Wind power frequency modulation control method, system and device based on VAE-GRU and medium

Abstract

The invention discloses a VAE-GRU-based wind power frequency modulation control method, system, equipment and medium, and relates to the technical field of automatic control and intelligent frequency modulation, and the method comprises the steps: collecting multi-working-condition power grid data, solving an optimal control parameter through a swarm intelligence optimization algorithm, constructing a high-dimensional sample set, and obtaining a high-dimensional sample set; a variable auto-encoder is utilized to encode state data into low-dimensional feature vectors, the vectors are input into a gating cycle unit network, a dynamic mapping relation between the vectors and control parameters is learned, power grid data are collected in real time during online application, and optimal control parameters can be adaptively output and wind power frequency modulation is realized through trained model processing. According to the method, the self-adaptive mapping relation from the power grid working condition to the control parameters is established, the kinetic energy recovery of the rotor is ensured, the secondary frequency drop of the wind power grid-connected system is effectively inhibited, and the frequency control accuracy and the system robustness are improved.

Description

Technical Field

[0001] This invention relates to the field of automatic control and intelligent frequency regulation technology, specifically to a wind power frequency regulation control method, system, equipment, and medium based on VAE-GRU. Background Technology

[0002] Existing wind turbines are connected to the grid via power electronic converters, decoupling the rotor speed from the grid frequency. This weakens the inherent inertia support and primary frequency regulation capability of traditional synchronous generators, reducing the system's equivalent inertia. Consequently, the grid's frequency change rate increases and the minimum frequency point decreases when subjected to disturbances. In severe cases, this may trigger a chain reaction, endangering the safe and stable operation of the system.

[0003] Despite the progress made in existing wind power frequency regulation technology, several technical bottlenecks remain to be addressed. First, traditional fixed-parameter control strategies struggle to adapt to complex and ever-changing grid operating conditions. When the system operating point, network topology, or disturbance type changes, fixed parameters cannot guarantee optimal control performance, potentially leading to insufficient frequency support or secondary frequency drops due to over-adjustment. Second, while some intelligent algorithm-based control methods possess a degree of adaptability, they suffer from significant deficiencies in data processing and model building. On one hand, grid state data is highly dimensional and noisy, and existing methods often lack effective feature extraction and noise reduction mechanisms, resulting in low-quality model input information and impacting decision accuracy. On the other hand, when establishing the mapping relationship from grid state to control parameters, simple neural networks or static models struggle to fully capture the temporal dependencies in the system's dynamic processes, failing to accurately characterize the complex evolution of control parameters under changing operating conditions, thus limiting the generalization ability and dynamic response performance of the control strategy. Summary of the Invention

[0004] In view of the above-mentioned existing problems, the present invention provides a wind power frequency regulation control method, system, equipment and medium based on VAE-GRU to solve the problems in the prior art that secondary frequency drops are caused by insufficient frequency support or over-adjustment, the quality of model input information is not high, the accuracy of decision-making is affected, and the complex laws of control parameters evolving with operating conditions cannot be accurately described.

[0005] To address the aforementioned technical issues, a wind power frequency regulation control method based on VAE-GRU is proposed, including: The system collects state data of the power grid under different operating conditions and, based on the goal of frequency stability, solves for the optimal control parameters under each condition using a swarm intelligence optimization algorithm. It then associates the state data with the optimal control parameters to construct a high-dimensional sample set. The state data from this high-dimensional sample set is input into a feature extraction and denoising model, which maps the high-dimensional input data to a low-dimensional latent space through an encoding network, generating low-dimensional feature vectors. These low-dimensional feature vectors are then input into a time-series dynamic modeling network, which learns the dynamic dependency between the feature sequence and the optimal control parameters through a gating mechanism, establishing a nonlinear mapping model. During the operation of the wind power grid-connected system, the system collects power grid state data, processes it through the feature extraction and denoising model and the time-series dynamic modeling network, adaptively outputs the optimal control parameters for the current operating condition, and applies them to the wind power frequency controller for frequency support and secondary sag suppression.

[0006] As a preferred embodiment of the wind power frequency regulation control method based on VAE-GRU described in this invention, the swarm intelligent optimization algorithm includes: constructing an objective function containing control parameter constraints with the goal of maximizing the minimum system frequency; initializing a set of candidate parameters as the search starting point; and simulating swarm cooperative behavior for multiple rounds of iteration. In each iteration, the adaptability of the candidate parameters is evaluated according to the objective function, and the position of the candidate parameters is updated based on the evaluation results. The parameters under different operating conditions are substituted into the system frequency response model to verify the effectiveness of the updated parameters. When the iteration result meets the preset convergence condition, the search is terminated and the current optimal parameters are output and combined with the corresponding working condition data to form a sample.

[0007] As a preferred embodiment of the wind power frequency regulation control method based on VAE-GRU described in this invention, the feature extraction and noise reduction model includes preprocessing the original high-dimensional state data to eliminate dimensional differences and remove abnormal samples. The preprocessed data is nonlinearly transformed by an encoding network, and the expected value vector and standard deviation vector representing the data distribution are output. Based on the reparameterization method, a low-dimensional feature vector is generated by combining a random vector sampled from the standard distribution with the expected value vector and the standard deviation vector. The low-dimensional feature vectors are reconstructed using a decoding network, and the difference between the reconstructed data and the original data is used as the optimization objective to train the encoding and decoding networks.

[0008] As a preferred embodiment of the wind power frequency regulation control method based on VAE-GRU described in this invention, the time-series dynamic modeling network includes inputting low-dimensional feature vectors into the network in a time sequence or feature dependency order. The degree of forgetting of historical information is controlled by resetting the gate, and candidate states are calculated in combination with the current input; The fusion ratio between the previous state and the candidate state is determined by updating the gate, and the hidden state at the current moment is output. Based on the final hidden state, the predicted control parameter values ​​are obtained through mapping in a fully connected layer.

