A wind turbine equivalent inertia prediction method

Abstract

This invention relates to the field of wind power generation and power system dynamic stability technology, and particularly to a method for predicting the equivalent inertia of wind turbine generators. The method first collects operating time-series data, preprocesses it, and divides the dataset to construct supervised samples. Next, it uses a temporal convolutional network and multi-head attention to extract multi-scale features. Then, a two-stage fusion mechanism is introduced to effectively fuse the information. Finally, a composite loss and optimization strategy are used to train the model to achieve high-precision prediction of the equivalent inertia. This invention can effectively improve the prediction accuracy and generalization ability of the equivalent inertia of wind turbine generators. The output equivalent inertia prediction results can be directly applied to real-time wind farm scheduling, grid frequency regulation, and ancillary service optimization, providing effective decision support for stable wind power grid connection and reliable power system operation.

Description

Technical Field

[0001] This invention relates to the field of wind power generation and dynamic stability technology of power systems, and in particular to a method for predicting the equivalent inertia of wind turbine generators. Background Technology

[0002] As the installed capacity of wind turbines increases year by year, the inertia level of the power system decreases significantly, leading to a marked decline in its frequency stability capability. Therefore, both domestic and international power grid operation rules require wind turbines to possess inertia response and frequency response capabilities during grid-connected operation. Currently, virtual inertia control strategies for wind turbines can utilize power converters to simulate inertia response to support grid frequency regulation. Accurately obtaining the system inertia is directly related to the implementation of the control strategy and the performance of frequency regulation, and is of great significance to improving the overall reliability of the power grid.

[0003] Existing wind turbine equivalent inertia prediction technologies face numerous unresolved issues. In data processing, traditional methods often neglect outliers and missing data, leading to dataset noise pollution and impacting model robustness. In feature extraction, single convolutional networks struggle to simultaneously capture local temporal dependencies and global long-range relationships, ignoring the multi-scale characteristics of wind power time-series data. Regarding fusion mechanisms, while existing attention models enhance global modeling, they lack effective methods for integrating multi-scale features extracted by temporal convolution and attention mechanisms. In terms of prediction accuracy, traditional models are susceptible to accumulated errors, exhibiting insufficient generalization ability and failing to meet the high precision and low latency requirements of real-time frequency control. Furthermore, existing composite loss functions are often limited to temporal errors, neglecting spectral regularization, leading to overfitting and underfitting under complex operating conditions.

[0004] Therefore, there is an urgent need for a method that can accurately and effectively predict the equivalent inertia of wind turbine units in order to improve the inertial support capacity of the power system and ensure the stability of the power grid frequency. Summary of the Invention

[0005] This invention addresses the shortcomings of existing technologies by proposing a method for predicting the equivalent inertia of wind turbine generators. Specifically, this invention addresses the problem of predicting the equivalent inertia of wind turbine generators in high-penetration wind power scenarios. First, it collects operational time-series data, preprocesses it, and divides the dataset to construct supervised samples. Next, it utilizes a temporal convolutional network and multi-head attention to extract multi-scale features. Then, it introduces a two-stage fusion mechanism to effectively fuse the information. Finally, it employs a composite loss and optimization strategy to train the model, achieving high-precision prediction of the equivalent inertia.

[0006] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0007] A method for predicting the equivalent inertia of a wind turbine generator set employs causal dilated convolution operations to calculate short-term correlation features in the time series, while simultaneously calculating scaled dot product self-attention to obtain long-term correlation features. These features are then dynamically integrated using a time-series gating unit and a frequency-domain adaptive adjustment module to obtain fused feature information. Finally, a fully connected layer is used to accurately estimate the equivalent inertia of the wind turbine generator set. The specific implementation steps include:

[0008] Step S1: Wind turbine operation data acquisition and preprocessing;

[0009] Step S2: Dataset partitioning and normalization;

[0010] Step S3: Construction of supervised learning samples;

[0011] Step S4: Temporal convolutional networks extract local dependency features;

[0012] Step S5: Extract global dependency features using a multi-head attention mechanism;

[0013] Step S6: Temporal gating mechanism achieves preliminary feature fusion;

[0014] Step S7: The spectrum adaptation mechanism achieves the final feature fusion;

[0015] Step S8: Linear mapping outputs the equivalent inertia prediction result;

[0016] Step S9: Model training optimization and performance evaluation;

[0017] Step S10: Deployment and practical application of the prediction model.

