CN-121997865-A - Double-fed fan parameter identification method based on depth space-time feature extraction and attention weighting

CN121997865ACN 121997865 ACN121997865 ACN 121997865ACN-121997865-A

Abstract

The invention discloses a doubly-fed fan parameter identification method based on deep space-time feature extraction and Attention weighting, which comprises the steps of constructing a data acquisition and high-correlation feature set, constructing a CNN-LSTM-Attention hybrid neural network identification model, training an identification model, optimizing parameters and identifying parameters of a doubly-fed fan controller, solving the problems that the doubly-fed fan controller is low in parameter identification precision, depends on artificial experience, is difficult to effectively extract deep space-time features from multivariable time sequence data and realizes multi-parameter high-precision synchronous identification at present under transient working conditions, and being capable of automatically extracting effective features from transient response data, accurately capturing time sequence dependency and realizing the intelligent method of multi-parameter high-precision synchronous identification so as to improve the reliability of the fan model and the accuracy of power grid simulation.

Inventors

ZHANG BIAO
Chao Shenghu
YE WEI
SHANG GUOBIN
YUAN KUN
ZHANG YUNFENG
WANG MINGDE
MA ZHIMING
CHANG JIE
MA HAIFENG

Assignees

青海德泓电力科技有限公司

Dates

Publication Date: 20260508
Application Date: 20251218

Claims (10)