[0009] As a preferred embodiment of the wind power frequency regulation control method based on VAE-GRU described in this invention, the method of solving for the optimal control parameters under each operating condition includes: constructing an input sample set using the gray wolf algorithm; calculating the distance vector between individual gray wolves and prey using the coordination coefficient vector; updating the position vector of gray wolves based on the distance; selecting α wolves, β wolves, and δ wolves as leader individuals based on fitness values; calculating the distance between the remaining gray wolves and the leader individuals; guiding the gray wolves to move based on the position of the leader individuals; and updating the candidate parameter positions. The search pattern is determined based on the convergence factor. The steps of finding prey, hunting, and attacking are repeated. After each iteration, the fitness value is recalculated and the leader individual is updated until the number of iterations reaches the maximum value. After the iteration terminates, the position of the α wolf is taken as the optimal over-firing power and over-firing duration under the current working condition and included in the sample set. The formula for updating the gray wolf's position vector is expressed as: in, Let be the distance vector between the gray wolf and its prey. As the swing factor, Let be the position of the prey in the t-th iteration. Let be the current position of the gray wolf in the t-th iteration. The convergence factor is Let be the current position of the gray wolf in the (t+1)th iteration. and A random number in the range [0,1]. The coefficients decrease linearly from 2 to 0 with the number of iterations; The formula for updating the candidate parameter position is expressed as: in, Let be the distance between the i-th gray wolf and the k-th leader wolf. Let the position of the leader wolf be defined in the t-th iteration. Let i be the position of the i-th ordinary gray wolf in the t-th iteration. Let be the convergence factor of the i-th wolf. Let be the swing factor of the i-th wolf. , and Three candidate positions are calculated for the gray wolf based on the positions of α, β, and δ wolves. This represents the final position of the gray wolf in the (t+1)th iteration.

[0010] As a preferred embodiment of the wind power frequency regulation control method based on VAE-GRU described in this invention, the generation of low-dimensional feature vectors includes a variational autoencoder network consisting of multiple fully connected layers, an input layer dimension matching the feature number of high-dimensional data, an intermediate layer undergoing nonlinear transformation through the ReLU activation function, and an output layer generating a mathematical expectation vector and a standard deviation vector, respectively. Random vectors are sampled from the standard normal distribution. The expected value vector and the standard deviation vector are combined with the random vectors, and latent variables are generated through element-level operations. The decoding network adopts a symmetric structure, inputs the latent variables into multiple fully connected layers, restores the original input dimension, and outputs reconstructed data. The objective function is to minimize the reconstruction error and the sum of KL divergence. The probabilistic relationship between latent variables and observed data is expressed as follows: in, For joint probability density distribution, The latent variable represents the low-dimensional feature vector extracted by the variational coding network. The observed variables are the original high-dimensional input data. Let be the mathematical expectation vector, and be the mean parameter output by the variational coding network. The standard deviation vector is the variance parameter output by the variational coding network. For vectors It is a diagonal matrix composed of diagonal elements. It follows a multivariate normal distribution. This is an element-wise multiplication of vectors; The weights of the encoding and decoding networks are iteratively optimized using the training sample set until the loss function converges. The high-dimensional data from the training and test sets are then input into the encoding network, which outputs a low-dimensional feature vector. This vector is then combined with the optimal control parameter label to form an enhanced sample set. The loss function is expressed as: in, Let the variational lower bound loss function be... For Kullback-Leibler divergence, This is the joint probability distribution, i.e., the joint distribution of the latent variable z and the observed data x under parameter θ. The variational posterior distribution, i.e., the parameters The approximate distribution of the true posterior distribution fitted by the sub-encoding network. To decode network parameters, To encode network parameters; The establishment of the nonlinear mapping model includes: calculating the reset gate output using the sigmoid activation function, calculating the update gate output using the sigmoid activation function, controlling the fusion ratio of historical information and current information, calculating candidate hidden states using the tanh activation function, fusing the previous hidden state and candidate hidden states through the update gate, outputting the hidden state at the current moment, and inputting the final hidden state into the fully connected layer, mapping it to the control parameter dimension, and outputting the predicted values ​​of over-generation power and over-generation duration.

[0011] As a preferred embodiment of the wind power frequency regulation control method based on VAE-GRU described in this invention, the optimal control parameters for the current operating condition are as follows: when the grid frequency drop is detected to exceed a preset threshold, the wind turbine is controlled to release rotor kinetic energy and increase electromagnetic power output to provide short-term frequency support based on the predicted over-generation power and over-generation duration. After the over-generation phase ends, the power output is restored according to the maximum power point tracking curve, and the rotor speed recovery and frequency stability are balanced by controlling the recovery rate. The control parameters are dynamically updated based on real-time operating condition data.

[0012] The beneficial effects of this preferred technical solution are that by establishing a dynamic mapping relationship through the gated loop unit, the temporal characteristics of the power grid state change are accurately captured, the adaptive prediction of control parameters is realized, and the predicted parameters are transformed into specific control actions through real-time control decision execution, thus coordinating the balance between frequency support and rotor kinetic energy recovery.

[0013] As a preferred embodiment of the wind power frequency regulation control system based on VAE-GRU described in this invention, it is characterized by including an optimal control parameter sample generation module, a high-dimensional data feature extraction and noise reduction module, a dynamic mapping relationship learning module, and a real-time adaptive control execution module.

[0014] The optimal control parameter sample generation module is used to establish a frequency response model of the wind power grid-connected system, with the optimization objective of maximizing the minimum system frequency. Using a swarm intelligence optimization algorithm, it performs simulation calculations under different wind speeds, wind power ratios, and load disturbance conditions, automatically finds the optimal SIC control parameters, and outputs the corresponding optimal parameters.

[0015] The high-dimensional data feature extraction and noise reduction module is used to utilize the nonlinear mapping capability of the variational autoencoder to map multi-dimensional operating condition data containing noise and redundant information to a low-dimensional, smooth latent feature space, thereby generating a low-dimensional feature vector.