[0018] Further, as a preferred technical solution of the present invention, the specific steps of step S1 are as follows: Collect wind turbine operating time series data, including: wind speed, active power, reactive power, rotor speed, pitch angle, voltage amplitude, frequency signal, and power disturbance characteristic data; calculate the equivalent inertia of the wind turbine as label data; obtain a set of operating data for each set of characteristic data collected and the label data calculated once, with a collection scale of 50,000 sets of operating data as the initial dataset, which covers multiple operating scenarios under different weather conditions; preprocess the collected initial dataset, firstly, based on the three-standard-deviation threshold identification method, remove outliers, specifically as follows: if a certain characteristic data x in the initial dataset meets the... In the formula, μ and σ are the mean and standard deviation of the feature, respectively. Then, the running data group containing the feature data x is removed, and the intermittent missing data is filled by linear front and back interpolation to finally obtain the preprocessed dataset.

[0019] Further, as a preferred embodiment of the present invention, step S2 specifically involves: dividing the preprocessed dataset obtained in step S1 into a training set, a validation set, and a test set according to timestamp order, with proportions of 70%, 10%, and 20%, respectively; and performing normalization processing on the training set, validation set, and test set respectively, specifically as follows: performing Min-Max normalization on each feature data and label data in the training set, and using the min(X) normalization of the training set to avoid data leakage. F ) and max(X F The validation set, test set, and normalization are performed, and the calculation process is as follows: ;

[0020] In the formula, X F x represents all values ​​of a certain feature or label data. F A single value in a specific feature or label data. The data points are normalized. Through the above process, we obtain the normalized training set, validation set, and test set, which include 50,000 normalized running data groups, each consisting of normalized feature data and label data.

[0021] Further, as a preferred embodiment of the present invention, step S3 specifically involves: based on the normalized training set, validation set, and test set from step S2, constructing predicted supervised learning samples using a sliding window method. The supervised learning samples include feature sequences and label sequences. The feature sequences consist of τ sets of feature data, expressed as follows: In the formula, the subscript τ is the window length, the subscript t is the time step (t > τ), x is a set of feature data in the normalized training set, validation set, and test set, and X is the feature sequence; the label sequence consists of N sets of label data, represented as: In the formula, the superscript n represents the nth group of label data, and H eq Let y be the normalized equivalent inertia, and y be the label sequence.

[0022] Furthermore, as a preferred embodiment of the present invention, step S4 specifically involves inputting the feature sequence X obtained in step S3 into a temporal convolutional network. This represents the output of the l-th layer of a multi-layer temporal convolutional network, where the 0th layer can be set to... Inter-layer computation first employs a causal dilated convolution operation, represented as... In the formula, This represents a causal dilated convolution operation, where α is the kernel index, A is the kernel size, and w α For learnable weights, d l Let be the dilation factor of layer l, 'a' be the input to the causal dilated convolution operation, and 't' be the time step. Then, using... The activation function introduces nonlinearity. Let b represent the activation function, and b represent the input to the activation function. Then, layer normalization is performed. In the formula, The layer normalization is represented by μ1 and σ1, which are the mean and variance of the input c, respectively, and γ and β are the learnable parameter 1 and learnable parameter 2, respectively. This indicates element-wise multiplication, where ε is a small constant.

[0023] Finally, residual connections are used to combine the normalized output of the temporal convolutional network layer with the input of the previous layer. The complete computation process is as follows: Finally, the output F of the temporal convolutional network can be obtained. TCN .

[0024] As a further preferred embodiment of the present invention, step S5 specifically involves: inputting the feature sequence X constructed in step S4 into the multi-head attention module, and the self-attention mechanism generating a query matrix through linear projection. Key matrix Value matrix In the formula, W Q W K W V These are the projection matrices of the query matrix Q, the key matrix K, and the value matrix V, respectively. The multi-head attention mechanism contains H self-attention mechanisms. For each attention head h, h=1~H, the scaled dot product self-attention head is calculated. h , In the formula, Q h K h V h Let d represent the query matrix, key matrix, and value matrix of the h-th head, respectively. k The dimension of the key matrix is ​​represented by Softmax, which is defined as: In the formula, s i The output is the i-th node, and j represents the traversal index of the denominator summation. The final output is obtained by concatenation and linear projection fusion to obtain the enhanced global dependency feature F. Att , represented as: In the formula, W represents dimensional concatenation. O This is for outputting the projection matrix.

