Doubly-fed fan parameter identification method based on depth spatial-temporal feature extraction and attention weighting

Abstract

The invention discloses a double-fed fan parameter identification method based on deep spatial-temporal feature extraction and attention weighting. The method comprises the steps of data acquisition and high-correlation feature set construction, CNN-LSTM-Attention hybrid neural network identification model construction, identification model training and parameter optimization and double-fed fan controller parameter identification. The method solves the problems that the parameter identification precision of the doubly-fed fan controller is low, the doubly-fed fan controller depends on artificial experience, deep spatio-temporal characteristics are difficult to effectively extract from multivariable time sequence data and multi-parameter high-precision synchronous identification is difficult to realize under the transient working condition at present, can automatically extract effective characteristics from transient response data, accurately captures a time sequence dependency relationship, and improves the identification accuracy of the parameters of the doubly-fed fan controller. And an intelligent method for multi-parameter high-precision synchronous identification is realized, so that the credibility of a fan model and the accuracy of power grid simulation are improved.

Description

Technical Field

[0001] This invention belongs to the field of power grid simulation technology, and in particular relates to a method for identifying parameters of doubly fed wind turbines based on deep spatiotemporal feature extraction and attention weighting. Background Technology

[0002] Doubly fed induction generators (DFIGs), as the mainstream wind turbine type, have dynamic characteristics primarily determined by the controller parameters of their back-to-back converters. However, in grid simulation and safety and stability analysis, the wind turbine models widely used are often "black box" or "gray box" encapsulated forms, with opaque internal controller parameters. This makes it impossible for researchers to build accurate wind turbine simulation models, posing a serious challenge to the dynamic characteristic analysis and control of power systems with high proportions of wind power integration. Therefore, achieving high-precision identification of wind turbine controller parameters is of significant engineering importance for improving the reliability of grid simulation and ensuring the safe and stable operation of the system.

[0003] Currently, research methods for parameter identification of doubly fed wind turbine controllers can be mainly divided into the following categories: First, methods based on excitation signals, such as superimposing external excitations like M-sequences into the measurement signals and then combining them with optimization algorithms for identification. Although these methods can avoid directly manipulating the internal variables of the controller, the application process of the excitation signal is complex and may interfere with the actual system operation. Second, methods based on optimization algorithms, such as the chaotic cuckoo algorithm and particle swarm optimization. These methods construct objective functions and iteratively search for optimal parameters. Although effective in some scenarios, they are sensitive to the initial value range of parameters and are prone to getting trapped in local optima, resulting in insufficient stability of the identification results. Third, methods based on traditional neural networks, such as BP networks and RNNs. These methods can learn mapping relationships from data, but when faced with complex multivariate, strongly coupled, and long-term time-dependent dynamic response data during wind turbine transient processes, it is difficult to effectively extract their deep spatiotemporal features, resulting in limited identification accuracy under transient conditions such as voltage drops.

[0004] Specifically, existing methods generally suffer from the following shortcomings when dealing with transient processes such as low-voltage ride-through: First, they do not fully utilize data features and fail to effectively distinguish the contribution of data from different stages of the transient process, such as the fault occurrence, duration, and recovery stages, to parameter identification; second, the model structure is not capable enough of capturing long-term dependencies in time series, making it difficult to characterize the complex nonlinear temporal correlation between controller parameters and dynamic response; third, computational efficiency and identification accuracy are often difficult to balance, especially under high-precision requirements, the model training and parameter optimization processes are time-consuming. Therefore, it is necessary to design a doubly-fed wind turbine parameter identification method based on deep spatiotemporal feature extraction and attention weighting to solve the above problems. Summary of the Invention

[0005] The technical problem to be solved by this invention is to provide a method for identifying parameters of doubly fed wind turbines based on deep spatiotemporal feature extraction and attention weighting. This method aims to solve the problems of low accuracy in identifying parameters of doubly fed wind turbine controllers under transient conditions, reliance on human experience, difficulty in effectively extracting deep spatiotemporal features from multivariate time series data, and achieving high-precision synchronous identification of multiple parameters.