[0016] The dynamic mapping relationship learning module is used to learn the nonlinear relationship between the low-dimensional feature sequence processed by the variational autoencoder and the optimal SIC parameters by using a gated cyclic unit time-series dynamic modeling network. The update gate and reset gate mechanism inside the gated cyclic unit captures the dynamic dependencies and long-term patterns in the data and predicts the control parameters under different disturbance scenarios.

[0017] The real-time adaptive control execution module is used to collect real-time status data of the power grid, process it through a feature extraction module and a dynamic mapping learning module, adaptively output the most suitable over-generation power and over-generation duration under the current operating conditions, and send the parameters into the wind power's stepwise inertial control system to dynamically adjust the control curve.

[0018] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of a method for wind power frequency regulation control based on VAE-GRU.

[0019] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a method for wind power frequency regulation control based on a VAE-GRU.

[0020] The beneficial effects of this invention are as follows: By collecting multi-condition data and using swarm intelligence optimization algorithms to solve for the optimal control parameters, this invention constructs a high-precision sample set, laying a solid data foundation for model learning; by using a variational autoencoder to extract features and reduce noise from high-dimensional state data, generating low-dimensional feature vectors, it improves the robustness of the model and reduces computational complexity; and by using a gated recurrent unit network to learn the dynamic dependency relationship between feature sequences and control parameters, it establishes an accurate nonlinear mapping model, overcoming the shortcomings of traditional methods in handling time-series correlations; when applied online, the model can adaptively output the optimal control parameters according to the real-time grid state, achieving effective frequency support and secondary drop suppression under different operating conditions, significantly improving the dynamic support capability of wind farms for the grid and the safety and stability level of the system. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 The above is a flowchart of a wind power frequency regulation control method based on VAE-GRU, which is provided as an embodiment of the present invention.

[0023] Figure 2 This is a schematic diagram of the principle structure of a variational encoder for a wind power frequency regulation control method based on VAE-GRU, provided as an embodiment of the present invention.

[0024] Figure 3 This is a schematic diagram of a gated loop network structure for a wind power frequency regulation control method based on VAE-GRU, provided as an embodiment of the present invention.

[0025] Figure 4 This is a schematic diagram illustrating the frequency response characteristics of a wind power frequency regulation control method based on VAE-GRU under different power increase amounts, as provided in an embodiment of the present invention.

[0026] Figure 5 This is a schematic diagram illustrating the frequency response characteristics of a wind power frequency regulation control method based on VAE-GRU under different durations, as provided in an embodiment of the present invention.

[0027] Figure 6 This is a schematic diagram comparing the RCS index of the power generation of a wind power frequency regulation control method based on VAE-GRU according to an embodiment of the present invention.

[0028] Figure 7 This is a schematic diagram comparing the RCS index of the optimal control duration of a wind power frequency regulation control method based on VAE-GRU, provided in an embodiment of the present invention.

[0029] Figure 8 The present invention provides a system scheme flowchart for a wind power frequency regulation control system based on VAE-GRU according to one embodiment of the present invention.

[0030] In the picture, Let n be the standard deviation vector. Let be an n-dimensional mathematical expectation vector. Detailed Implementation

[0031] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0032] Example 1, referring to Figure 1 As an embodiment of the present invention, a wind power frequency regulation control method based on VAE-GRU is provided, comprising: S100: Collects state data of the power grid under different operating conditions, and based on the frequency stability target, solves the optimal control parameters under each operating condition through a swarm intelligence optimization algorithm, and associates the state data with the optimal control parameters to construct a high-dimensional sample set.

[0033] S200: Input the state data of the high-dimensional sample set into the feature extraction and denoising model, and map the high-dimensional input data to the low-dimensional latent space through the encoding network to generate low-dimensional feature vectors.

[0034] S300: Input low-dimensional feature vectors into a time-series dynamic modeling network, learn the dynamic dependency between feature sequences and optimal control parameters through a gating mechanism, and establish a nonlinear mapping model.

[0035] S400: During the operation of the wind power grid-connected system, it collects grid status data, processes it through feature extraction and noise reduction models and time-series dynamic modeling networks, adaptively outputs the optimal control parameters under the current operating conditions, and applies them to the wind power frequency regulation controller to provide frequency support and secondary drop suppression.

[0036] It should be noted that this invention constructs a high-quality training sample set through a swarm intelligence optimization algorithm, establishing a reliable data foundation covering all operating conditions. It also uses a variational autoencoder for feature extraction and noise reduction, effectively addressing noise interference and feature redundancy issues in power grid measurement data. It extracts core features strongly correlated with frequency regulation control, establishes a dynamic mapping relationship through a gated cyclic unit to capture the temporal characteristics of power grid state changes, achieves adaptive prediction of control parameters, and coordinates the balance between frequency support and rotor kinetic energy recovery through real-time control decision execution.

[0037] Example 2, refer to Figures 1-3 This is a second embodiment of the present invention, which provides a wind power frequency regulation control method based on VAE-GRU, including: In this embodiment of the application, in step S100, the construction of the high-dimensional sample set includes steps S101 to S105: S101: With the optimization objective of maximizing the minimum system frequency, construct an objective function that includes control parameter constraints. The objective function is derived based on the system frequency response model to ensure that the frequency dynamic characteristics are accurately represented. The objective function and constraints are expressed as follows: in, The first frequency drop depth represents the frequency deviation at the end of the inertial overshoot phase. The second frequency drop depth represents the frequency drop deviation that occurs during the power recovery phase. For inertial over-power generation, This represents the load disturbance. To control the time parameters, The minimum frequency safety threshold, This refers to the power drop. The MPPT power after the speed decreases. The coefficients for the optimal power curve are... The initial rotor speed, For the duration of inertial overfiring, For frequency conversion coefficients, Pi This is the initial output power. For inertial over-power generation, The MPPT power after the speed decreases. This represents the power drop.

[0038] S102: Initialize a set of candidate parameters as the starting point of the search, simulate group cooperative behavior for multiple rounds of iteration, and dynamically adjust the search direction and step size based on the fitness evaluation results in each round of iteration.