[0025] Furthermore, as a preferred embodiment of the present invention, step S6 specifically involves: the timing gating mechanism first converting the F obtained in step S4... TCN And F obtained in step S5 Att Element-wise gating is performed along the time dimension to obtain the time-dependent gating coefficient G. t : In the formula, Pool represents the average pooling operation, defined as: x pFor average pooling input, W g and b g Here are the learnable weights and biases of the gated network, and sigmoid is the activation function, defined as: x s The input is used as the activation function input; then, a weighted mixture is used to obtain the preliminary fusion representation F. mix : .

[0026] Furthermore, as a preferred embodiment of the present invention, step S7 specifically involves: the spectrum adaptive mechanism first converting the F obtained in step S6... mix Projecting onto the discrete cosine domain yields the spectrum represented as S mix Specifically, it is expressed as: In the formula, k S For S mix Frequency domain index, k F For F mix The time-domain index, K F For F mix The sequence length is determined; then, the adaptive spectral weights are calculated, and the spectrum is weighted by frequency band to obtain the weighted spectrum, specifically represented as follows: In the formula, MLP is a fully connected layer mapping; finally, it is reconstructed back to the time domain and then combined with F. mix Perform residual fusion to obtain the final fused feature information F. fused Specifically, it is expressed as: W Re This represents the projection matrix for the reconstruction operation.

[0027] Furthermore, as a preferred embodiment of the present invention, step S8 specifically involves: processing the fusion feature F obtained in step S7... fused Perform 1×1 convolution, pass through two layers of feedforward network with residuals, and then output an equivalent inertia sequence of length N through linear mapping. In the formula, The equivalent inertia sequence obtained from the calculation, The specific normalized value of the equivalent inertia of length N is obtained; and the corresponding variance vector is also obtained. In the formula It is the variance vector. This represents the specific value of the corresponding variance vector.

[0028] Furthermore, as a preferred embodiment of the present invention, step S9 specifically involves: using the output of step S8. Calculate the composite loss L with the label sequence y n The composite loss of the nth output is expressed as: In the formula, λ mes Let λ be the hyperparameter of the mean square error term. specThe weights are defined as the spectral regularization terms. The gradient is obtained through the loss function, and the model parameters are updated using the AdamW optimizer via backpropagation, thus optimizing the model. A linear warm-up and cosine annealing strategy is employed to dynamically adjust the learning rate. After each round of model parameter optimization, the validation set loss is calculated and performance changes are monitored. If the validation set loss does not significantly improve within 20 consecutive training cycles, an early stopping mechanism is triggered, and the model reverts to the best-performing weighted state. The performance of the best-performing weighted state model is validated using a test set, ultimately yielding a well-trained wind turbine equivalent inertia prediction model.

[0029] As a further preferred technical solution of the present invention, the specific steps of step S10 are as follows: the wind turbine equivalent inertia prediction model trained in step S9 is deployed in the real-time operating environment for wind turbine equivalent inertia prediction; during operation, the same wind turbine operating characteristic data as the model training is collected, and after preprocessing in step S1 and normalization processing in step S2, it is input into the trained wind turbine equivalent inertia prediction model. The model outputs the wind turbine equivalent inertia normalized value, and further inverse normalization can obtain the equivalent inertia prediction value.

[0030] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0031] This invention first collects operational time-series data, preprocesses it, and divides the dataset to construct supervised samples. Next, it utilizes a temporal convolutional network and multi-head attention to extract multi-scale features. Then, a two-stage fusion mechanism is introduced to effectively fuse the information. Finally, a composite loss and optimization strategy are used to train the model, achieving high-precision prediction of equivalent inertia. This invention can effectively improve the prediction accuracy and generalization ability of wind turbine equivalent inertia. Its output equivalent inertia prediction results can be directly applied to real-time wind farm scheduling, grid frequency regulation, and ancillary service optimization, providing effective decision support for stable wind power grid connection and reliable power system operation. Attached Figure Description