[0006] To achieve the above-mentioned technical effects, the technical solution adopted by the present invention is as follows: A method for identifying parameters of doubly fed wind turbines based on deep spatiotemporal feature extraction and attention weighting includes the following steps: S1, Data Acquisition and Construction of Highly Correlated Feature Sets: The RT-LAB hardware-in-the-loop test platform was used to collect multivariable dynamic response time series data of a doubly-fed wind turbine under transient operating conditions with different voltage dip depths. Based on the physical characteristics of the low voltage ride-through process, the entire process is divided into multiple time intervals. Statistical features of key electrical quantities in each interval are extracted, and the Pearson correlation coefficient method is used to screen features that are strongly correlated with controller parameters to construct a high-quality input feature set. S2, Construct a CNN-LSTM-Attention hybrid neural network recognition model: The CNN-LSTM-Attention hybrid neural network recognition model is a deep learning model that integrates convolutional neural networks, long short-term memory networks, and attention mechanisms. It uses convolutional neural networks to extract local spatiotemporal features from the input feature sequence, uses long short-term memory networks to capture long-term time dependencies in the dynamic response of the system, and introduces attention mechanisms to adaptively weight and focus on key recognition periods. Finally, it outputs the estimated values ​​of all controller parameters to be identified through fully connected layers. S3, Recognition Model Training and Parameter Optimization: The hybrid neural network model established in step S2 is trained using the feature set constructed in step S1. The model parameters are iteratively updated through the backpropagation algorithm and optimizer, so that the model output gradually approaches the true controller parameter values ​​until the model converges. S4, Doubly fed fan controller parameter identification: The dynamic response data of the wind turbine under transient operating conditions to be identified is input into the trained CNN-LSTM-Attention model after undergoing the same feature processing procedure as in step S1. After forward propagation, the model directly outputs the high-precision identification results of all controller parameters of the rotor-side and grid-side converters.

[0007] Preferably, in step S1, data acquisition and construction of a highly correlated feature set specifically include: Using the RT-LAB hardware platform, different voltage drop depths (such as 20% and 80% of rated voltage) and drop start and duration were set to collect multivariate time series data of the doubly fed wind turbine during the entire low voltage ride-through process, including DC voltage udc, d-axis current idg and q-axis current iqg.

[0008] Preferably, in step S1, the data acquisition and construction of the highly correlated feature set further includes: Based on the "Guidelines for Modeling Electrical Simulation Models of Wind Turbine Units" and the physical characteristics of transient processes, the entire low-voltage ride-through process is divided into 6 time intervals: startup dynamic A, steady-state operation B, voltage drop transient C, voltage drop steady-state D, voltage recovery E, and steady-state recovery F.

[0009] Preferably, in step S1, the data acquisition and construction of the highly correlated feature set further includes: Within each time series interval, the statistical characteristics of the collected key electrical quantities are calculated, including mean, maximum, minimum, variance, and peak value. The original time series data are then converted into static feature vectors that can characterize the dynamic characteristics of each stage, forming an initial high-dimensional feature set. Feature selection was performed using the Pearson correlation coefficient method. Calculate the Pearson correlation coefficient P between each extracted feature and each controller parameter to be identified (such as the scaling coefficients Kp1~Kp7 and the integral coefficients Ki1~Ki7). The calculation formula is as follows: ; in, For the feature sequence sample values, For the sample values ​​of the parameter sequence to be identified, , These are their mean values, and n is the number of samples. Set a correlation threshold or select features with the highest absolute values ​​of correlation coefficients, and remove irrelevant or redundant features to form the final model input feature set.

[0010] Preferably, in step S2, the CNN-LSTM-Attention hybrid neural network recognition model includes: The input layer receives the fixed-dimensional feature vectors after filtering in step S1. Let the input feature sequence be X∈RT×DX∈RT×D, where T is the time step and D is the feature dimension. The CNN feature extraction module consists of one-dimensional convolutional layers; The LSTM temporal modeling module inputs the feature sequences extracted by the CNN into the LSTM layer; The attention mechanism module processes the hidden state sequence H output by the LSTM. h 1,h 2,..., h T Apply attention mechanisms; The output layer takes the context vector c generated by the attention mechanism and inputs it into the fully connected layer, ultimately outputting the predicted values ​​^y for all the controller parameters to be identified. ; Among them, W o and b o These are the weights and biases for the output layer.

[0011] Preferably, in the CNN feature extraction module, the input feature sequence is X=[x1,x2,...,x...]. T ] T The convolution kernel is ω∈R K ( K (where y is the kernel size), then the convolution output y at time step t is... t for: ; in, b The bias term is used; multiple convolutional kernels of different sizes are used to capture local spatiotemporal features at different scales, and nonlinearity is introduced through the ReLU activation function.