[0039] S103: In each iteration, the candidate parameters and the current operating condition data are substituted into the system frequency response model to calculate the frequency deviation response, and the adaptability of the candidate parameters is evaluated based on the objective function value. The formula for calculating the complex frequency domain response of the frequency deviation is expressed as: in, This represents the system frequency deviation, indicating the amount of deviation relative to the rated frequency. The natural oscillation frequency, For wind power penetration rate, This is the generator damping coefficient. This is the governor droop coefficient. This is the load damping coefficient. The reheat time constant is This is the power disturbance. For the Laplace operator, For the damping ratio, Let be the system's inertial time constant. This refers to the power ratio of the high-pressure cylinder; The formula for calculating the complex time-domain response of the frequency deviation is expressed as: in, For time variables, The oscillation amplitude coefficient, The frequency of the damped oscillation. The phase angle is the initial phase of the frequency oscillation.

[0040] S104: Based on the evaluation results, the positions of the candidate parameters are adjusted through a position update mechanism to gradually approach the global optimal solution, while verifying the effectiveness of the updated parameters under different perturbation scenarios.

[0041] S105: When the iteration result meets the preset convergence condition, terminate the search and output the current optimal parameter, combine it with the corresponding working condition data to form a sample, and include it in the training sample set. The training sample set includes the voltage amplitude of each node in the power grid before and after the disturbance, the voltage phase angle of each node in the power grid before and after the disturbance, the power of each generator in the power grid, the load size of each node in the power grid, the proportion of wind power in the power grid, the amount of user load disturbance, and the wind speed.

[0042] In an optional implementation, in step S100, constructing the high-dimensional sample set further includes calculating the area under the frequency deviation curve, considering the weight allocation for different time periods, and optimizing the parameters to minimize the integral.

[0043] In another optional implementation, in step S100, constructing the high-dimensional sample set may further include setting a frequency change rate threshold, adding a secondary drop depth constraint, and performing multi-objective trade-off optimization.

[0044] Furthermore, in this embodiment of the application, in step S104, the solution for the optimal control parameters under each operating condition includes steps A1~A4: A1: Construct the input sample set using the gray wolf algorithm, calculate the distance vector between individual gray wolves and prey using the coordination coefficient vector, and represent the distance between prey and gray wolves as: in, Let be the distance vector between the gray wolf and its prey. As the swing factor, Let be the position of the prey in the t-th iteration. Let t be the current position of the gray wolf in the t-th iteration.

[0045] A2: Update the position vector of the gray wolf based on the distance, select α wolf, β wolf and δ wolf as leader individuals based on fitness value, calculate the distance between the remaining gray wolves and the leader individuals, guide the gray wolves to move based on the position of the leader individuals, and update the candidate parameter positions. The formula for updating the gray wolf's position vector is expressed as: in, The convergence factor is Let be the current position of the gray wolf in the (t+1)th iteration. and A random number in the range [0,1]. The coefficient decreases linearly from 2 to 0 with the number of iterations.

[0046] A3: Determine the search mode based on the convergence factor, and repeatedly execute the steps of finding prey, encircling and attacking. After each iteration, recalculate the fitness value and update the leader individual until the number of iterations reaches the maximum value. Determine the search mode based on the value of A. When |A|>1, perform a global exploration. When |A|<1, perform a local fine-grained search, gradually narrowing the search range.

[0047] A4: After the iteration terminates, the position of α wolf is taken as the optimal over-generation power and over-generation duration under the current working condition, and included in the sample set; The formula for updating the candidate parameter position is expressed as: in, Let be the distance between the i-th gray wolf and the k-th leader wolf. Let the position of the leader wolf be defined in the t-th iteration. Let i be the position of the i-th ordinary gray wolf in the t-th iteration. Let be the convergence factor of the i-th wolf. Let be the swing factor of the i-th wolf. , and Three candidate positions are calculated for the gray wolf based on the positions of α, β, and δ wolves. This represents the final position of the gray wolf in the (t+1)th iteration.

[0048] In an optional implementation, in step S104, the solution for the optimal control parameters under each working condition further includes optimizing the parameters by simulating the flight and foraging behavior of bird flocks, initializing the position and velocity of the particle swarm, calculating the fitness (lowest frequency value) of each particle, updating the individual optimal and global optimal positions, adjusting the particle positions according to the velocity update formula, and iterating to the maximum number of times to output the optimal parameters.

[0049] In another optional implementation, in step S104, solving for the optimal control parameters under each operating condition may further include: using a genetic algorithm to optimize parameters, initializing the population, selecting individuals with high fitness, performing crossover and mutation operations to generate a new generation of population, and iteratively evolving to output the optimal solution.

[0050] In S200, the feature extraction and noise reduction model includes steps S201 to S205: S201: Preprocess the original high-dimensional state data, including data standardization to eliminate dimensional differences and statistical methods to remove outliers, to ensure the quality and consistency of the input data.

[0051] S202: The preprocessed data is nonlinearly transformed by an encoding network to output the mathematical expectation vector and standard deviation vector representing the potential distribution of the data. Dimension compression and feature abstraction are achieved through a multi-layer neural network.

[0052] S203: Based on the reparameterization method, a low-dimensional feature vector in the latent space is generated by combining a random vector sampled from the standard distribution with the mathematical expectation vector and the standard deviation vector, ensuring that the sampling process is differentiable and easy to optimize.

[0053] S204: The low-dimensional feature vector is reconstructed through the decoding network, and the reconstructed data is output. The difference between the reconstructed data and the original data is used as the optimization objective. The encoding and decoding networks are trained together so that the latent variables retain key working condition information after dimensionality reduction.

[0054] S205: After training, the original high-dimensional data is converted into low-dimensional feature vectors through the encoding network to form an enhanced sample set for subsequent modeling.