[0032] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0033] Figure 1 This is a flowchart illustrating the overall method of the present invention;

[0034] Figure 2 For preprocessing the dataset;

[0035] Figure 3 To supervise the learning sample construction process;

[0036] Figure 4 This is a flowchart of the forward propagation process of the model of this invention;

[0037] Figure 5 This is a comparison chart of the equivalent inertia predictions of the present invention (sunny day scene);

[0038] Figure 6 This is a comparison chart of the equivalent inertia predictions of the present invention (multi-cloud scenario);

[0039] Figure 7 This is a comparison chart of the equivalent inertia prediction of the present invention (rainy scene). Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. Of course, the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0041] like Figure 1 As shown, this invention proposes a method for predicting the equivalent inertia of wind turbines. It employs causal dilated convolution operations to calculate short-term correlation features in the time series, while simultaneously calculating scaled dot product self-attention to obtain long-term correlation features. These features are then dynamically integrated using a temporal gating unit and a frequency domain adaptive adjustment module to obtain fused feature information. Finally, a fully connected layer is used to accurately estimate the equivalent inertia of the wind turbine. The specific implementation steps include: Step S1: Wind turbine operation data acquisition and preprocessing; Step S2: Dataset partitioning and normalization; Step S3: Supervised learning sample construction; Step S4: Temporal convolutional network extracts local dependency features; Step S5: Multi-head attention mechanism extracts global dependency features; Step S6: Temporal gating mechanism achieves preliminary feature fusion; Step S7: Spectrum adaptive mechanism achieves final feature fusion; Step S8: Linear mapping outputs the equivalent inertia prediction result; Step S9: Model training optimization and performance evaluation.

[0042] Step S10: Deployment and practical application of the prediction model.

[0043] Step S1 involves collecting time-series operating data of the wind turbine, including characteristic data such as wind speed, active power, reactive power, rotor speed, pitch angle, voltage amplitude, frequency signal, and power disturbance. The equivalent inertia of the wind turbine is then calculated as label data. Each set of characteristic data collected and label data calculated yields one set of operating data. The collection scale is 50,000 sets of operating data, serving as the initial dataset. The dataset is as follows: Figure 2 As shown, Figure 2 The x-axis represents the timestamp, and the y-axis represents the feature data. There are a total of 9 line charts; the first 8 are line charts of 8 types of feature data, and the last one is a line chart of label data. Figure 2It can be seen that the nine data types exhibit strong randomness and show significant fluctuations overall. This initial dataset covers various operational scenarios under different weather conditions, including sunny, cloudy, and rainy days. The collected initial dataset undergoes preprocessing. First, outliers are identified and removed using a three-standard-deviation threshold method, specifically as follows: If a feature data x in the initial dataset meets... In the formula, μ and σ are the mean and standard deviation of the feature, respectively. Then, the running data group containing the feature data x is removed, and the intermittent missing data is filled by linear front and back interpolation to finally obtain the preprocessed dataset.

[0044] Step S2 involves dividing the preprocessed dataset obtained in Step S1 into training, validation, and test sets according to timestamp order, with proportions of 70%, 10%, and 20%, respectively. Normalization is then performed on the training, validation, and test sets as follows: Min-Max normalization is applied to each feature and label data in the training set. To avoid data leakage, the min(X) normalization of the training set is used. F ) and max(X F The validation set, test set, and normalization are performed, and the calculation process is as follows: In the formula, X F x represents all values ​​of a certain feature or label data. F A single value in a specific feature or label data. These are the normalized data points. Through the above process, we can obtain the normalized training set, validation set, and test set, which include 50,000 normalized running data sets, each consisting of normalized feature data and label data.

[0045] The specific steps of step S3 are as follows: Based on the normalized training set, validation set, and test set from step S2, predictive supervised learning samples are constructed using the sliding window method. The sliding window construction process for supervised learning is as follows: Figure 3 As shown, the supervised learning samples contain feature sequences and label sequences. The feature sequences consist of τ sets of feature data, which can be represented as: In the formula, the subscript τ is the window length, the subscript t is the time step (t > τ), x is a set of feature data in the normalized training set, validation set, and test set, and X is the feature sequence; the label sequence consists of N sets of label data, which can be represented as: In the formula, the superscript n represents the nth group of label data, and H eq Let y be the normalized equivalent inertia, and y be the label sequence.