[0012] Preferably, the LSTM unit updates the cell state through a gating mechanism. c t and hidden state h t Its core calculations are as follows: ; in, i t , f t , o t These are the input gate, forget gate, and output gate, respectively; σ is the sigmoid activation function; ⊙ represents element-wise multiplication; W i W f W o W c and b i ,b f ,b o ,b c These are learnable parameters; this module is used to learn the long-term dependencies of features over time.

[0013] Preferably, the attention mechanism module first calculates the attention score for each time step. e t : ; Among them, W a ,b a ,v a These are learnable parameters; then the attention weights are obtained by normalization using the softmax function. α t : ; Finally, the hidden states are weighted and summed to obtain the context vector c: ; The attention mechanism module enables the model to adaptively focus on time segments that are more critical for parameter identification.

[0014] Preferably, in step S3, the specific process of identifying model training and parameter optimization includes: The loss function uses the mean squared error (MSE) to measure the difference between the model's predicted parameters and the true parameters. Its calculation formula is as follows: ; Where N is the number of samples and M is the number of parameters to be identified. For the i-th sample, the j-th parameter is the true value. The corresponding predicted value; The optimizer uses the Adam algorithm for adaptive moment estimation, with an initial learning rate of 0.001, and can use a learning rate decay strategy based on the training situation. The batch size is set to 32, and the training epochs adopt an early stopping strategy based on the model's performance on the validation set to prevent overfitting. The collected dataset is divided into training, validation, and test sets according to a set ratio for model training, hyperparameter tuning, and final performance evaluation.

[0015] Preferably, in step S4, the application process of doubly-fed wind turbine controller parameter identification is as follows: For new and unknown transient operating conditions of doubly fed wind turbine dynamic response data, there is no need to retrain the model. The data is transformed into the input format specified by the model through the standardized feature processing process in step S1, including time interval division, statistical feature extraction and feature screening based on Pearson correlation coefficient. Then, it is input into the pre-trained CNN-LSTM-Attention model to output the identification results of all target controller parameters.

[0016] The beneficial effects of this invention are as follows: 1. Traditional methods for identifying parameters of doubly-fed induction generator (DFIG) wind turbine controllers often rely on optimization algorithms or manual feature design, which suffers from problems such as sensitivity to initial values, susceptibility to local optima, and difficulty in extracting complex temporal dependencies during transient processes. This invention takes a data-driven approach, constructing a CNN-LSTM-Attention hybrid neural network. It utilizes convolutional neural networks to automatically extract local spatiotemporal features from the dynamic response, effectively captures long-term temporal dependencies of the system through a long short-term memory network, and adaptively focuses on key time periods for parameter identification using an attention mechanism. This method achieves high-precision end-to-end mapping from raw dynamic data to controller parameters, eliminating the need for manual feature design or multiple iterative optimizations. It significantly improves the automation and efficiency of identification, solving the problems of reliance on expert experience and complex computational processes inherent in traditional methods.

[0017] 2. Traditional neural network methods such as BP and RNN often struggle to balance identification accuracy and generalization ability when processing high-dimensional, nonlinear, and strongly coupled time-series data such as wind turbine transient responses. This invention introduces an attention mechanism, enabling the model to differentiate the weighting of features at different time steps, strengthening the information contribution of key identification periods, and suppressing interference from noise and irrelevant periods, thereby significantly improving the accuracy and robustness of parameter identification. Experimental results show that under two typical transient conditions of 20% and 80% voltage drop, the proposed method can control the identification error of most of the 14 control parameters of the rotor-side and grid-side converters to within 1%, which is significantly better than traditional BP and RNN models, verifying its effectiveness and superiority in achieving high-precision synchronous identification of multiple parameters in complex transient processes.

