A bo-mvmd-bigru wind power prediction method combined with a wind power converter loss model

Abstract

A BO-MVMD-BiGRU wind power prediction method combined with wind power converter loss model is provided, first, a multivariate variational mode decomposition (MVMD) is proposed to decompose wind speed and power respectively, and a Bayesian optimization algorithm (BO) is used to optimize the MVMD-BiGRU parameters; then, the submodes obtained by decomposition are input into BiGRU for prediction, and the results are superimposed to obtain the preliminary prediction results; finally, a BP neural network is used to construct an equivalent model of converter loss, and the preliminary prediction results are input into the equivalent model for correction to obtain the final prediction results. The simulation results show that the proposed model can comprehensively consider the influence of dynamic change of converter loss, and effectively improve the accuracy of wind power prediction model.

Description

Technical Field

[0001] This invention relates to the field of new energy technology, and in particular to large-scale wind power generation, grid connection technology, and wind turbine power generation prediction technology. Specifically, it relates to a BO-MVMD-BiGRU wind power prediction method that combines a wind power converter loss model. Background Technology

[0002] Vigorously developing renewable energy sources such as wind power is one of the important ways to achieve the national strategic goal of "carbon peaking and carbon neutrality," alleviate energy resource constraints, and reduce environmental pollution. Among these, wind power generation, as the main form of wind energy utilization, holds a key strategic position in the process of energy structure transformation and the construction of a new power system. However, wind power generation has significant intermittency and random fluctuations, and its output power is easily affected by a variety of meteorological and environmental factors, such as drastic fluctuations in wind speed, frequent changes in wind direction, and changes in atmospheric pressure and temperature.

[0003] With the continuous increase in installed wind power capacity and grid connection ratio, the uncertainty and volatility of wind power output have become increasingly prominent in impacting the safe and stable operation of the power system. On the one hand, random fluctuations in wind power increase the demand for peak-shaving power sources and reserve capacity, raising system operating costs. On the other hand, large-scale wind power grid connection may also trigger voltage fluctuations, frequency shifts, and abnormal power flow distribution in local grids, exacerbating the complexity of grid planning, dispatching, and real-time control. Against this backdrop, accurately predicting the power generation of wind turbines at different time scales and reducing the adverse impact of wind power output on grid operation has become a key aspect of wind power grid integration and the construction of high-proportion renewable energy power systems. High-precision wind power forecasting not only helps improve the reliability and economy of grid dispatching plans but also provides important support for optimized operation control of wind farms, flexible resource allocation, and the safe and stable operation of new power systems.

[0004] Against this backdrop, achieving high-precision prediction of wind power generation has become a key technical aspect supporting optimized grid scheduling and operation control. This invention proposes a BO-MVMD-BiGRU wind power prediction model that incorporates a wind power converter loss model. First, a Multivariate Variational Mode Decomposition (MVMD) method is proposed to decompose wind speed and power separately, and Bayesian Optimization (BO) is used to optimize the MVMD-BiGRU parameters. Then, the decomposed sub-modes are input into the BiGRU for prediction, and the results are superimposed to obtain a preliminary prediction. Finally, a backpropagation (BP) neural network is used to construct an equivalent converter loss model, and the preliminary prediction results are input into the equivalent model for correction to obtain the final prediction result. Summary of the Invention

