A method and system for dynamic equivalent modeling of wind farms based on GRU-LSTM-FC combined network

Abstract

This invention discloses a method and system for dynamic equivalent modeling of wind farms based on a GRU-LSTM-FC combined network. First, a hybrid wind farm containing both doubly-fed induction generators (DFIGs) and direct-drive turbines is constructed. Second, the output power of the wind farm is simulated under different wind speeds, wind directions, and voltage dips at the grid connection point, establishing a historical dataset, which is then divided into a training set and a test set. Third, the GRU-LSTM-FC combined network is trained using the training set data to obtain the combined network model parameters. Next, a genetic algorithm is used to optimize the number of fully connected (FC) layers and the number of neurons in each layer of the combined network, obtaining optimized combined network model parameters. Finally, the optimized combined network is used to predict the test set data, obtaining the predicted dynamic response of the wind farm. This invention uses a data-driven approach for dynamic wind farm modeling, demonstrating good adaptability to wind farms with different types of turbines or complex terrain, thus improving modeling accuracy.

Description

Technical Field

[0001] This invention belongs to the field of power system modeling, and in particular relates to a method and system for dynamic equivalent modeling of wind farms based on GRU-LSTM-FC combined network. Background Technology

[0002] Dynamic equivalent modeling of renewable energy power plants is of great significance for ensuring the safe and stable operation of the power system. Equivalent modeling methods include three categories: mechanistic modeling, non-mechanistic modeling, and data-driven modeling. Currently, most research focuses on mechanistic and non-mechanistic modeling; however, both of these methods depend on operating conditions and specific disturbances, making it difficult to obtain universally applicable equivalent models for wind farms.

[0003] In recent years, with the continuous advancement of big data technology, data-driven modeling methods have received increasing attention. This method not only utilizes response data under a specific perturbation but also fully mines the characteristic information of historical data, exhibiting better adaptability compared to non-mechanistic modeling methods based on transfer functions. Currently, data-driven modeling methods mainly focus on the field of deep learning, with the core being the determination of a suitable deep learning network for dynamic modeling of wind farms. However, it also faces challenges such as network structure adjustment and hyperparameter optimization. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to provide a highly adaptable method and system for dynamic equivalent modeling of wind farms based on a gated cyclic unit-long short-term memory-fully connected (GRU-LSTM-FC) combined network.

[0005] Technical solution: The wind farm dynamic equivalent modeling method based on GRU-LSTM-FC combined network of the present invention obtains the dynamic response time series of active power and reactive power of the target wind farm by performing the following steps;

[0006] Step 1: Construct a target wind farm that includes both doubly fed induction generator (DFIG) and direct-drive wind turbine (PMSG);

[0007] Step 2: Based on the wind speed and direction data measured by the meteorological tower, calculate the wind speed data of each wind turbine in the target wind farm. Based on the simulation of the disturbance, obtain the voltage time series of the grid connection point of the target wind farm. Use the wind speed data, wind direction data of each wind turbine and the voltage time series of the grid connection point as input, and the dynamic response time series of the active power and reactive power of the target wind farm as output to construct the training set and the test set.

[0008] Step 3: Train the GRU-LSTM-FC combined network using the training set data to obtain the GRU-LSTM-FC combined network model parameters. Based on the GRU-LSTM-FC combined network model parameters, use a genetic algorithm to optimize the number of FC layers and the number of neurons in each layer of the GRU-LSTM-FC combined network to obtain the optimized GRU-LSTM-FC combined network model.

[0009] Step 4: Input the wind speed data, wind direction data of each wind turbine in the test set and the voltage time series of the grid connection point into the optimized GRU-LSTM-FC combined network model, and output the dynamic response time series of the active power and reactive power of the target wind farm.

[0010] Furthermore, in step 1, the target wind farm is connected to the CEPRI-36 node system after being boosted by two stages.