[0018] 3. The method proposed in this invention possesses good engineering applicability and generalization ability. By acquiring and training the model using measured data through the RT-LAB hardware-in-the-loop platform, the model can adapt to different voltage dip conditions and exhibits stable identification performance under various operating states. This method can not only be used for offline parameter identification but also provides a feasible technical approach for future online monitoring and adaptive control, playing a crucial supporting role in improving the simulation accuracy and operational safety of wind power grid-connected systems. Attached Figure Description

[0019] Figure 1 This is a control structure diagram of a doubly fed wind turbine and its grid-side converter. Figure 2 This is a schematic diagram illustrating the time interval division of the low-voltage ride-through process. Figure 3 Pearson correlation coefficients and schematic diagrams for each feature and the parameter to be identified; Figure 4 The measured waveform of DC capacitor voltage under an 80% voltage drop condition is shown. Figure 5The measured waveform of DC capacitor voltage under a 20% voltage drop condition is shown. Figure 6 This is a graph showing the change in the loss function during model training. Figure 7 A comparison chart of the output responses of different models under an 80% voltage drop condition; Figure 8 A comparison chart of the output responses of different models under a 20% voltage drop condition; Figure 9 This is a flowchart of the overall process for parameter identification of doubly fed wind turbines based on CNN-LSTM-Attention. Detailed Implementation

[0020] Example 1: like Figure 1 As shown, the parameter identification method for doubly fed wind turbines based on deep spatiotemporal feature extraction and attention weighting includes the following steps: S1, Data Acquisition and Construction of Highly Correlated Feature Sets: The RT-LAB hardware-in-the-loop test platform was used to collect multivariable dynamic response time series data of a doubly-fed wind turbine under transient operating conditions with different voltage dip depths. Based on the physical characteristics of the low voltage ride-through process, the entire process is divided into multiple time intervals. Statistical features of key electrical quantities in each interval are extracted, and the Pearson correlation coefficient method is used to screen features that are strongly correlated with controller parameters to construct a high-quality input feature set. S2, Construct a CNN-LSTM-Attention hybrid neural network recognition model: The CNN-LSTM-Attention hybrid neural network recognition model is a deep learning model that integrates convolutional neural networks, long short-term memory networks, and attention mechanisms. It uses convolutional neural networks to extract local spatiotemporal features from the input feature sequence, uses long short-term memory networks to capture long-term time dependencies in the dynamic response of the system, and introduces attention mechanisms to adaptively weight and focus on key recognition periods. Finally, it outputs the estimated values ​​of all controller parameters to be identified through fully connected layers. S3, Recognition Model Training and Parameter Optimization: The hybrid neural network model established in step S2 is trained using the feature set constructed in step S1. The model parameters are iteratively updated through the backpropagation algorithm and optimizer, so that the model output gradually approaches the true controller parameter values ​​until the model converges. S4, Doubly fed fan controller parameter identification: The dynamic response data of the wind turbine under transient operating conditions to be identified is input into the trained CNN-LSTM-Attention model after undergoing the same feature processing procedure as in step S1. After forward propagation, the model directly outputs the high-precision identification results of all controller parameters of the rotor-side and grid-side converters.

[0021] Preferably, in step S1, data acquisition and construction of a highly correlated feature set specifically include: Using the RT-LAB hardware platform, different voltage drop depths (such as 20% and 80% of rated voltage) and drop start and duration were set to collect multivariate time series data of the doubly fed wind turbine during the entire low voltage ride-through process, including DC voltage udc, d-axis current idg and q-axis current iqg.

[0022] Preferably, in step S1, the data acquisition and construction of the highly correlated feature set further includes: Based on the "Guidelines for Modeling Electrical Simulation Models of Wind Turbine Units" and the physical characteristics of transient processes, the entire low-voltage ride-through process is divided into 6 time intervals: startup dynamic A, steady-state operation B, voltage drop transient C, voltage drop steady-state D, voltage recovery E, and steady-state recovery F.

[0023] Preferably, in step S1, the data acquisition and construction of the highly correlated feature set further includes: Within each time series interval, the statistical characteristics of the collected key electrical quantities are calculated, including mean, maximum, minimum, variance, and peak value. The original time series data are then converted into static feature vectors that can characterize the dynamic characteristics of each stage, forming an initial high-dimensional feature set. Feature selection was performed using the Pearson correlation coefficient method. Calculate the Pearson correlation coefficient P between each extracted feature and each controller parameter to be identified (such as the scaling coefficients Kp1~Kp7 and the integral coefficients Ki1~Ki7). The calculation formula is as follows: ; in, For the feature sequence sample values, For the sample values ​​of the parameter sequence to be identified, , These are their mean values, and n is the number of samples. Set a correlation threshold or select features with the highest absolute values ​​of correlation coefficients, and remove irrelevant or redundant features to form the final model input feature set.

