standard transient current time series, The length of the cross-correlation sequence, l The first time series l item, Standard time series The first in item, Standard time series The first in item, For the length of the sequence shift, Normalization yields , representing the time series of transient currents of wind turbine generators and The degree of similarity.

3. The data-driven equivalent modeling method for permanent magnet direct-drive wind farms according to claim 1, characterized in that: Step 3 includes the following steps: Step 3.1: Construct a multidimensional distance measure for the standardized multidimensional transient current time series. The similarity of time series is mapped to a distance index, and then dimensionality reduction is performed. ； ； Where V is the multidimensional distance measure Dimensions For the first Distance measure for 3D time series To find the centroid sequence that minimizes the sum of squares of the distance metrics between the distance metrics of all transient current time series of the wind turbine, we use the reduced-dimensional distance metric. To optimize the objective, the distance measure between the centroid transient current sequence and other transient current time sequences is calculated iteratively. ； in, For the k-th centroid sequence, For the k-th type of wind turbine cluster, the distance measure of each transient current time series is compared with the current centroid series. Based on the comparison results, the wind turbine cluster is divided to update the class members and obtain the dynamic behavior category label of the wind turbine.

4. The data-driven equivalent modeling method for permanent magnet direct-drive wind farms according to claim 1, characterized in that: Step 4 includes the following steps: Step 4.1: Select the operating condition parameters of the wind turbine and the pre-calculated voltage drop parameters as input features to construct an input feature vector that is both accessible and universal. Step 4.2: Use a backpropagation neural network for modeling. The output of the neural network model is the dynamic behavior category label of the wind turbine obtained in step 3.

5. The data-driven equivalent modeling method for permanent magnet direct-drive wind farms according to claim 4, characterized in that: The backpropagation neural network model contains two hidden layers. The neurons in the hidden layers use the Sigmoid activation function and have an output range of [0,1], which is used to enhance the smoothness and robustness of the classification results. The training process of a backpropagation neural network model is based on the forward propagation formula: ； ； ； ； In the formula, Indicates the first The weight matrix of the layer, Indicates the first Weighted sum of layers, Indicates the first Layer bias terms, This represents the activation function. Indicates the first The layer's activation output.

6. The data-driven equivalent modeling method for permanent magnet direct-drive wind farms according to claim 1, characterized in that: Step 5 includes the following steps: Step 5.1: During the training sample generation process, the wind speed is set to be distributed in steps of 0.1 m / s within the range from the cut-in to the rated wind speed, and short-circuit faults of the same type and different severity are simulated at different nodes to obtain diverse transient response data. Step 5.2: Divide the clustering labels and the input feature data of wind speed and voltage drop into training set, validation set and test set according to the proportion, and use them for model training, overfitting control and performance evaluation, respectively. Step 5.3: Train the model using the backpropagation algorithm, with the mean squared error (MSE) as the loss function. ； in, For the output results Dimension The actual value; During backpropagation, the weights and biases of each layer are updated according to the gradient descent rule. The gradient calculation formula is as follows: ； ； in, Let i be the loss function value of the i-th layer. The activation output of the i-th layer, and These are the gradients of the weights and the bias term, respectively. Adjust the values ​​of the repetition values ​​and bias terms of each neuron according to the gradient to reduce the error of the neural network surface model; Step 5.4: After each round of training, construct an equivalent model and compare it with the original unit simulation response. If the error exceeds the set threshold, it will be fed back to the training phase.

7. The data-driven equivalent modeling method for permanent magnet direct-drive wind farms according to claim 1, characterized in that: Step 6 includes the following steps: Step 6.1: The capacity, active power, reactive power, etc. of the equivalent wind turbine are obtained by directly summing the corresponding quantities of each wind turbine. The internal parameters of the equivalent wind turbine are obtained by weighted summation based on the capacity of each turbine. The internal converter and its control parameters of the equivalent turbine are consistent with those of the original unit. Step 6.2: For the generator terminal transformers of the units in the station, they are equivalent to an expansion transformer located at the generator terminal of the equivalent unit. Its capacity is the sum of the capacities of the generator terminal transformers of each unit, and its impedance is the parallel equivalent of the impedances of the transformers of each unit. ； in, The equivalent impedance of the collector network; , These are the terminal voltage and output current of fan k, respectively. ; This refers to the bus voltage at the fan port. is the ratio of the rated capacity of wind turbine k to the rated capacity of wind turbine, and N is the number of wind turbine units in the wind farm; Step 6.3: The equivalent parameters of the collection line are calculated based on the principle of power loss consistency, and connected to the equivalent unit outlet in the form of series impedance to realize the equivalent modeling of the collection network.

Images