[0011] Furthermore, step 2 specifically includes the following steps:

[0012] (a) Steady-state power calculation of wind turbines considering wake effect

[0013] Based on the wind speed and direction measured by the meteorological tower, the input wind speed of each turbine in the wind farm, taking into account the wake effect, is calculated as follows:

[0014]

[0015] In the formula, v0 is the wind speed measured by the anemometer tower; v d d represents the wind speed at a distance d from the meteorological tower, affected only by the wake of the wind turbine at the meteorological tower; r is the radius of the wind turbine. C is the wake descent constant; T v is the thrust coefficient; i v represents the actual wind speed of the wind turbine. i0 To disregard the wake velocity acting on the i-th fan; v j-i To account for the wind speed when the j-th fan acts on the i-th fan during wake conditions; n is the total number of fans, and the fan terminal wind speed is determined based on the wind direction v at the anemometer tower. wd It changes with the changes;

[0016] Based on the terminal wind speed of each wind turbine, the steady-state power P0 of each wind turbine is calculated as follows:

[0017]

[0018] In the formula, v is the wind speed at the turbine end; v in and v out These are the cut-in wind speed and the cut-out wind speed, respectively; v N Rated wind speed; P N This refers to the rated power of the wind turbine generator set;

[0019] (b) Simulation setup of LVRT process in wind farm

[0020] At different wind speeds and wind directions, a three-phase short-circuit fault occurs at the grid connection point of the wind farm at time t0, and the fault is cleared at time t1. By adjusting the grounding impedance, different voltage drop depths are obtained, and the active power and reactive power time series of the wind farm are obtained through simulation.

[0021] (c) Construction of training and test sets

[0022] The input data includes wind speed and direction data measured by the meteorological tower, as well as the voltage time series at the wind farm's grid connection point; the output data includes the active and reactive power time series of the wind farm.

[0023] The training set data includes: wind speeds within the range of 3 m / s cut-in wind speed and 12 m / s rated wind speed, with a wind speed interval of 1 m / s; wind direction [0, 180°] (since the wake effect caused by wind direction changes within [180°, 360°] is consistent with that within [0, 180°], only the wind direction within [0, 180°] is considered), with a direction interval of 10°. The voltage drop range is [0.1, 0.9] U. n The drop interval is 0.1U. n The total sampling time is 1 second, and the sampling interval is 0.001 seconds. Therefore, there are 1710 sets of sample data (10×19×9) as the training set, and the length of each data point is 1000.

[0024] The test set data includes three types of data: wind speed, wind direction, and voltage drop at the PCC point. All data are presented at smaller scales of variation (i.e., 0.1 m / s, 1°, and 0.01 U). n The data were randomly combined into 100 groups to form the test data.

[0025] Furthermore, step 3 specifically includes the following steps:

[0026] (a) Construction of GRU-LSTM-FC combined network

[0027] The constructed combined network consists of three layers: the first layer is a GRU layer, the second layer is an LSTM layer, and the third layer is a FC layer. Data passes through the GRU layer, LSTM layer, and FC layer in sequence.

[0028] (b) Performance evaluation metrics for combined networks

[0029] The mean squared error (MSE) metric is used to evaluate the degree of matching between the output and the actual response of the GRU-LSTM-FC combined network. The absolute error at each time point is used as another metric to describe the instantaneous error over the entire transient interval, as detailed below:

[0030]

[0031]

[0032] In the formula, y i Let i be the actual response value of the wind farm at time i; The output of the GRU-LSTM-FC combined network at time i; N = 1000, where N is the total number of data samples taken during the entire time-domain simulation.

[0033] (c) Determining the number of layers in the combined network

[0034] The performance of five networks—single-layer LSTM, single-layer GRU, two-layer LSTM, two-layer GRU, and single-layer GRU-single-layer LSTM—is evaluated based on iterative efficiency and the mean squared error of the training and test sets. In all five cases, the number of fully connected layers (FC) is uniformly set to 1, and the total number of hidden layer neurons is set to 50. The iteration termination error for training the combined network is set to 10. -2 The maximum number of iterations is set to 30,000, and the model learning rate is I. r It can be expressed as an inverse proportional function with the number of iterations (epochs) as the independent variable, specifically:

[0035]

[0036] In the formula, I r0 =10 -2 This is the initial learning rate of the model; epoch is the number of iterations during the model training process.