[0024] Preferably, in step S2, the CNN-LSTM-Attention hybrid neural network recognition model includes: The input layer receives the fixed-dimensional feature vectors after filtering in step S1. Let the input feature sequence be X∈RT×DX∈RT×D, where T is the time step and D is the feature dimension. The CNN feature extraction module consists of one-dimensional convolutional layers; The LSTM temporal modeling module inputs the feature sequences extracted by the CNN into the LSTM layer; The attention mechanism module processes the hidden state sequence H output by the LSTM. h 1, h 2,..., h T Apply attention mechanisms; The output layer takes the context vector c generated by the attention mechanism and inputs it into the fully connected layer, ultimately outputting the predicted values ​​^y for all the controller parameters to be identified. ; Among them, W o and b o These are the weights and biases for the output layer.

[0025] Preferably, in the CNN feature extraction module, the input feature sequence is X=[x1,x2,...,x...]. T ] T The convolution kernel is ω∈R K ( K (where y is the kernel size), then the convolution output y at time step t is... t for: ; in, b The bias term is used; multiple convolutional kernels of different sizes are used to capture local spatiotemporal features at different scales, and nonlinearity is introduced through the ReLU activation function.

[0026] Preferably, the LSTM unit updates the cell state through a gating mechanism. c t and hidden state h t Its core calculations are as follows: ; in, i t , f t , o t These are the input gate, forget gate, and output gate, respectively; σ is the sigmoid activation function; ⊙ represents element-wise multiplication; W i W f W o W c and b i,b f ,b o ,b c These are learnable parameters; this module is used to learn the long-term dependencies of features over time.

[0027] Preferably, the attention mechanism module first calculates the attention score for each time step. e t : ; Among them, W a ,b a ,v a These are learnable parameters; then the attention weights are obtained by normalization using the softmax function. α t : ; Finally, the hidden states are weighted and summed to obtain the context vector c: ; The attention mechanism module enables the model to adaptively focus on time segments that are more critical for parameter identification.

[0028] Preferably, in step S3, the specific process of identifying model training and parameter optimization includes: The loss function uses the mean squared error (MSE) to measure the difference between the model's predicted parameters and the true parameters. Its calculation formula is as follows: ; Where N is the number of samples and M is the number of parameters to be identified. For the i-th sample, the j-th parameter is the true value. The corresponding predicted value; The optimizer uses the Adam algorithm for adaptive moment estimation, with an initial learning rate of 0.001, and can use a learning rate decay strategy based on the training situation. The batch size is set to 32, and the training epochs adopt an early stopping strategy based on the model's performance on the validation set to prevent overfitting. The collected dataset is divided into training, validation, and test sets according to a set ratio for model training, hyperparameter tuning, and final performance evaluation.

[0029] Preferably, in step S4, the application process of doubly-fed wind turbine controller parameter identification is as follows: For new and unknown transient operating conditions of doubly fed wind turbine dynamic response data, there is no need to retrain the model. The data is transformed into the input format specified by the model through the standardized feature processing process in step S1, including time interval division, statistical feature extraction and feature screening based on Pearson correlation coefficient. Then, it is input into the pre-trained CNN-LSTM-Attention model to output the identification results of all target controller parameters.

[0030] Example 2: like Figure 9 As shown, the parameter identification method for doubly fed wind turbines based on deep spatiotemporal feature extraction and attention weighting provided in this embodiment includes the following steps: Step 1: Data Collection and Construction of Highly Correlated Feature Sets Taking a 1.5MW doubly-fed induction generator as an example, the dynamic response data of this generator during low-voltage ride-through was obtained using the RT-LAB hardware-in-the-loop simulation platform. Two voltage sag conditions were set: sag to 20% of rated voltage and sag to 80% of rated voltage, with a sag initiation time of 0.9s and a duration of 0.7s. 100 sets of data were collected for each condition, resulting in a total of 200 sets of data for model training and 40 sets for testing. Key electrical quantities collected included DC-side voltage. u dc d-axis current of grid-side converter i dg q-axis current i qg Grid connection point voltage u g The sampling frequency is 10kHz.

[0031] According to the "Guidelines for Modeling Electrical Simulation Models of Wind Turbine Units", the entire low voltage ride-through process is divided into 6 time intervals with clear physical meaning.

[0032] like Figure 2 As shown: A (start-up dynamic, 0-0.5s), B (steady-state operation, 0.5-0.9s), C (voltage drop transient, 0.9-1.0s), D (voltage drop steady-state, 1.0-1.6s), E (voltage recovery, 1.6-2.2s), F (steady-state recovery, 2.2-2.7s).