[0055] Furthermore, in this embodiment of the application, in step 200, the generation of the low-dimensional feature vector includes steps S211~S213: S211: The variational autoencoder network consists of multiple fully connected layers. The input layer dimension matches the feature number of high-dimensional data, the intermediate layer performs nonlinear transformation through the ReLU activation function, and the output layer generates the expected value vector and the standard deviation vector, respectively.

[0056] S212: Random vectors are sampled from the standard normal distribution. The expected value vector and the standard deviation vector are combined with the random vectors, and latent variables are generated through element-level operations. The decoding network adopts a symmetric structure, inputs the latent variables into multiple fully connected layers, restores the original input dimension, and outputs reconstructed data. The objective function is to minimize the reconstruction error and the sum of KL divergence. The probabilistic relationship between latent variables and observed data is expressed as follows: in, For joint probability density distribution, The latent variable represents the low-dimensional feature vector extracted by the variational coding network. The observed variables are the original high-dimensional input data. Let be the mathematical expectation vector, and be the mean parameter output by the variational coding network. The standard deviation vector is the variance parameter output by the variational coding network. For vectors It is a diagonal matrix composed of diagonal elements. It follows a multivariate normal distribution. This is the element-wise multiplication of vectors.

[0057] S213: Iteratively optimize the weights of the encoding and decoding networks using the training sample set until the loss function converges. Input the high-dimensional data from the training and test sets into the encoding network, output a low-dimensional feature vector, and combine it with the optimal control parameter label to form an enhanced sample set. The loss function is expressed as: in, Let the variational lower bound loss function be... For Kullback-Leibler divergence, This is the joint probability distribution, i.e., the joint distribution of the latent variable z and the observed data x under parameter θ. The variational posterior distribution, i.e., the parameters The approximate distribution of the true posterior distribution fitted by the sub-encoding network. To decode network parameters, To encode network parameters.

[0058] In an optional implementation, in step 200, generating the low-dimensional feature vector further includes extracting features layer by layer through a multi-layer encoder, pre-training each layer of the autoencoder, fine-tuning the network, and using a denoising autoencoder to enhance robustness, and extracting abstract features layer by layer.

[0059] In another optional implementation, in step 200, generating the low-dimensional feature vector may further include: first separating the noise components using wavelet transform, then performing linear dimensionality reduction using PCA, wavelet thresholding for denoising, calculating the covariance matrix, obtaining principal components through eigenvalue decomposition, and selecting the principal components to form a feature subset.

[0060] In this embodiment of the application, step S300, the temporal dynamic modeling network includes steps S301 to S304: S301: The reset gate determines how much information should be retained from the previous hidden state based on the current input, and calculates the candidate hidden state using the following formula: in, To reset the gate output vector, the value range is [0,1]. It is the sigmoid activation function. To reset the weight matrix of the gate, For time indexing, for The hidden state at all times The input features at the current time are the low-dimensional feature vectors after VAE processing. To reset the door's bias, This is a vector concatenation operation.

[0061] S302: The update gate evaluates the current input, determines how to fuse historical and new information, and calculates the candidate hidden state based on the control result of the reset gate. The formula is as follows: in, To update the gate output vector, To update the gate weight matrix, To update the gate's bias term, As a candidate hidden state, a new state is calculated based on the current input and some historical information. The hyperbolic tangent activation function is used. Here is the weight matrix for the candidate states. For the bias term of the candidate state, This is element-wise multiplication.

[0062] S303: The update gate, based on the calculated ratio, weights and fuses the hidden state from the previous time step with the newly generated candidate hidden states to obtain the final hidden state at the current time step. The formula is expressed as: in, The current hidden state. To retain a certain percentage of historical information, The proportion of new information adopted.

[0063] S304: To train the network efficiently, the ADAM optimization algorithm is used, which accelerates convergence and avoids getting trapped in local optima by dynamically adjusting the learning rate.

[0064] In an optional implementation, in step S300, the temporal dynamic modeling network further includes calculating three gating signals, updating the cell state, and outputting the hidden state.

[0065] In another optional implementation, in step S300, the temporal dynamic modeling network may further include constructing causal convolutions to ensure temporality, multi-layer dilated convolutions to expand the receptive field, and residual connections to prevent gradient vanishing.

[0066] Furthermore, in step S304, the ADAM optimization algorithm includes steps B1 to B3: B1: The cumulative historical gradient direction is estimated by calculating the first moment of the gradient, providing momentum for parameter updates. The formula is as follows: in, For the first moment estimate of the current time step u, This is an estimate of the first moment of the previous time step u-1. Let u be the gradient at the current time step. The first-order moment decay rate is set to 0.9.

[0067] B2: Calculate the second moment estimate of the cumulative historical gradient magnitude, and adaptively adjust the learning step size for each parameter. The calculation formula is expressed as: in, For the second moment estimate of the current time step u, This is an estimate of the second moment of the previous time step u-1. The second-order moment decay rate is set to 0.999.

[0068] B3: Perform moment estimation bias correction, and update the model parameters using the bias-corrected first-order and second-order moment estimates, as expressed by the formula: in, This is the first-order moment estimate after bias correction. This is the second-order moment estimate after bias correction. and The decay rate is a power of u, which decreases further as the number of training steps increases. For time indexing, The updated model parameters include the weights and biases of the GRU and VAE networks. These are the model parameters before the update. Based on the learning rate, For smoothing terms, This is the adaptive learning rate term, which adjusts the update magnitude for each parameter based on the gradient history.

[0069] In step S400, outputting the optimal control parameters under the current operating condition includes steps S401~S403: S401: Inertial Over-generation Stage: When the grid frequency drop is detected to exceed the preset threshold, the wind turbine is controlled to release rotor kinetic energy based on the predicted over-generation power and over-generation duration, thereby increasing electromagnetic power output and providing short-term frequency support.

[0070] S402: Power recovery stage: After the over-generation stage ends, the power output is gradually restored according to the maximum power point tracking curve to avoid power drop. The recovery rate is controlled to balance the rotor speed recovery and frequency stability, and to suppress secondary frequency drop.