[0046] Figure 4This is a flowchart of the model's forward propagation, mainly illustrating the process of feature sequences propagating forward within the model. The model primarily consists of three parts: a temporal convolutional network, an attention mechanism, and a two-stage temporal gating-spectral adaptive mechanism. The detailed computational flow of the model is as follows. Step S4 involves inputting the feature sequence X obtained in step S3 into the temporal convolutional network. This represents the output of the l-th layer of a multi-layer temporal convolutional network, where the 0th layer can be set to... Inter-layer computation first employs a causal dilated convolution operation, which can be represented as: In the formula, This represents a causal dilated convolution operation, where α is the kernel index, A is the kernel size, and w α For learnable weights, d l Let be the dilation factor of layer l, 'a' be the input to the causal dilated convolution operation, and 't' be the time step. Then, using... The activation function introduces nonlinearity. Let b represent the activation function, and b represent the input to the activation function. Then, layer normalization is performed. In the formula, The layer normalization is represented by μ1 and σ1, which are the mean and variance of the input c, respectively, and γ and β are the learnable parameter 1 and learnable parameter 2, respectively. This represents element-wise multiplication, where ε is a small constant. Finally, residual connections are used to combine the normalized output of the temporal convolutional network layer with the input of the previous layer. The complete computation process is as follows: Finally, the output F of the temporal convolutional network can be obtained. TCN .

[0047] The specific steps of step S5 are as follows: input the feature sequence X constructed in step S4 into the multi-head attention module, and the self-attention mechanism generates the query matrix through linear projection. Key matrix Value matrix In the formula, W Q W K W V These are the projection matrices of the query matrix Q, the key matrix K, and the value matrix V, respectively. The multi-head attention mechanism contains H self-attention mechanisms. For each attention head h (h=1~H), the scaled dot product self-attention head is calculated. h , In the formula, Q h K h V h Let d represent the query matrix, key matrix, and value matrix of the h-th head, respectively. k The dimension of the key matrix is ​​represented by Softmax, which is defined as: In the formula, s iThe output is the i-th node, and j represents the traversal index of the denominator summation. The final output is obtained by concatenation and linear projection fusion to obtain the enhanced global dependency feature F. Att , can be represented as: In the formula, W represents dimensional concatenation. O This is for outputting the projection matrix.

[0048] The specific steps of step S6 are as follows: The timing gating mechanism first processes the F obtained in step S4... TCN And F obtained in step S5 Att Element-wise gating is performed along the time dimension to obtain the time-dependent gating coefficient G. t : In the formula, Pool represents the average pooling operation, defined as: x p For average pooling input, W g and b g Here are the learnable weights and biases of the gated network, and sigmoid is the activation function, defined as: x s The input is used as the activation function input. Then, a weighted mixture is used to obtain the preliminary fusion representation F. mix : .

[0049] The specific steps of step S7 are as follows: The spectrum adaptive mechanism first uses the F obtained in step S6... mix Projecting onto the discrete cosine domain yields the spectrum represented as S mix Specifically, it can be expressed as: In the formula, k S For S mix Frequency domain index, k F For F mix The time-domain index, K F For F mix The sequence length is determined; then, the adaptive spectral weights are calculated, and the spectrum is weighted by frequency band to obtain the weighted spectrum, which can be specifically expressed as: In the formula, MLP is a fully connected layer mapping; finally, it is reconstructed back to the time domain and then combined with F. mix Perform residual fusion to obtain the final fused feature information F. fused Specifically, it can be expressed as: W Re This represents the projection matrix for the reconstruction operation.

[0050] The specific steps of step S8 are as follows: The fusion feature F obtained in step S7... fused Perform 1×1 convolution, pass through two layers of feedforward network with residuals, and then output an equivalent inertia sequence of length N through linear mapping. In the formula, The equivalent inertia sequence obtained from the calculation, The specific normalized value of the equivalent inertia of length N is obtained; and the corresponding variance vector is also obtained. In the formula It is the variance vector. This represents the specific value of the corresponding variance vector.