[0037] Based on the mean squared errors of the five networks on the training and test sets, it was found that the error was minimized when GRU had 1 layer and LSTM had 1 layer. Therefore, it was determined that GRU had 1 layer and LSTM had 1 layer.

[0038] Furthermore, step 4 specifically includes the following steps:

[0039] (a) Hyperparameters to be optimized in the combined network

[0040] The hyperparameters to be optimized include the number of FC layers, and the number of neurons in GRU, LSTM, and FC.

[0041] (b) Hyperparameter optimization based on genetic algorithm

[0042] Population chromosome encoding is performed, and the chromosome C structure of individuals in the population is constructed based on the number of FC layers and the number of neurons in each hidden layer as follows:

[0043] C = [L] FC N GRU N LSTM NFC1 .,...,N FCn ]

[0044] In the formula, L FC N represents the number of layers in the FC (Full-Functional) layer. GRU N represents the number of neurons in the GRU layer. LSTM N represents the number of neurons in the LSTM layer. FC1 ,…,N FCn This represents the number of neurons in each layer of the FC layer.

[0045] The fitness function is designed to comprehensively evaluate the performance of the combined network based on three metrics: accuracy, generalization, and training efficiency. The fitness function J is designed as the sum of these three metrics, and its expression is as follows:

[0046]

[0047] In the formula, J is the fitness function; Loss train0 and Loss train The ratio of the mean squared error of the training set to the total mean squared error is used to characterize the accuracy; Loss test0 and Loss test T0 represents the mean squared error of the test set and the total mean squared error, respectively. The ratio of these two values ​​is used to characterize the model's generalization ability. T0 represents the training duration, which is the average duration of multiple training iterations. start and T end These are the start and end times of training, in minutes, and the ratio T0 / (T) is... end -T start The training efficiency of the network is represented by α, β, and γ, which are the weight coefficients of each index. Based on the iteration termination condition, the optimized hyperparameter results are obtained.

[0048] This invention also discloses a dynamic equivalent modeling system for wind farms based on a GRU-LSTM-FC combined network, comprising:

[0049] Modeling Unit: Based on the PSASP platform, construct a target wind farm containing n wind turbine units, and connect the target wind farm to the CEPRI-36 node system after two-stage voltage boosting;

[0050] Simulation Unit: Based on wind speed data, wind direction data, and voltage drop disturbance at the grid connection point, the target wind farm is simulated to obtain the output power of the wind farm, and training and testing sets are constructed.

[0051] Optimization Unit: The GRU-LSTM-FC combined network is trained using the training set data. The number of GRU layers and LSTM layers in the GRU-LSTM-FC combined network is determined by performance metrics. The number of FC layers and the number of neurons in each layer of the combined network are optimized using a genetic algorithm to obtain the optimized GRU-LSTM-FC combined network model parameters.

[0052] Prediction Unit: Based on the input data of a certain wind speed, wind direction and voltage drop disturbance in the test set, the optimized GRU-LSTM-FC combined network is used to make predictions and obtain the dynamic response time series of active and reactive power of the wind farm.

[0053] Furthermore, in the modeling unit, the target wind farm is boosted in two stages and then connected to the CEPRI-36 node system via the BUS30 bus.

[0054] Beneficial Effects: Compared with existing technologies, this invention has the following significant advantages: The data-driven modeling method based on the GRU-LSTM-FC combined network proposed in this invention can mine and learn the internal features of a large amount of historical data. The richer the historical data, the higher the modeling accuracy. At the same time, the model has strong universality and good adaptability to wind farms with different types of turbines or wind farms with complex terrain. Attached Figure Description

[0055] Figure 1 This is a flowchart illustrating an embodiment of the present invention;

[0056] Figure 2 This is a diagram of the wind farm topology.