[0033] Within each time interval, the collected data u dc , i dg , i qg The mean, maximum, minimum, variance, and peak-to-peak value of the three key electrical quantities were calculated. Initially, a total of 90-dimensional feature vectors were obtained (3 electrical quantities × 6 intervals × 5 features).

[0034] The Pearson correlation coefficient method was used for feature selection. The above 90 features and 14 controller parameters to be identified were calculated. K p1 , K i1 ,..., K p7 , K i7 The correlation coefficient between the features and the parameters was used, and a threshold of 0.3 was set to remove features with an absolute correlation coefficient less than 0.3. Finally, 11 highly correlated features were selected to form the input feature set. The Pearson coefficients of each feature and parameter were calculated as follows: Figure 3 As shown.

[0035] Step 2: Construct a CNN-LSTM-Attention hybrid neural network recognition model: The constructed CNN-LSTM-Attention model structure specifically includes the following modules: (1) Input layer: receives tensors with dimensions batch_size, time_steps=6, features=11, where time_steps corresponds to 6 time intervals.

[0036] (2) CNN Feature Extraction Module: This module contains two one-dimensional convolutional layers. The first layer uses 32 convolutional kernels of size 3, and the second layer uses 64 convolutional kernels of size 3. The activation function for both layers is ReLU, with a stride of 1. "Same" padding is used to maintain the length of the time dimension. This module is used to automatically extract local correlation patterns between features in different time intervals.

[0037] (3) LSTM Temporal Modeling Module: Consists of two stacked LSTM layers, each containing 64 hidden units, used to capture long-term temporal dependencies across intervals in dynamic response data. The calculation of LSTM units is as described in the specification.

[0038] (4) Attention Mechanism Module: An attention mechanism is applied to the hidden states of all time steps in the output of the second-layer LSTM. The attention score for each time step is calculated using a learnable weight matrix, and then normalized to weights using softmax. Finally, the weighted sum is used to obtain the context vector. This mechanism allows the model to focus on time intervals that are more critical for parameter identification (such as the steady-state interval D of voltage drop).

[0039] (5) Output layer: Input the context vector into a fully connected layer and output a 14-dimensional vector, which corresponds to the estimated values ​​of 14 controller parameters.

[0040] Step 3: Identification Model Training and Parameter Optimization: The model was trained using the 200 sets of training data prepared in step 1. Before training, the input features were Z-score standardized. The training parameters were set as follows: batch size of 32, initial learning rate of 0.001, Adam optimizer, and mean squared error (MSE) loss function. The training process lasted for 200 epochs, and early stopping was performed using a validation set (20% of the training data) to prevent overfitting.

[0041] The curve of the loss function during training is shown below. Figures 4-6 As shown, the model's loss decreases rapidly in the early stages of training (approximately the first 50 epochs) and then stabilizes. The loss curves on the training and validation sets converge with the same trend and do not show significant divergence, indicating that the model has good generalization ability. Finally, the model's loss value on the training set stabilizes at around 0.0032, meeting the requirements for high-precision parameter identification.

[0042] Step 4: Parameter identification test and verification of the doubly fed wind turbine controller: Forty sets of test data that were not used in training and validation were input into the trained model for parameter identification. The test data covered two voltage drop depths: 20% and 80%. The identification results and errors of some key controller parameters are shown in Table 1 below: Table 1: Identification results and errors of some key controller parameters;

[0043] As shown in Table 1 above, the identification error of most parameters is controlled within 1%, with only Ki4 having a slightly higher error. Overall, the identification accuracy meets the requirements of engineering applications.

[0044] To verify the superiority of the proposed method, it was compared with traditional BP neural network and RNN models on the same dataset. The structure and training parameters of the BP and RNN models are shown in the table in the specification. A comparison of the output response (taking DC voltage as an example) of the three models under voltage drop conditions of 80% and 20% with the measured waveforms is shown below. Figure 7 and Figure 8 As shown.

[0045] from Figure 7 and Figure 8 It can be clearly seen that: The identification output of the BP model (red dashed line) deviates significantly from the measured curve (black solid line) during transient processes (especially the voltage recovery stage), indicating insufficient tracking capability.

[0046] The RNN model (blue dotted line) outperforms the BP model, but still exhibits significant lag and error during periods of intense fluctuation.