[0071] S403: Parameter Adaptive Adjustment: Dynamically updates control parameters based on real-time operating data to ensure optimized frequency response under different wind speeds, wind power ratios, and load disturbances.

[0072] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

[0073] Example 3, referring to Figures 4-7 This is the third embodiment of the present invention, which provides a wind power frequency regulation control method based on VAE-GRU. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through experiments.

[0074] Step 1: An improved IEEE-39 node system model was built in PSCAD / EMTDC simulation software. The model replaced the synchronous unit at node 31 in the original system with an equivalent wind farm model. In the wind power grid connection system model, the equivalent wind farm was set to contain a total of 120 wind turbines, each with a rated capacity of 0.05 pu.

[0075] To simulate the changing proportion of wind power in the power grid, the equivalent wind farm's turbine capacity was set to vary between 60 and 120 units, and the turbine operating wind speed was set to vary between 5 m / s and 10 m / s. Using Python scripts, simulations were conducted on all generators and all loads in the system under various disturbance scenarios, including random turbine switching of 5%, 10%, 15%, 20%, 25%, and 30% for single and multiple generators, random load increases of 5%, 10%, 15%, 20%, 25%, and 30% for single and multiple loads, or a combination of both.

[0076] Step 2: For each scenario, use the wind speed, wind power ratio, network parameters, and the operating status of each node before and after the disturbance (including voltage amplitude, phase angle, generator power, load size, etc.) from the simulation analysis as sample input information, and use the simulation parameters to update the optimal wind power frequency regulation control parameter optimization model and set the optimal frequency regulation control parameters. , The lowest grid transient frequency is 49.8Hz. The frequency modulation control parameters corresponding to the current scenario are obtained through the GWO algorithm as sample output parameter information. A total of 5000 sets of original sample data are obtained to form the original sample set, which is divided into training sample set and test sample set in a 4:1 ratio.

[0077] To verify the necessity of the proposed method for optimizing wind power frequency regulation control parameters, based on the sample data acquisition method, an equivalent wind farm with 100 wind turbines was set, the wind speed of the turbines was 8 m / s, and the grid disturbances were 15% generator tripping at node 30, 20% load increase at node 15, and 5% load increase at node 23 as simulation conditions for a single case analysis. A fixed optimal control duration was set. At that time, under different power increase amounts, we obtained Figure 4 The frequency dynamic response characteristics are shown; similarly, a fixed optimal control power boost is set. When different durations are taken, the results are obtained. Figure 5 The frequency dynamic response characteristics are shown.

[0078] according to Figure 4 It can be seen that when the optimal control time When stationary, larger wind turbines have a stronger ability to support frequencies after disturbances, but at the same time, they experience a sudden drop in power during the speed recovery phase. The larger the frequency drop, the more severe the secondary frequency drop during the frequency recovery phase after the disturbance, resulting in poor overall wind power frequency regulation control; furthermore, through Figure 5 It can be seen that the wind turbine's ability to support frequency is limited after the disturbance is fixed. While a larger wind turbine can withstand disturbances for a longer period, it also leads to a more severe secondary frequency drop during the speed recovery phase and a longer overall frequency drop time. Analysis shows that it is necessary to obtain the optimal frequency regulation parameters of the wind turbine during wind power frequency regulation to ensure that the grid frequency can be quickly restored after the disturbance, and to minimize the degree of secondary frequency drop of the wind turbine during the frequency recovery phase, so as to ensure the best overall frequency regulation effect of wind power.

[0079] Step 3: Performance Validation of the Control Parameter Prediction Model. To evaluate the effectiveness of obtaining the optimal control parameters for wind turbine frequency regulation in this paper, the root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), coefficient of determination, and 90% absolute error quantile are used as performance evaluation indicators for the control parameter prediction model. The calculation formula is expressed as follows: in, The root mean square error, For the number of test samples, Let be the true value of the j-th sample, i.e., the optimal control parameter obtained by the GWO algorithm. This represents the predicted value for the j-th sample, i.e., the output of the VAE-GRU model. The mean absolute error, For absolute error, The mean absolute percentage error, This is a relative error. As the coefficient of determination, For quantile index, The average of the true values. The 90th percentile of absolute error. For the sorted absolute error sequence, The j-th error value, sorted in ascending order, where j is the position index. The length of the error sequence is the number of test samples.

[0080] To further evaluate the predictive performance of the models, the evaluation indicators for the predictive performance of each model were calculated according to the indicator formula, and the comparison is shown in Table 1 below: Table 1 Comparison of Prediction Performance of Each Model

[0081] When predicting optimal frequency regulation control parameters for wind power using the same training and test samples, the VAE-GRU model constructed in this invention significantly outperforms the comparative models VAE-SVR, VAE-RNN, and the single GRU model in all prediction performance evaluation metrics. Among them, the second-best model, GRU, performs best in MAPE, a key indicator reflecting prediction accuracy. The prediction error was reduced by 37.5%. The prediction error was reduced by 53.3%, which verifies that the VAE-GRU model proposed in this invention, with its powerful feature extraction capability for high-dimensional complex state data of the power grid by variational autoencoder, can effectively mine its deep potential features and correlation information. Combined with the modeling advantage of GRU for complex dependencies between features, it can significantly improve the prediction accuracy of optimal frequency regulation control parameters for wind power.

[0082] Considering that voltage, power, and other measurement signals in actual power grids may be affected by noise interference, impacting the prediction of optimal frequency regulation control parameters, to evaluate the robustness of the proposed method under noise interference, Gaussian white noise with different signal-to-noise ratios (SNR) levels was randomly added to the voltage and power data of the original test set. Prediction performance was tested at different SNR levels. To facilitate the evaluation of the comprehensive prediction performance of each model, a comprehensive index was designed, expressed by the following formula: in, The robustness score is calculated based on the overall performance.