[0051] The specific steps of step S9 are as follows: using the output of step S8 Calculate the composite loss L with the label sequence y n The composite loss of the nth output can be expressed as: In the formula, λ mes Let λ be the hyperparameter of the mean square error term. spec The weights are for the spectral regularization term. The gradient is obtained through the loss function, and training uses the AdamW optimizer for parameter updates via backpropagation. This optimizer combines Adam's adaptive learning rate and weight decay to effectively mitigate the gradient explosion / vanishing problem and prevent parameter drift. The optimizer parameters are set as follows: momentum parameters β1=0.9, β2=0.999, and weight decay coefficient w. d =1×10 -4 Gradient updates follow the standard backpropagation process: forward propagation calculation and Backpropagation calculation ∇L n Then execute In the formula, θ represents the model parameters, and m t and v t For AdamW's first and second moment estimation, ε1 is a small constant to accelerate convergence and avoid local optima. This invention employs a two-stage learning rate scheduling: linear preheating and cosine annealing. The learning rate in the linear preheating stage starts from η. min Linear growth to η max The specific growth process is as follows: In the formula, the subscript e is the training loop index, and v e For the e-th training cycle, η e Let V be the learning rate for the e-th training cycle. w The number of warm-up steps is used; the learning rate during the cosine annealing phase decays to η according to a cosine function. min The specific attenuation process is as follows: In the formula, V a This represents the total number of training iterations; to prevent overfitting, the composite loss L is calculated on the validation set after each training iteration. n If L in 20 consecutive training cycles nIf there is no improvement, early stopping is triggered: the model reverts to the optimal weight state, saves the best model through model checkpoints, and terminates training. The performance of the best-performing weight state model is validated using a test set, ultimately obtaining a well-trained wind turbine equivalent inertia prediction model. The method of this invention is compared with four other methods: temporal convolutional networks, Transformers, and a sequential structure of temporal convolutional networks and Transformers. Figure 5 , Figure 6 , Figure 7 The results show the comparison between the actual and predicted values ​​of samples in different scenarios. Figure 5 For sunny scenes, Figure 6 For multi-cloud scenarios, Figure 7 A rainy scene. Figure 5 (a) Figure 6 (a) Figure 7 (a) represents the prediction results of temporal convolutional networks. Figure 5 (b) Figure 6 (b) Figure 7 (b) Both are Transformer prediction results. Figure 5 (c) Figure 6 (c) Figure 7 (c) All are prediction results from a serial Transformer network in the temporal convolutional domain. Figure 5 (d) Figure 6 (d) Figure 7 (d) All figures represent the predicted effects of this invention, with the x-axis representing the timestamp and the y-axis representing H. eq The predicted curves of this invention closely match the actual curves, with the maximum prediction deviation controlled within ±0.06s. The mean absolute error on the test set is 0.0428s, and the root mean square error is 0.0539. The overall trend is accurately captured, demonstrating significantly higher accuracy compared to other models, thus verifying the accuracy of this invention's model in predicting the equivalent inertia of wind turbines.

[0052] The specific steps of step S10 are as follows: Deploy the wind turbine equivalent inertia prediction model trained in step S9 in the real-time operating environment for wind turbine equivalent inertia prediction; during operation, collect the same wind turbine operating characteristic data as the model training, and input the preprocessing in step S1 and the normalization processing in step S2 into the trained wind turbine equivalent inertia prediction model. The model outputs the wind turbine equivalent inertia normalized value, and further inverse normalization can obtain the equivalent inertia prediction value.

[0053] The predictive model designed in this invention provides high-precision decision support for the dynamic control of virtual inertia of wind turbine units, which helps to optimize wind power frequency support strategies and improve the stability and reliable operation of new power systems.

[0054] The specific implementation schemes described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific implementation schemes of the present invention and are not intended to limit the scope of the present invention. Any equivalent changes and modifications made by those skilled in the art without departing from the concept and principles of the present invention should fall within the scope of protection of the present invention.

Claims

1. A method for predicting the equivalent inertia of a wind turbine generator set, characterized in that, Includes the following steps: Step S1: Wind turbine operation data acquisition and preprocessing; Step S2: Dataset partitioning and normalization; Step S3: Construction of supervised learning samples; Step S4: Temporal convolutional networks extract local dependency features; Step S5: Extract global dependency features using a multi-head attention mechanism; Step S6: The temporal gating mechanism achieves preliminary feature fusion; Step S7: The spectrum adaptation mechanism achieves the final feature fusion; Step S8: Linear mapping outputs the equivalent inertia prediction result; Step S9: Model training optimization and performance evaluation; Step S10: Deployment and practical application of the prediction model.