[0005] This invention addresses two prominent problems commonly found in existing wind power prediction methods: first, insufficient consideration of the loss mechanism of wind power converters, failing to reflect the impact of dynamic changes in converter efficiency on the actual output of the unit; and second, inadequate decomposition and processing of time-series data such as wind speed and power, making it difficult to fully characterize the fluctuation characteristics of wind power. Therefore, this invention proposes a BO method that combines a wind power converter loss model. MVMD BiGRU wind power prediction method.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A BO-MVMD-BiGRU wind power prediction method combining a wind power converter loss model includes the following steps: Step 1: Perform preprocessing on historical operating data and meteorological data of wind turbines during operation to obtain a preprocessed dataset. Use Pearson correlation analysis to perform feature correlation analysis on the measured meteorological data to reduce the feature dimensionality of the data. Step 2: Use the multivariate variational mode decomposition algorithm to perform multimode decomposition on the wind power and wind speed time series signals in the preprocessed dataset of Step 1, decompose the wind power sequence and wind speed variable into several intrinsic mode components, so as to achieve separation of characteristics at different time scales and frequency bands. Step 3: Construct a bidirectional gated recurrent unit (BiGRU) neural network prediction sub-model for each intrinsic mode component in Step 2. Use the BiGRU neural network to extract temporal features simultaneously along the forward and backward time directions to obtain the predicted values ​​of each mode component. Step 4: Jointly optimize the decomposition parameters of the multivariate variational mode decomposition algorithm in Step 2, as well as the structural parameters and training hyperparameters of the BiGRU neural network in Step 3, based on the Bayesian optimization algorithm; Step 5: Utilize the MVMD optimized by the Bayesian optimization algorithm in Step 4 The BiGRU model predicts the preprocessed data to obtain the prediction results of each modal component, and reconstructs the prediction results of each modal component to obtain the predicted wind power of the wind turbine. Step 6: Build a simulation circuit model of the wind power converter based on the MATLAB platform; construct a BP neural network model for network training to establish a nonlinear mapping model between converter power and efficiency; Step 7: Use the wind power prediction value from Step 5 as the input to the BP neural network in Step 6. By performing real-time efficiency correction on the predicted power, a wind power value closer to the actual output of the unit can be obtained, improving the accuracy of the wind power prediction model and providing more accurate results for subsequent system optimization configuration.

[0007] It also includes step 8: calculating prediction error indices to evaluate model performance, including mean square error (MSE), root mean square error (RMSE), and coefficient of determination (R²).

[0008] In step 1, the operational data includes wind speed, wind direction, ambient temperature, and atmospheric pressure; preprocessing operations include data cleaning, outlier removal, missing value imputation, time-series alignment, and normalization / standardization; step 1 includes the following steps: Step 1-1) Clean the raw wind power data to remove extreme outliers caused by equipment malfunctions or sudden behavior. Apply the interquartile range (IQR) clipping method to detect and remove outliers, effectively ensuring the rationality of data distribution. Steps 1-2) Divide the preprocessed dataset into training, validation, and test sets according to a certain ratio; Steps 1-3) perform data type unification processing, including standardization of time fields, encoding conversion of categorical variables, and simplification of data types for numerical variables.

[0009] Step 2 includes the following steps: Step 2-1) Using the multivariable variational mode decomposition method, the wind power and wind speed sequences are combined and decomposed into several intrinsic mode functions with finite bandwidth under the same framework; Step 2-2) Randomly initialize each modal component, center frequency, and Lagrange multiplier; determine the center frequency and bandwidth of each component through continuous iteration, and introduce Lagrange multipliers. and secondary penalty items Solve the problem; Steps 2-3) Calculate the error of the modal components in adjacent iterations, determine whether the termination iteration condition has been met, and output the modal components with finite bandwidth and their corresponding center frequencies.

[0010] In step 2-1), the model obtained after decomposition is: (1); In the formula: For the decomposition of the first k One modal component; The corresponding center frequency; For Dirac functions; This is the initial wind power or wind speed signal. K To preset the number of modal components, For the constructed analytic signal, For time derivative.

[0011] In step 2-2), Lagrange multipliers are introduced. and secondary penalty items Solve the equation as shown in equation (2): (2); in, Let Lagrange multipliers be the functions of the Lagrange multipliers. It represents the amplitude change of a certain frequency component.

[0012] In steps 2-3), the iteration uses the alternating direction multiplier method, as shown in equation (3): (3); In the formula: , and They are respectively , and Fourier transform; For frequency, n This represents the number of iterations.

[0013] In step 5, the specific steps are as follows: Variational mode decomposition is applied to the wind speed sequence and wind power sequence respectively to decompose the non-stationary signal into several intrinsic mode functions to characterize the power fluctuation characteristics at different time scales; on this basis, BO is introduced to jointly optimize the key hyperparameters of MVMD and BiGRU, and BiGRU prediction sub-models are constructed for each decomposed mode component, and the optimal parameters obtained by BO are used to complete the training; finally, the prediction results of each mode component are superimposed point by point on the power scale to reconstruct the power prediction value of the wind turbine, thereby achieving high-precision prediction of wind power generation.