[0083] pass Figure 6 and Figure 7 The RCS comparison shows that when the grid voltage and power measurement signals are interfered with by Gaussian white noise, the RCS of each prediction model fluctuates to varying degrees. Specifically, in a 5dB strong noise environment, the comprehensive index of the VAE-SVR model drops sharply to around 0.15, while our invention maintains a high accuracy of over 0.85. When the signal-to-noise ratio is below 15dB, the performance degradation of the VAE-RNN and GRU models intensifies significantly, indicating that conventional deep learning methods are insufficiently adaptable to strong noise. However, our invented VAE-GRU method maintains a comprehensive index of over 0.85 for each parameter prediction across a wide noise range of 5dB to 25dB. Furthermore, in a typical noise level of 15dB, the comprehensive index is more than 40.0% higher than the optimal comparison model GRU, and the performance degradation rate is much lower than GRU. This is mainly attributed to the fact that VAE effectively filters high-frequency interference components in measurement noise by performing nonlinear feature extraction and dimensionality reduction on high-dimensional grid state data, while the robust mapping relationship established by GRU further ensures the robustness of control parameter prediction.

[0084] Example 4, refer to Figure 8This is the fourth embodiment of the present invention. This embodiment provides a wind power frequency regulation control system based on VAE-GRU, including an optimal control parameter sample generation module, a high-dimensional data feature extraction and noise reduction module, a dynamic mapping relationship learning module, and a real-time adaptive control execution module.

[0085] The optimal control parameter sample generation module is used to establish a frequency response model of the wind power grid-connected system, with the optimization objective of maximizing the minimum system frequency. Using a swarm intelligence optimization algorithm, it performs simulation calculations under different wind speeds, wind power ratios, and load disturbance conditions, automatically finds the optimal SIC control parameters, and outputs the corresponding optimal parameters.

[0086] The high-dimensional data feature extraction and noise reduction module is used to utilize the nonlinear mapping capability of the variational autoencoder to map multi-dimensional operating condition data containing noise and redundant information to a low-dimensional, smooth latent feature space, thereby generating a low-dimensional feature vector.

[0087] The dynamic mapping relationship learning module is used to learn the nonlinear relationship between the low-dimensional feature sequence processed by the variational autoencoder and the optimal SIC parameters by using a gated cyclic unit time-series dynamic modeling network. The update gate and reset gate mechanism inside the gated cyclic unit captures the dynamic dependencies and long-term patterns in the data and predicts the control parameters under different disturbance scenarios.

[0088] The real-time adaptive control execution module is used to collect real-time status data of the power grid, process it through a feature extraction module and a dynamic mapping learning module, adaptively output the most suitable over-generation power and over-generation duration under the current operating conditions, and send the parameters into the wind power's stepwise inertial control system to dynamically adjust the control curve.

[0089] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

[0090] Example 5, the fifth embodiment of the present invention, differs from the previous four embodiments in that: If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0091] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0092] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0093] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

Claims

1. A wind power frequency regulation control method based on VAE-GRU, characterized in that: include, The system collects state data of the power grid under different operating conditions, and based on the frequency stability target, solves the optimal control parameters under each operating condition through a swarm intelligence optimization algorithm. The state data is then correlated with the optimal control parameters to construct a high-dimensional sample set. The state data of a high-dimensional sample set is input into the feature extraction and denoising model. The high-dimensional input data is mapped to a low-dimensional latent space through an encoding network to generate a low-dimensional feature vector. Low-dimensional feature vectors are input into a time-series dynamic modeling network, and the dynamic dependency between feature sequences and optimal control parameters is learned through a gating mechanism to establish a nonlinear mapping model. During the operation of the wind power grid-connected system, grid status data is collected, processed by feature extraction and noise reduction models and time-series dynamic modeling networks, and the optimal control parameters under the current operating conditions are adaptively output and applied to the wind power frequency regulation controller for frequency support and secondary drop suppression.

2. The wind power frequency regulation control method based on VAE-GRU as described in claim 1, characterized in that: The swarm intelligence optimization algorithm includes: constructing an objective function with control parameter constraints, with the optimization objective being to maximize the minimum system frequency; initializing a set of candidate parameters as the search starting point; and simulating swarm cooperative behavior for multiple rounds of iteration. In each iteration, the adaptability of the candidate parameters is evaluated according to the objective function, and the position of the candidate parameters is updated based on the evaluation results. The parameters under different operating conditions are substituted into the system frequency response model to verify the effectiveness of the updated parameters. When the iteration result meets the preset convergence condition, the search is terminated and the current optimal parameters are output and combined with the corresponding working condition data to form a sample.

3. The wind power frequency regulation control method based on VAE-GRU as described in claim 2, characterized in that: The feature extraction and noise reduction model includes preprocessing the original high-dimensional state data to eliminate dimensional differences and remove abnormal samples. The preprocessed data is nonlinearly transformed by an encoding network, and the expected value vector and standard deviation vector representing the data distribution are output. Based on the reparameterization method, a low-dimensional feature vector is generated by combining a random vector sampled from the standard distribution with the expected value vector and the standard deviation vector. The low-dimensional feature vectors are reconstructed using a decoding network, and the difference between the reconstructed data and the original data is used as the optimization objective to train the encoding and decoding networks.

4. The wind power frequency regulation control method based on VAE-GRU as described in claim 3, characterized in that: The time-series dynamic modeling network includes inputting low-dimensional feature vectors into the network in a time sequence or feature dependency order; The degree of forgetting of historical information is controlled by resetting the gate, and candidate states are calculated in combination with the current input; The fusion ratio between the previous state and the candidate state is determined by updating the gate, and the hidden state at the current moment is output. Based on the final hidden state, the predicted control parameter values ​​are obtained through mapping in a fully connected layer.