[0057] Figure 3 For the CEPRI-36 node system;

[0058] Figure 4 The mean squared error of the five network structures during the training process;

[0059] Figure 5 The fitness percentage of different FC layer numbers in each generation of the genetic population;

[0060] Figure 6 Mean squared error for modeling test samples before and after GA optimization;

[0061] Figure 7 The mean square error of the four networks on the test sample data is given. Detailed Implementation

[0062] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0063] Example 1:

[0064] The implementation process is attached. Figure 1 .

[0065] Step 1: Build a hybrid wind farm with both doubly fed induction generators (DFIG) and direct-drive wind turbines (PMSG) based on the PSASP platform.

[0066] The structure of the hybrid wind farm is attached. Figure 2 As shown, the wind farm consists of 18 doubly-fed wind turbines with a rated power of 1.5MW and 18 direct-drive permanent magnet wind turbines with a rated power of 1.5MW, with each turbine spaced 500m apart. The wind farm is connected to the BUS30 bus of the CEPRI-36 node system via a two-stage boost-up system of terminal transformers and main transformers.

[0067] The CEPRI-36 node system architecture is attached. Figure 3 Detailed parameters of the wind turbine and collector lines are shown in Table 1.

[0068] Table 1 Wind Farm Model Parameters

[0069]

[0070] Step 2: Calculate the wind speed of each wind turbine in the target wind farm based on the wind speed and direction measured by the wind measurement tower. Simulate the voltage, active power and reactive power time series of the grid connection point of the target wind farm based on the disturbance. Obtain a historical dataset consisting of {wind speed, wind direction, voltage time series, active power time series and reactive power time series}. Divide the historical dataset into training set and test set.

[0071] (a) Steady-state power calculation of wind turbines considering wake effect

[0072] Based on the wind speed and direction measured by the meteorological tower, the input wind speed of each turbine in the wind farm, taking into account the wake effect, is calculated as follows:

[0073]

[0074] In the formula, v0 is the wind speed measured by the anemometer tower; v d d represents the wind speed at a distance d from the meteorological tower, affected only by the wake of the wind turbine at the meteorological tower; r is the radius of the wind turbine. C is the wake descent constant; T v is the thrust coefficient; i v represents the actual wind speed of the wind turbine. i0 To disregard the wake velocity acting on the i-th fan; v j-i To account for the wind speed when the j-th fan acts on the i-th fan during the wake vortex; n is the total number of fans. Simultaneously, the wind direction v at the anemometer tower... wd The wind speeds at the turbine terminals are different.

[0075] Based on the terminal wind speed of each wind turbine, the steady-state power P0 of each wind turbine is calculated as follows:

[0076]

[0077] In the formula, v is the wind speed at the turbine end; v in and v out These are the cut-in wind speed and the cut-out wind speed, respectively; v N Rated wind speed; P N This refers to the rated power of the wind turbine.

[0078] (b) Simulation setup of LVRT process in wind farm

[0079] At different wind speeds and wind directions, a three-phase short-circuit fault occurs at the grid connection point of the wind farm at t=0.1s, and the fault is cleared at 0.25s. By adjusting the grounding impedance, different voltage drop depths are obtained, and the active power and reactive power trajectories of the wind farm are obtained through simulation.

[0080] (c) Construction of training and test sets

[0081] The input data includes wind speed and direction data measured by the meteorological tower, as well as the voltage time series at the wind farm's grid connection point; the output data includes the active and reactive power time series of the wind farm.

[0082] The training set data includes: wind speeds within the range of 3 m / s cut-in wind speed and 12 m / s rated wind speed, with a wind speed interval of 1 m / s; wind direction [0, 180°] (since the wake effect caused by wind direction changes within [180°, 360°] is consistent with that within [0, 180°], only the wind direction within [0, 180°] is considered), with a direction interval of 10°. The voltage drop range is [0.1, 0.9] U. n The drop interval is 0.1U. n The total sampling time is 1 second, and the sampling interval is 0.001 seconds. Therefore, there are 1710 sets of sample data (10×19×9) as the training set, and the length of each data point is 1000.