2. The method for predicting the equivalent inertia of a wind turbine generator according to claim 1, characterized in that, The specific steps of step S1 are as follows: Collect wind turbine operating time series data, including wind speed, active power, reactive power, rotor speed, pitch angle, voltage amplitude, frequency signal, and power disturbance characteristic data; calculate the equivalent inertia of the wind turbine as label data. Each time a set of characteristic data is collected and the label data is calculated, a set of operating data is obtained. The collection scale is 50,000 sets of operating data, serving as the initial dataset. This initial dataset covers various operating scenarios under different weather conditions. Preprocess the collected initial dataset. First, an outlier is removed based on a three-standard-deviation threshold identification method, specifically as follows: If a certain characteristic data x in the initial dataset meets the... In the formula, μ and σ are the mean and standard deviation of the feature, respectively. Then, the running data group containing the feature data x is removed, and the intermittent missing data is filled by linear front and back interpolation to finally obtain the preprocessed dataset.

3. The method for predicting the equivalent inertia of a wind turbine generator according to claim 2, characterized in that, The specific steps of step S2 are as follows: Divide the preprocessed dataset obtained in step S1 into a training set, a validation set, and a test set according to timestamp order, with proportions of 70%, 10%, and 20%, respectively; perform normalization processing on the training set, validation set, and test set respectively, specifically as follows: Perform Min-Max normalization on each feature data and label data in the training set. To avoid data leakage, use the min(X) normalization of the training set. F ) and max(X F The validation set, test set, and normalization are performed, and the calculation process is as follows: ; In the formula, X F x represents all values ​​of a certain feature or label data. F A single value in a specific feature or label data. These are the normalized data points; Through the above process, we obtain the normalized training set, validation set, and test set, which include 50,000 normalized running data sets, each consisting of normalized feature data and label data.

4. The method for predicting the equivalent inertia of a wind turbine generator according to claim 3, characterized in that, The specific steps of step S3 are as follows: Based on the normalized training set, validation set, and test set from step S2, a predicted supervised learning sample is constructed using the sliding window method. The supervised learning sample contains a feature sequence and a label sequence. The feature sequence consists of τ sets of feature data, represented as follows: In the formula, the subscript τ is the window length, the subscript t is the time step (t>τ), x is a set of feature data in the normalized training set, validation set and test set, and X is the feature sequence; The label sequence consists of N sets of label data, represented as: In the formula, the superscript n represents the nth group of label data, and H eq Let y be the normalized equivalent inertia, and y be the label sequence.

5. The method for predicting the equivalent inertia of a wind turbine generator according to claim 4, characterized in that, The specific steps of step S4 are as follows: input the feature sequence X obtained in step S3 into the temporal convolutional network. This represents the output of the l-th layer of a multi-layer temporal convolutional network, where the 0th layer can be set to... Inter-layer computation first employs a causal dilated convolution operation, represented as... In the formula, This represents a causal dilated convolution operation, where α is the kernel index, A is the kernel size, and w α For learnable weights, d l Let be the dilation factor of layer l, 'a' be the input to the causal dilated convolution operation, and 't' be the time step. Then, using... The activation function introduces nonlinearity. Let b represent the activation function, and b represent the input to the activation function. Then, layer normalization is performed. In the formula, The layer normalization is represented by μ1 and σ1, which are the mean and variance of the input c, respectively, and γ and β are the learnable parameter 1 and learnable parameter 2, respectively. This indicates element-wise multiplication, where ε is a small constant. Finally, residual connections are used to combine the normalized output of the temporal convolutional network layer with the input of the previous layer. The complete computation process is as follows: Finally, the output F of the temporal convolutional network can be obtained. TCN .