5. The wind power frequency regulation control method based on VAE-GRU as described in claim 4, characterized in that: The solution for the optimal control parameters under each working condition includes: constructing an input sample set using the gray wolf algorithm; calculating the distance vector between individual gray wolves and prey using the coordination coefficient vector; updating the position vector of gray wolves based on the distance; selecting α wolves, β wolves, and δ wolves as leader individuals based on fitness values; calculating the distance between the remaining gray wolves and the leader individuals; guiding the gray wolves to move based on the position of the leader individuals; and updating the candidate parameter positions. The search pattern is determined based on the convergence factor. The steps of finding prey, hunting, and attacking are repeated. After each iteration, the fitness value is recalculated and the leader individual is updated until the number of iterations reaches the maximum value. After the iteration terminates, the position of the α wolf is taken as the optimal over-firing power and over-firing duration under the current working condition and included in the sample set. The formula for updating the gray wolf's position vector is expressed as: in, Let be the distance vector between the gray wolf and its prey. As the swing factor, Let be the position of the prey in the t-th iteration. Let be the current position of the gray wolf in the t-th iteration. The convergence factor is Let be the current position of the gray wolf in the (t+1)th iteration. and A random number in the range [0,1]. The coefficients decrease linearly from 2 to 0 with the number of iterations; The formula for updating the candidate parameter position is expressed as: in, Let be the distance between the i-th gray wolf and the k-th leader wolf. Let the position of the leader wolf be defined in the t-th iteration. Let i be the position of the i-th ordinary gray wolf in the t-th iteration. Let be the convergence factor of the i-th wolf. Let be the swing factor of the i-th wolf. , and Three candidate positions are calculated for the gray wolf based on the positions of α, β, and δ wolves. This represents the final position of the gray wolf in the (t+1)th iteration.

6. The wind power frequency regulation control method based on VAE-GRU as described in claim 5, characterized in that: The generation of low-dimensional feature vectors includes a variational autoencoder network consisting of multiple fully connected layers. The input layer dimension matches the number of features in high-dimensional data. The intermediate layer undergoes nonlinear transformation through the ReLU activation function. The output layer generates a mathematical expectation vector and a standard deviation vector, respectively. Random vectors are sampled from the standard normal distribution. The expected value vector and the standard deviation vector are combined with the random vectors, and latent variables are generated through element-level operations. The decoding network adopts a symmetric structure, inputs the latent variables into multiple fully connected layers, restores the original input dimension, and outputs reconstructed data. The objective function is to minimize the reconstruction error and the sum of KL divergence. The probabilistic relationship between latent variables and observed data is expressed as follows: in, For joint probability density distribution, The latent variable represents the low-dimensional feature vector extracted by the variational coding network. The observed variables are the original high-dimensional input data. Let be the mathematical expectation vector, and be the mean parameter output by the variational coding network. The standard deviation vector is the variance parameter output by the variational coding network. For vectors It is a diagonal matrix composed of diagonal elements. It follows a multivariate normal distribution. This is an element-wise multiplication of vectors; The weights of the encoding and decoding networks are iteratively optimized using the training sample set until the loss function converges. The high-dimensional data from the training and test sets are then input into the encoding network, which outputs a low-dimensional feature vector. This vector is then combined with the optimal control parameter label to form an enhanced sample set. The loss function is expressed as: in, Let the variational lower bound loss function be... For Kullback-Leibler divergence, This is the joint probability distribution, i.e., the joint distribution of the latent variable z and the observed data x under parameter θ. The variational posterior distribution, i.e., the parameters The approximate distribution of the true posterior distribution fitted by the sub-encoding network. To decode network parameters, To encode network parameters; The establishment of the nonlinear mapping model includes: calculating the reset gate output using the sigmoid activation function, calculating the update gate output using the sigmoid activation function, controlling the fusion ratio of historical information and current information, calculating candidate hidden states using the tanh activation function, fusing the previous hidden state and candidate hidden states through the update gate, outputting the hidden state at the current moment, and inputting the final hidden state into the fully connected layer, mapping it to the control parameter dimension, and outputting the predicted values ​​of over-generation power and over-generation duration.

7. The wind power frequency regulation control method based on VAE-GRU as described in claim 6, characterized in that: The optimal control parameters for the current operating condition include: when the grid frequency drop exceeds a preset threshold, controlling the wind turbine to release rotor kinetic energy and increase electromagnetic power output to provide short-term frequency support based on the predicted over-generation power and over-generation duration; after the over-generation phase ends, restoring power output based on the maximum power point tracking curve, and balancing rotor speed recovery and frequency stability by controlling the recovery rate, and dynamically updating control parameters based on real-time operating condition data.

8. A wind power frequency regulation control system based on VAE-GRU, employing the wind power frequency regulation control method based on VAE-GRU as described in any one of claims 1 to 7, characterized in that, It includes an optimal control parameter sample generation module, a high-dimensional data feature extraction and noise reduction module, a dynamic mapping relationship learning module, and a real-time adaptive control execution module; The optimal control parameter sample generation module is used to establish a frequency response model of the wind power grid-connected system, with the optimization objective of maximizing the minimum system frequency. Using a swarm intelligence optimization algorithm, it performs simulation calculations under different wind speeds, wind power ratios, and load disturbance conditions, automatically finds the optimal SIC control parameters, and outputs the corresponding optimal SIC control parameters. The high-dimensional data feature extraction and noise reduction module is used to utilize the nonlinear mapping capability of the variational autoencoder to map multi-dimensional operating condition data containing noise and redundant information to a low-dimensional, smooth latent feature space, thereby generating a low-dimensional feature vector. The dynamic mapping relationship learning module is used to learn the nonlinear relationship between the low-dimensional feature sequence processed by the variational autoencoder and the optimal SIC parameters by using a gated cyclic unit time-series dynamic modeling network. The update gate and reset gate mechanism inside the gated cyclic unit captures the dynamic dependencies and long-term patterns in the data and predicts the control parameters under different disturbance scenarios. The real-time adaptive control execution module is used to collect real-time status data of the power grid, process it through a feature extraction module and a dynamic mapping learning module, adaptively output the most suitable over-generation power and over-generation duration under the current operating conditions, and send the parameters into the wind power's stepwise inertial control system to dynamically adjust the control curve.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the wind power frequency regulation control method based on VAE-GRU as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the wind power frequency regulation control method based on VAE-GRU as described in any one of claims 1 to 7.

Images