[0083] The test set data includes three types of data: wind speed, wind direction, and voltage drop at the PCC point. All data are presented at smaller scales of variation (i.e., 0.1 m / s, 1°, and 0.01 U). n The data were randomly combined into 100 groups to form the test data.

[0084] Step 3: Train the GRU-LSTM-FC combined network using the training set data to obtain the GRU-LSTM-FC combined network model parameters;

[0085] (a) Construction of GRU-LSTM-FC combined network

[0086] The constructed combined network consists of three layers: the first layer is a GRU layer, the second layer is an LSTM layer, and the third layer is an FC layer.

[0087] (b) Performance evaluation metrics for combined networks

[0088] The mean squared error (MSE) metric is used to evaluate the degree of matching between the output and the actual response of the GRU-LSTM-FC combined network. Simultaneously, the absolute error at each time point is used as another metric to describe the instantaneous error over the entire transient interval, as follows:

[0089]

[0090]

[0091] In the formula, y i Let i be the actual response value of the wind farm at time i; The output of the GRU-LSTM-FC combined network at time i; N = 1000, where N is the total number of data samples taken during the entire time-domain simulation.

[0092] (c) Determining the number of layers in the combined network

[0093] To analyze the impact of the number of GRU and LSTM layers in the combined network on the performance of the wind farm equivalent model, we evaluate the performance of five networks—single-layer LSTM, single-layer GRU, two-layer LSTM, two-layer GRU, and single-layer GRU-single-layer LSTM—from the perspectives of iteration efficiency and mean squared error of the training and test sets. In all five cases, the number of fully connected (FC) layers is uniformly set to 1, and the total number of hidden layer neurons is set to 50. The iteration termination error for training the combined network is set to 10. -2 The maximum number of iterations is set to 30,000, and the model learning rate is I. r It can be expressed as an inverse proportional function with the number of iterations (epochs) as the independent variable, specifically:

[0094]

[0095] In the formula, I r0 =10 -2 It is the initial learning rate of the model; epoch is the number of iterations during the model training process.

[0096] The mean squared errors of the five networks during training are shown in the appendix. Figure 4 The mean squared errors on the training and test sets are shown in Table 2. Figure 4 As shown in Table 2, the error is minimized when both the GRU and LSTM layers are 1. Therefore, we determine that both the GRU and LSTM layers should be 1.

[0097] Table 2 shows the mean squared errors of the five network architectures on the training and test sets.

[0098]

[0099] Step 4: Use a genetic algorithm (GA) to optimize the number of FC layers and the number of neurons in each layer of the combined network to obtain the optimized GRU-LSTM-FC combined network model parameters.

[0100] (a) Hyperparameters to be optimized in the combined network

[0101] The hyperparameters to be optimized include the number of FC layers, and the number of neurons in GRU, LSTM, and FC.

[0102] (b) Hyperparameter optimization based on GA algorithm

[0103] First, population chromosome encoding is performed. Based on the number of FC layers and the number of neurons in each hidden layer, the chromosome C structure of individuals in the population is constructed as follows:

[0104] C = [L] FC N GRU N LSTM N FC1 .,...,N FCn ]

[0105] In the formula, L FC N represents the number of layers in the FC (Full-Functional) layer. GRU N represents the number of neurons in the GRU layer. LSTM N represents the number of neurons in the LSTM layer. FC1 ,…,N FCn This represents the number of neurons in each layer of the FC layer.

[0106] Then, the fitness function is designed. Based on the comprehensive evaluation of the combined network's performance using three metrics—accuracy, generalization, and training efficiency—the fitness function J is designed as the sum of these three metrics, and its expression is:

[0107]

[0108] In the formula, J is the fitness function; Loss train0 and Loss train The ratio of the mean squared error of the training set to the total mean squared error is used to characterize the accuracy; Loss test0 and Loss test T0 represents the mean squared error of the test set and the total mean squared error, respectively. The ratio of these two values ​​is used to characterize the model's generalization ability. T0 represents the training duration, which is the average duration of multiple training iterations. start and T end These are the start and end times of training, in minutes, and the ratio T0 / (T) is... end-T start The training efficiency of the network is represented by α, β, and γ, which are the weight coefficients of each indicator.