1. The doubly-fed fan parameter identification method based on depth space-time feature extraction and attention weighting is characterized by comprising the following steps of: S1, constructing a data acquisition and high-correlation feature set: Collecting multivariable dynamic response time sequence data of the doubly fed wind turbine under voltage drop transient conditions at different depths by using a semi-physical test platform; Dividing the whole process into a plurality of time sequence intervals according to the physical characteristics of the low voltage ride through process, extracting the statistical characteristics of key electric quantity in each interval, screening the characteristics strongly related to the parameters of the controller by adopting a Pearson correlation coefficient method, and constructing a high-quality input characteristic set; S2, constructing a CNN-LSTM-attribute hybrid neural network identification model: The CNN-LSTM-Attention hybrid neural network identification model is a deep learning model integrating a convolutional neural network, a long-short-time memory network and an Attention mechanism, local space-time characteristics in an input characteristic sequence are extracted by using the convolutional neural network, long-time dependency in dynamic response of a system is captured by using the long-short-time memory network, the Attention mechanism is introduced to carry out self-adaptive weighted focusing on a key identification period, and finally estimated values of all to-be-identified controller parameters are output through a full-connection layer; s3, training an identification model and optimizing parameters: Training the hybrid neural network model established in the step S2 by utilizing the feature set established in the step S1, and iteratively updating model parameters through a back propagation algorithm and an optimizer to enable the model output to gradually approach to a real controller parameter value until the model converges; s4, identifying parameters of the doubly-fed fan controller: And (3) inputting dynamic response data of the fan to be identified under the transient working condition into a trained CNN-LSTM-Attention model after the same characteristic processing flow of the step S1, and directly outputting high-precision identification results of all controller parameters of the rotor side and the network side converters after forward propagation of the model.
2. The method for identifying parameters of a doubly-fed wind turbine based on depth spatiotemporal feature extraction and attention weighting according to claim 1, wherein in step S1, data acquisition and high correlation feature set construction specifically comprises: Different voltage drop depths, drop starting time and drop duration are set through the RT-LAB semi-physical platform, and multivariable time sequence data of the doubly-fed fan in the whole low-voltage ride through process are collected, wherein the multivariable time sequence data comprise direct-current voltage u dc , d-axis current i dg and q-axis current i qg .
3. The method for identifying parameters of a doubly-fed wind turbine based on depth spatiotemporal feature extraction and attention weighting according to claim 2, wherein in step S1, the construction of the data collection and high correlation feature set specifically further comprises: the whole low voltage ride through process is divided into 6 time sequence intervals, namely starting 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 parameters of a doubly-fed wind turbine based on depth spatiotemporal feature extraction and attention weighting as claimed in claim 3, wherein in step S1, the construction of the data collection and high correlation feature set further comprises: In each time sequence interval, calculating statistical characteristics of the collected key electric quantity data, including mean value, maximum value, minimum value, variance and peak value, and converting the original time sequence data into static characteristic vectors capable of representing dynamic characteristics of each stage to form an initial Gao Weite collection; And (3) adopting a Pearson correlation coefficient method to perform feature screening: Calculating a Pearson correlation coefficient P between each extracted feature and each controller parameter to be identified, wherein the calculation formula is as follows: ; Wherein, the For the sample values of the feature sequence, For the sample values of the parameter sequence to be identified, 、 And setting a correlation threshold or selecting the feature with the absolute value of the correlation coefficient ranked at the front, and eliminating irrelevant or redundant features to form a final model input feature set.
5. The method for identifying parameters of a doubly-fed fan based on depth spatiotemporal feature extraction and Attention weighting according to claim 1, wherein in step S2, the CNN-LSTM-Attention hybrid neural network identification model comprises: An input layer for receiving the fixed dimension feature vector screened in the step S1, and setting an input feature sequence as X epsilon RT X DX epsilon RT X D, wherein T is a time step and D is a feature dimension; the CNN feature extraction module consists of a one-dimensional convolution layer; the LSTM time sequence modeling module inputs the feature sequence extracted by the CNN into the LSTM layer; An attention mechanism module, which applies an attention mechanism to the hidden state sequence h= [ H 1 ,h 2 ,...,h T ] output by the LSTM; the output layer inputs the context vector c generated by the attention mechanism into the full connection layer, and finally outputs the predicted values of all the controller parameters to be recognized: ; Where W o and b o are the weights and biases of the output layers.
6. The method for identifying parameters of a doubly-fed fan based on depth spatiotemporal feature extraction and attention weighting according to claim 5, wherein in the CNN feature extraction module, assuming that the input feature sequence is x= [ X 1 ,x 2 ,...,x T ] T , the convolution kernel is ω e R K (K is the convolution kernel size), the convolution output y t at time step t is: ; Wherein b is a bias term, a plurality of convolution kernels of different sizes are used to capture local spatiotemporal features of different scales, and nonlinearities are introduced by the ReLU activation function.
7. The method for identifying parameters of a doubly-fed fan based on depth spatiotemporal feature extraction and attention weighting according to claim 5, wherein the LSTM unit updates the cell state c t and the hidden state h t through a gating mechanism, and the core calculation is as follows: ; Wherein i t 、f t 、o t is an input gate, a forgetting gate and an output gate respectively, sigma is a sigmoid activation function, as well as indicates element-by-element multiplication, and W i ,W f ,W o ,W c and b i ,b f ,b o ,b c are learnable parameters, and the module is used for learning long-term dependency of features in a time dimension.
8. The method for doubly fed fan parameter identification based on depth spatiotemporal feature extraction and attention weighting according to claim 5 wherein the attention mechanism module first calculates an attention score e t for each time step: ; wherein W a ,b a ,v a is a learnable parameter, and then normalizing by a softmax function to obtain an attention weight alpha t : ; finally, the hidden states are weighted and summed to obtain a context vector c: ; The attention mechanism module enables the model to adaptively focus on time segments that are more critical to parameter identification.
9. The method for identifying parameters of a doubly-fed wind turbine based on depth spatiotemporal feature extraction and attention weighting according to claim 1, wherein in step S3, the specific process of training an identification model and optimizing parameters includes: the loss function adopts a mean square error MSE, and is used for measuring the difference between a model prediction parameter and a real parameter, and the calculation formula is as follows: ; Wherein N is the number of samples, M is the number of parameters to be identified, For the actual value of the j parameter of the i sample, Is the corresponding predicted value; The optimizer adopts an adaptive moment estimation Adam algorithm, the initial learning rate is set to be 0.001, and a learning rate attenuation strategy can be used according to training conditions; the batch size is set to 32, and training rounds adopt an early-stop strategy according to the performance of the model on the verification set to prevent overfitting; the collected data set is proportionally divided into a training set, a verification set and a test set for model training, super-parameter tuning and final performance evaluation.
10. The method for identifying parameters of a doubly-fed fan based on depth spatiotemporal feature extraction and attention weighting according to claim 1, wherein in step S4, the doubly-fed fan controller parameter identification is applied as follows: And (3) for the dynamic response data of the doubly-fed wind turbine under the new and unknown transient conditions, a model is not required to be retrained, the dynamic response data is converted into an input format specified by the model through the standardized feature processing flow of the step S1, including time sequence interval division, statistical feature extraction and feature screening based on Pearson correlation coefficients, and then the input format is input into a trained CNN-LSTM-Attention model to output identification results of all target controller parameters.