[0047] The output response of the CNN-LSTM-Attention model (green dashed line) proposed in this embodiment has the highest degree of agreement with the measured curve, especially during the critical low voltage ride-through period (1.0-1.6s), it can more accurately reproduce the dynamic characteristics of the system.

[0048] Step 5: Comprehensive analysis of identification efficiency and accuracy: Table 2 below compares the training time and average recognition error of the three models under the same hardware platform (Intel Core i7, 16GB RAM, NVIDIA GTX 1060 GPU) and software environment: Table 2: Training time and average recognition error of the three models;

[0049] As shown in Table 2, although the training time of the model in this embodiment is slightly longer than that of the comparison model, its root mean square error (RMSE) and maximum parameter error on the test set are significantly reduced. This indicates that the method in this embodiment, by introducing CNN and attention mechanisms, achieves a significant improvement in recognition accuracy at an acceptable time cost, effectively solving the core problem of insufficient accuracy of traditional methods under transient conditions.

[0050] In summary, this embodiment provides an efficient and high-precision intelligent identification scheme for doubly-fed induction generator (DFIG) controller parameters through systematic data feature processing, innovative deep learning model architecture, and thorough experimental verification. It has significant practical value for improving the accuracy of wind turbine modeling and simulation and the level of power grid safety and stability analysis.

Claims

1. A method for identifying parameters of a doubly fed wind turbine based on deep spatiotemporal feature extraction and attention weighting, characterized in that, Includes the following steps: S1, Data Acquisition and Construction of Highly Correlated Feature Sets: Multivariable dynamic response time-series data of a doubly fed wind turbine under transient operating conditions with different voltage dip depths were collected using a hardware-in-the-loop test platform. Based on the physical characteristics of the low voltage ride-through process, the entire process is divided into multiple time intervals. Statistical features of key electrical quantities in each interval are extracted, and the Pearson correlation coefficient method is used to screen features that are strongly correlated with controller parameters to construct a high-quality input feature set. S2, Construct a CNN-LSTM-Attention hybrid neural network recognition model: The CNN-LSTM-Attention hybrid neural network recognition model is a deep learning model that integrates convolutional neural networks, long short-term memory networks, and attention mechanisms. It uses convolutional neural networks to extract local spatiotemporal features from the input feature sequence, uses long short-term memory networks to capture long-term time dependencies in the dynamic response of the system, and introduces attention mechanisms to adaptively weight and focus on key recognition periods. Finally, it outputs the estimated values ​​of all controller parameters to be identified through fully connected layers. S3, Recognition Model Training and Parameter Optimization: The hybrid neural network model established in step S2 is trained using the feature set constructed in step S1. The model parameters are iteratively updated through the backpropagation algorithm and optimizer, so that the model output gradually approaches the true controller parameter values ​​until the model converges. S4, Doubly fed fan controller parameter identification: The dynamic response data of the wind turbine under transient operating conditions to be identified is input into the trained CNN-LSTM-Attention model after undergoing the same feature processing procedure as in step S1. After forward propagation, the model directly outputs the high-precision identification results of all controller parameters of the rotor-side and grid-side converters.

2. The method for identifying doubly fed wind turbine parameters based on deep spatiotemporal feature extraction and attention weighting according to claim 1, characterized in that, In step S1, data collection and construction of a highly relevant feature set specifically include: Using the RT-LAB hardware-in-the-loop platform, different voltage sag depths, sag initiation times, and durations were set to collect multivariate time-series data of the doubly-fed induction generator (DFIG) during the entire low-voltage ride-through process, including DC voltage. u dc d-axis current i dg and q-axis current i qg .

3. The method for identifying doubly fed wind turbine parameters based on deep spatiotemporal feature extraction and attention weighting according to claim 2, characterized in that, In step S1, data collection and construction of highly correlated feature sets specifically include: The entire low-voltage ride-through process is divided into 6 time intervals: startup dynamic A, steady-state operation B, voltage drop transient C, voltage drop steady-state D, voltage recovery E, and steady-state recovery F.