[0109] The fitness percentage of GA during the iteration process and the number of iterations are shown in the appendix for different numbers of fully connected (FC) layers (including the number of neurons in each layer under the corresponding FC layer). Figure 5 In the example, the GA algorithm converged after the 10th iteration. It can be seen that during the iteration process, the fitness ratio is the highest when the number of FC layers is 6, reaching as high as 88.684% after the 10th iteration. Therefore, the FC layer is determined to be 6 layers. At this time, the optimal chromosome value is C = [6,71,66,48,79,45,26,77,49], that is, the FC layer is 6 layers. The number of neurons in GRU and LSTM are 71 and 66, respectively, and the number of neurons in each layer of FC are 48, 79, 45, 26, 77 and 49, respectively.

[0110] Step 5: Use the optimized GRU-LSTM-FC combined network to predict the test set data and obtain the wind farm dynamic response prediction results.

[0111] Based on the optimized GRU-LSTM-FC combined network described above, a dynamic equivalent model of the wind farm was performed. The mean square error of the output power of the GRU-LSTM-FC combined network under 100 test set samples is shown in the appendix. Figure 6 The figure also shows the mean square error of the output power before optimization of the combined network. It can be seen that the GRU-LSTM-FC combined network optimized based on GA can significantly improve the accuracy of equivalent modeling of wind farms.

[0112] To verify the equivalent effect of the present invention based on the GRU-LSTM-FC combined network, the above examples and training data are used as examples, employing BP neural network, RNN network, and nonlinear autoregressive neural network (NARX) with external input for modeling. The mean square errors of active and reactive power for the four models in 100 test samples are shown in the appendix. Figure 7 and Table 3. (Appendix) Figure 7 Both Table 3 and Table 4 show that the combined model of the present invention not only has high equivalence accuracy, but also the training time is within an acceptable range.

[0113] Table 3 Comparison of modeling performance of the four models

[0114] Network Structure MSE(P) MSE(Q) Average training time / s BP 4.00215 0.04439 91 RNN 1.04354 0.41124 257 NARX 0.51146 0.63713 289 This article's model 0.01941 0.00889 173

[0115] Example 2:

[0116] This invention provides a dynamic equivalent modeling system for wind farms based on a GRU-LSTM-FC combined network, comprising:

[0117] Modeling Unit: Based on the PSASP platform, a wind farm containing 36 wind turbine units is constructed, and the wind farm is connected to the CEPRI-36 node system through the BUS30 bus after two-stage voltage boost.

[0118] Simulation Unit: Based on wind speed, wind direction, and voltage dip disturbances at the grid connection point, the simulation unit obtains the output power time series of the target wind farm, generating a historical dataset consisting of {wind speed, wind direction, voltage time series, active power time series, and reactive power time series}. This historical dataset is then divided into a training set and a test set.

[0119] Optimization Unit: The GRU-LSTM-FC combined network is trained using training set data. The number of GRU layers and LSTM layers in the GRU-LSTM-FC combined network is determined by performance metrics. Furthermore, the genetic algorithm (GA) is used to optimize the number of FC layers and the number of neurons in each layer of the combined network to obtain the optimized GRU-LSTM-FC combined network model parameters.

[0120] Prediction Unit: Based on the input data of a certain wind speed, wind direction and voltage drop disturbance in the test set, the optimized GRU-LSTM-FC combined network is used to make predictions and obtain the dynamic response time series of active and reactive power of the wind farm.

[0121] For specific implementation schemes of each module of the device of the present invention, please refer to the specific implementation process of the above method.