Description

Double-fed fan parameter identification method based on depth space-time feature extraction and attention weighting Technical Field The invention belongs to the technical field of power grid simulation, and particularly relates to a doubly-fed fan parameter identification method based on deep space-time feature extraction and attention weighting. Background Doubly Fed Induction Generators (DFIGs) are currently the dominant wind turbine type, the dynamics of which are mainly determined by the controller parameters of the back-to-back converters. However, in the power grid simulation and the safety and stability analysis, most of widely used fan models are in a black box or gray box packaging form, and the parameters of an internal controller are opaque, so that researchers cannot establish an accurate fan simulation model, and the dynamic characteristic analysis and control of a power system under high-proportion wind power access are severely challenged. Therefore, the high-precision identification of the parameters of the fan controller is realized, and the method has important engineering significance for improving the simulation reliability of the power grid and ensuring the safe and stable operation of the system. At present, research methods for identifying parameters of a doubly-fed fan controller mainly can be divided into methods based on excitation signals, such as superposition of external excitation of M sequences and the like in measurement signals, identification by combining an optimization algorithm, wherein the methods can avoid direct operation of internal variables of the controller, but the application process of the excitation signals is complex and can possibly cause interference to the actual system operation, methods based on the optimization algorithm, such as a chaotic cuckoo algorithm, a particle swarm algorithm and the like, are effective in partial scenes by constructing an objective function and iteratively searching for optimal parameters, are sensitive to the initial value range of the parameters and easily fall into a local optimal solution, and are insufficient in stability of identification results, and methods based on traditional neural networks, such as BP networks, RNNs and the like, can learn mapping relations from data, but face complex multivariable, strong coupling and long-time-sequence dependent dynamic response data in the transient process of the fan, are difficult to effectively extract deep space-time characteristics, and result in limited identification accuracy under transient conditions such as voltage drop. The method is characterized in that the method is not sufficient in utilization of data characteristics, contribution degree of data to parameter identification in different stages such as fault occurrence, persistence and recovery in the transient process cannot be effectively distinguished, capturing capability of a model structure on long-term dependency in a time sequence is insufficient, complex nonlinear time sequence correlation between controller parameters and dynamic response is difficult to characterize, and the calculation efficiency and the identification precision are difficult to compromise, and particularly under the high-precision requirement, time consumption of model training and parameter optimization is long. Disclosure of Invention The invention aims to solve the technical problems that the doubly-fed fan controller parameter identification precision is low, manual experience is relied on, deep space-time characteristics are difficult to be effectively extracted from multivariable time sequence data, and multi-parameter high-precision synchronous identification is realized. In order to achieve the technical effects, the technical scheme adopted by the invention is as follows: A doubly-fed fan parameter identification method based on depth space-time feature extraction and attention weighting comprises the following steps: S1, constructing a data acquisition and high-correlation feature set: The method comprises the steps of collecting multivariable dynamic response time sequence data of the doubly fed fans under voltage drop transient conditions at different depths by using an RT-LAB semi-physical test platform; Dividing the whole process into a plurality of time sequence intervals according to the physical characteristics of the low voltage ride through process, extracting the statistical characteristics of key electric quantity in each interval, screening the characteristics strongly related to the parameters of the controller by adopting a Pearson correlation coefficient method, and constructing a high-quality input characteristic set; S2, constructing a CNN-LSTM-attribute hybrid neural network identification model: The CNN-LSTM-Attention hybrid neural network identification model is a deep learning model integrating a convolutional neural network, a long-short-time memory network and an Attention mechanism, local space-time cha