4. The method for identifying doubly fed wind turbine parameters based on deep spatiotemporal feature extraction and attention weighting according to claim 3, characterized in that, In step S1, data collection and construction of highly correlated feature sets specifically include: Within each time series interval, the statistical characteristics of the collected key electrical quantities are calculated, including mean, maximum, minimum, variance, and peak value. The original time series data are then converted into static feature vectors that can characterize the dynamic characteristics of each stage, forming an initial high-dimensional feature set. Feature selection was performed using the Pearson correlation coefficient method. The Pearson correlation coefficient P between each extracted feature and each controller parameter to be identified is calculated using the following formula: ； in, For the feature sequence sample values, For the sample values ​​of the parameter sequence to be identified, , These are their mean values, and n is the number of samples. Set a correlation threshold or select features with the highest absolute values ​​of correlation coefficients, and remove irrelevant or redundant features to form the final model input feature set.

5. The method for identifying doubly fed wind turbine parameters based on deep spatiotemporal feature extraction and attention weighting according to claim 1, characterized in that, In step S2, the CNN-LSTM-Attention hybrid neural network recognition model includes: The input layer receives the fixed-dimensional feature vectors after filtering in step S1. Let the input feature sequence be X∈RT×DX∈RT×D, where T is the time step and D is the feature dimension. The CNN feature extraction module consists of one-dimensional convolutional layers; The LSTM temporal modeling module inputs the feature sequences extracted by the CNN into the LSTM layer; The attention mechanism module processes the hidden state sequence H output by the LSTM. h 1, h 2,..., h T Apply attention mechanisms; The output layer takes the context vector c generated by the attention mechanism and inputs it into the fully connected layer, ultimately outputting the predicted values ​​^y for all the controller parameters to be identified. ； Among them, W o and b o These are the weights and biases for the output layer.

6. The method for identifying doubly fed wind turbine parameters based on deep spatiotemporal feature extraction and attention weighting according to claim 5, characterized in that, In the CNN feature extraction module, let the input feature sequence be X=[x1,x2,...,x...]. T ] T The convolution kernel is ω∈R K ( K (where is the kernel size), then the convolution output y at time step t is... t for: ； in, b The bias term is used; multiple convolutional kernels of different sizes are used to capture local spatiotemporal features at different scales, and nonlinearity is introduced through the ReLU activation function.

7. The method for identifying doubly fed wind turbine parameters based on deep spatiotemporal feature extraction and attention weighting according to claim 5, characterized in that, LSTM units update cell state through a gating mechanism. c t and hidden state h t Its core calculations are as follows: ； in, i t , f t , o t These are the input gate, forget gate, and output gate, respectively; σ is the sigmoid activation function; ⊙ represents element-wise multiplication; W i W f W o W c and b i ,b f ,b o ,b c These are learnable parameters; this module is used to learn the long-term dependencies of features over time.

8. The method for identifying doubly fed wind turbine parameters based on deep spatiotemporal feature extraction and attention weighting according to claim 5, characterized in that, The attention mechanism module first calculates the attention score at each time step. e t : ； Among them, W a ,b a ,v a These are learnable parameters; then the attention weights are obtained by normalization using the softmax function. α t : ； Finally, the hidden states are weighted and summed to obtain the context vector c: ； The attention mechanism module enables the model to adaptively focus on time segments that are more critical for parameter identification.

9. The method for identifying doubly fed wind turbine parameters based on deep spatiotemporal feature extraction and attention weighting according to claim 1, characterized in that, In step S3, the specific process of identifying model training and parameter optimization includes: The loss function uses the mean squared error (MSE) to measure the difference between the model's predicted parameters and the true parameters. Its calculation formula is as follows: ； Where N is the number of samples and M is the number of parameters to be identified. For the i-th sample, the j-th parameter is the true value. The corresponding predicted value; The optimizer uses the Adam algorithm for adaptive moment estimation, with an initial learning rate of 0.001, and can use a learning rate decay strategy based on the training situation. The batch size is set to 32, and the training epochs adopt an early stopping strategy based on the model's performance on the validation set to prevent overfitting. The collected dataset is divided into training, validation, and test sets according to a set ratio for model training, hyperparameter tuning, and final performance evaluation.

10. The method for identifying doubly fed wind turbine parameters based on deep spatiotemporal feature extraction and attention weighting according to claim 1, characterized in that, In step S4, the application process of doubly-fed wind turbine controller parameter identification is as follows: For new and unknown transient operating conditions of doubly fed wind turbine dynamic response data, there is no need to retrain the model. The data is transformed into the input format specified by the model through the standardized feature processing process in step S1, including time interval division, statistical feature extraction and feature screening based on Pearson correlation coefficient. Then, it is input into the pre-trained CNN-LSTM-Attention model to output the identification results of all target controller parameters.

Images