[0122] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0123] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0124] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0125] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0126] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for dynamic equivalent modeling of wind farms based on a GRU-LSTM-FC combined network, characterized in that, For a target wind farm, the following steps are performed to obtain the dynamic response time series of the active and reactive power of the target wind farm; Step 1: Construct a target wind farm that includes both doubly fed induction generator (DFIG) and direct-drive wind turbine (PMSG); Step 2: Based on the wind speed and direction data measured by the meteorological tower, calculate the wind speed data of each wind turbine in the target wind farm. Based on the simulation of the disturbance, obtain the voltage time series of the grid connection point of the target wind farm. Use the wind speed data, wind direction data of each wind turbine and the voltage time series of the grid connection point as input, and the dynamic response time series of the active power and reactive power of the target wind farm as output to construct the training set and the test set. Step 2 specifically includes the following steps: (a) Steady-state power calculation of wind turbines considering wake effect Based on the wind speed and direction measured by the meteorological tower, the input wind speed of each turbine in the wind farm, taking into account the wake effect, is calculated as follows: In the formula, v0 is the wind speed measured by the anemometer tower; v d d represents the wind speed at a distance d from the meteorological tower, affected only by the wake of the wind turbine at the meteorological tower; r is the radius of the wind turbine. C is the wake descent constant; T v is the thrust coefficient; i v represents the actual wind speed of the wind turbine. i0 To disregard the wake velocity acting on the i-th fan; v j-i To consider the wind speed when the j-th fan acts on the i-th fan during wake generation; n is the total number of units, n = n D +n P n D n represents the number of DFIG (Doubly Injected Feeder) wind turbines. P The number of PMSGs for direct-drive wind turbines is given, and the turbine terminal wind speed is determined based on the wind direction v at the meteorological tower. wd The steady-state power P0 of each wind turbine is calculated based on the wind speed at the turbine terminals, as follows: In the formula, v is the wind speed at the turbine end; v in and v out These are the cut-in wind speed and the cut-out wind speed, respectively; v N Rated wind speed; P N This refers to the rated power of the wind turbine generator set; (b) Simulation setup of LVRT process in wind farm At different wind speeds and wind directions, a three-phase short-circuit fault occurs at the grid connection point of the wind farm at time t0, and the fault is cleared at time t1. By adjusting the grounding impedance, different voltage drop depths are obtained, and the active power and reactive power time series of the wind farm are obtained through simulation. Step 3: Train the GRU-LSTM-FC combined network using the training set data to obtain the GRU-LSTM-FC combined network model parameters. Based on the GRU-LSTM-FC combined network model parameters, use a genetic algorithm to optimize the number of FC layers and the number of neurons in each layer of the GRU-LSTM-FC combined network to obtain the optimized GRU-LSTM-FC combined network model. Step 4: Input the wind speed data, wind direction data of each wind turbine in the test set and the voltage time series of the grid connection point into the optimized GRU-LSTM-FC combined network model, and output the dynamic response time series of the active power and reactive power of the target wind farm.

2. The wind farm dynamic equivalent modeling method based on GRU-LSTM-FC combined network according to claim 1, characterized in that, In step 1, the target wind farm is built on the PSASP platform and includes a doubly fed induction generator (DFIG) and a direct-drive turbine (PMSG). The target wind farm is connected to the CEPRI-36 node system after two-stage voltage boosting.

3. The wind farm dynamic equivalent modeling method based on GRU-LSTM-FC combined network according to claim 1, characterized in that, Step 3 involves training the GRU-LSTM-FC combined network using the training set data to obtain the GRU-LSTM-FC combined network model parameters. This specifically includes the following steps: (a) Construction of GRU-LSTM-FC combined network The constructed combined network consists of three layers: the first layer is a GRU layer, the second layer is an LSTM layer, and the third layer is a FC layer. Data passes through the GRU layer, LSTM layer, and FC layer in sequence. (b) Performance evaluation metrics for combined networks The mean squared error (MSE) metric is used to evaluate the degree of matching between the output and the actual response of the GRU-LSTM-FC combined network. The absolute error at each time point is used as another metric to describe the instantaneous error over the entire transient interval, as detailed below: In the formula, y i Let i be the actual response value of the wind farm at time i; The output of the GRU-LSTM-FC combined network at time i; N = 1000, where N is the total number of data samples taken during the entire time-domain simulation. (c) Determining the number of layers in the combined network The performance of five networks—single-layer LSTM, single-layer GRU, two-layer LSTM, two-layer GRU, and single-layer GRU-single-layer LSTM—is evaluated based on iterative efficiency and the mean squared error of the training and test sets. In all five cases, the number of fully connected layers (FC) is uniformly set to 1, and the total number of hidden layer neurons is set to 50. The iteration termination error for training the combined network is set to 10. -2 The maximum number of iterations is set to 30,000, and the model learning rate is I. r It can be expressed as an inverse proportional function with the number of iterations (epochs) as the independent variable, specifically: In the formula, I r0 =10 -2 This is the initial learning rate of the model; epoch is the number of iterations during the model training process. Based on the mean squared errors of the five networks on the training and test sets, it was found that the error was minimized when GRU had 1 layer and LSTM had 1 layer. Therefore, it was determined that GRU had 1 layer and LSTM had 1 layer.