6. The method for predicting the equivalent inertia of a wind turbine generator according to claim 5, characterized in that, The specific steps of step S5 are as follows: input the feature sequence X constructed in step S4 into the multi-head attention module, and the self-attention mechanism generates a query matrix through linear projection. Key matrix Value matrix In the formula, W Q W K W V These are the projection matrices of the query matrix Q, the key matrix K, and the value matrix V, respectively. The multi-head attention mechanism contains H self-attention mechanisms. For each attention head h, h=1~H, the scaled dot product self-attention head is calculated. h , In the formula, Q h K h V h Let d represent the query matrix, key matrix, and value matrix of the h-th head, respectively. k The dimension of the key matrix is ​​represented by Softmax, which is defined as: In the formula, s i The output is the i-th node, and j represents the traversal index of the denominator summation. The final output is obtained by concatenation and linear projection fusion to obtain the enhanced global dependency feature F. Att , represented as: In the formula, W represents dimensional concatenation. O This is for outputting the projection matrix.

7. The method for predicting the equivalent inertia of a wind turbine generator according to claim 6, characterized in that, The specific steps of step S6 are as follows: The timing gating mechanism first processes the F obtained in step S4... TCN And F obtained in step S5 Att Element-wise gating is performed along the time dimension to obtain the time-dependent gating coefficient G. t : In the formula, Pool represents the average pooling operation, defined as: x p For average pooling input, W g and b g Here are the learnable weights and biases of the gated network, and sigmoid is the activation function, defined as: x s The input is used as the activation function input; then, a weighted mixture is used to obtain the preliminary fusion representation F. mix : .

8. The method for predicting the equivalent inertia of a wind turbine generator according to claim 7, characterized in that, The specific steps of step S7 are as follows: The spectrum adaptive mechanism first processes the F obtained in step S6... mix Projecting onto the discrete cosine domain yields the spectrum represented as S mix Specifically, it is expressed as: In the formula, k S For S mix Frequency domain index, k F For F mix The time-domain index, K F For F mix The sequence length is determined; then, the adaptive spectral weights are calculated, and the spectrum is weighted by frequency band to obtain the weighted spectrum, specifically represented as follows: In the formula, MLP is a fully connected layer mapping; finally, it is reconstructed back to the time domain and then combined with F. mix Perform residual fusion to obtain the final fused feature information F. fused Specifically, it is expressed as: W Re This represents the projection matrix for the reconstruction operation.

9. The method for predicting the equivalent inertia of a wind turbine generator according to claim 8, characterized in that, The specific steps of step S8 are as follows: The fusion feature F obtained in step S7 is processed... fused Perform 1×1 convolution, pass through two layers of feedforward network with residuals, and then output an equivalent inertia sequence of length N through linear mapping. In the formula, The equivalent inertia sequence obtained from the calculation, The specific normalized value of the equivalent inertia of length N is obtained; and the corresponding variance vector is also obtained. In the formula It is the variance vector. This represents the specific value of the corresponding variance vector.

10. The method for predicting the equivalent inertia of a wind turbine generator according to claim 9, characterized in that, The specific steps of step S9 are as follows: using the output of step S8. Calculate the composite loss L with the label sequence y n The composite loss of the nth output is expressed as: In the formula, λ mes Let λ be the hyperparameter of the mean square error term. spec The weights are defined as the spectral regularization terms. The gradient is obtained through the loss function, and the model parameters are updated using the AdamW optimizer via backpropagation, thus optimizing the model. A linear warm-up and cosine annealing strategy is employed to dynamically adjust the learning rate. After each round of model parameter optimization, the validation set loss is calculated and performance changes are monitored. If the validation set loss does not significantly improve within 20 consecutive training cycles, an early stopping mechanism is triggered, and the model reverts to the best-performing weighted state. The performance of the best-performing weighted state model is validated using a test set, ultimately yielding a well-trained wind turbine equivalent inertia prediction model.

11. The method for predicting the equivalent inertia of a wind turbine generator according to claim 10, characterized in that, The specific steps of step S10 are as follows: the wind turbine equivalent inertia prediction model trained in step S9 is deployed in the real-time operating environment for wind turbine equivalent inertia prediction; during operation, the same wind turbine operating characteristic data as the model training is collected, and after preprocessing in step S1 and normalization in step S2, it is input into the trained wind turbine equivalent inertia prediction model. The model outputs the wind turbine equivalent inertia normalized value, and further inverse normalization can obtain the equivalent inertia prediction value.

Images