4. The wind farm dynamic equivalent modeling method based on GRU-LSTM-FC combined network according to claim 1, characterized in that, Step 3 involves optimizing the number of FC layers and the number of neurons in each layer of the GRU-LSTM-FC ensemble network using a genetic algorithm to obtain the optimized GRU-LSTM-FC ensemble network model. This process includes the following steps: (a) Hyperparameters to be optimized in the combined network The hyperparameters to be optimized include the number of FC layers, and the number of neurons in GRU, LSTM, and FC. (b) Hyperparameter optimization based on genetic algorithm Population chromosome encoding is performed, and the chromosome C structure of individuals in the population is constructed based on the number of FC layers and the number of neurons in each hidden layer as follows: C=[L FC ,N GRU ,N LSTM ,N FC1. ,...,N FCn ] In the formula, L FC N represents the number of layers in the FC (Full-Functional) layer. GRU N represents the number of neurons in the GRU layer. LSTM N represents the number of neurons in the LSTM layer. FC1 ,…,N FCn This represents the number of neurons in each layer of the FC layer. The fitness function is designed to comprehensively evaluate the performance of the combined network based on three metrics: accuracy, generalization, and training efficiency. The fitness function J is designed as the sum of these three metrics, and its expression is as follows: In the formula, J is the fitness function; Loss train0 and Loss train The ratio of the mean squared error of the training set to the total mean squared error is used to characterize the accuracy; Loss test0 and Loss test T0 represents the mean squared error of the test set and the total mean squared error, respectively. The ratio of these two values ​​is used to characterize the model's generalization ability. T0 represents the training duration, which is the average duration of multiple training iterations. start and T end These are the start and end times of training, in minutes, and the ratio T0 / (T) is... end -T start The training efficiency of the network is represented by α, β, and γ, which are the weight coefficients of each index. Based on the iteration termination condition, the optimized hyperparameter results are obtained.

5. A wind farm dynamic equivalent modeling system based on a GRU-LSTM-FC combined network, used to implement the method as described in claim 1, characterized in that, include: Modeling Unit: Based on the PSASP platform, construct a target wind farm containing n wind turbine units, and connect the target wind farm to the CEPRI-36 node system after two-stage voltage boosting; Simulation Unit: Based on wind speed data, wind direction data, and voltage drop disturbance at the grid connection point, the target wind farm is simulated to obtain the output power of the wind farm, and training and testing sets are constructed. Optimization Unit: The GRU-LSTM-FC combined network is trained using the training set data. The number of GRU layers and LSTM layers in the GRU-LSTM-FC combined network is determined by performance metrics. The number of FC layers and the number of neurons in each layer of the combined network are optimized using a genetic algorithm to obtain the optimized GRU-LSTM-FC combined network model parameters. Prediction Unit: Based on the input data of a certain wind speed, wind direction and voltage drop disturbance in the test set, the optimized GRU-LSTM-FC combined network is used to make predictions and obtain the dynamic response time series of active and reactive power of the wind farm.

6. The wind farm dynamic equivalent modeling system based on GRU-LSTM-FC combined network according to claim 5, characterized in that, In the modeling unit, the target wind farm is boosted in two stages and then connected to the CEPRI-36 node system via the BUS30 bus.

Images