[0014] Step 6 includes the following steps: Step 6-1: Build a simulation circuit model of the wind power converter based on the MATLAB platform, and set the characteristic parameters of each power device according to the above loss analysis results to accurately simulate the actual loss characteristics of the circuit. Step 6-2: Configure the simulation solver parameters and iteration step size, perform simulation calculations under different input power conditions, and obtain the input power and output efficiency dataset of the converter; Step 6-3: Construct a BP neural network model, using power data obtained from simulation as input and efficiency data as output for network training, thereby establishing a nonlinear mapping model between converter power and efficiency.

[0015] Step 7 specifically includes: the BP neural network architecture employs five hidden layers, with the number of neurons in each hidden layer being 80, 60, 40, 20, and 10 respectively, forming a deep feedforward network that "goes from wide to narrow." The first three layers are wider to fully characterize the complex nonlinear mapping relationship between input wind power and efficiency, while the last two layers gradually shrink to achieve feature dimensionality reduction, thereby improving the model's generalization ability. By coupling the BO-MVMD-BiGRU wind power prediction model with the wind power converter efficiency model, real-time efficiency correction can be applied to the predicted power, resulting in wind power output that is closer to the actual output of the unit.

[0016] Step 3 specifically includes: The Bidirectional Gated Recurrent Unit (BiGRU) builds upon this foundation by introducing a reverse recurrent structure. It consists of two GRU layers, forward and backward, encoding the input sequence from the forward and reverse directions respectively. The hidden state vectors from both directions are concatenated as the final output, thus capturing the global temporal features of the sequence more comprehensively. The principle of BiGRU is expressed as follows: (4); In the formula: w t and v t They are respectively t The output weights of the front and back hidden layers of the time series at each time step; and They are t- The output states of the forward and backward GRUs at time 1.

[0017] Step 4 specifically includes: The core of the Bayesian optimization algorithm consists of two parts: Gaussian process regression and a data acquisition function. A surrogate model is constructed using randomly acquired data points, and the local optima of the model are found based on the data acquisition function. The surrogate model is essentially an approximation of the actual function. The algorithm automatically searches for hyperparameter combinations with the objective function of minimizing the mean absolute percentage error, and selects the optimal hyperparameters based on a comprehensive evaluation index on the validation set.

[0018] Step 8 specifically includes: To verify that the power and wind speed sequences decomposed by MVMD can effectively aid model training and improve prediction accuracy, and that the converter efficiency model can further enhance the accuracy of the results, this paper employs mean absolute percentage error (MAPE), root mean square error (RMSE), and the coefficient of determination. R2 To evaluate the model's performance, the expression is as follows: (5); (6); (7); In the formula: n The total number of samples, This is the actual value. For predicted values, This is the average of the actual values.

[0019] Compared with the prior art, the present invention has the following technical effects: 1) Based on Pearson correlation analysis, meteorological features with strong correlation to wind power were screened out, which effectively reduced the dimensionality of input features and reduced the modeling and computational complexity brought about by large-scale wind power data while retaining key information. 2) The MVMD method is used to decompose the wind power signal and wind speed signal into a combination, which decomposes the non-stationary original sequence into several finite bandwidth modes, effectively reducing the non-stationarity and complexity of the data. 3) By jointly optimizing the key hyperparameters of MVMD and BiGRU using BO, the optimal parameter combination was obtained, resulting in the BO-MVMD-BiGRU model achieving MAPE = 2.436% and RMSE = 0.0458 on the test set. R 2 =0.9979, which significantly improves the accuracy of wind power prediction; 4) Based on high-precision wind power prediction, the BO-MVMD-BiGRU model is coupled with the wind power converter efficiency model to form an overall wind power prediction system, which makes the prediction results closer to the actual wind power and can provide more accurate and effective support for subsequent system optimization. Attached Figure Description

[0020] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a system flowchart of the present invention; Figure 2 This is a variational mode decomposition diagram of wind power output. Figure 3 It is a variational mode decomposition wind speed diagram; Figure 4 This is a schematic diagram of a bidirectional gated recurrent unit neural network structure; Figure 5 This is a flowchart illustrating the principle of the Bayesian optimization algorithm; Figure 6It is a solution for improving the efficiency of wind power converters; Figure 7 This is a diagram of a BP neural network structure; Figure 8 This is a graph showing the wind power prediction results in an example of the present invention; Figure 9 This is a corrected diagram of wind power prediction results in an example of the present invention. Detailed Implementation

[0021] like Figure 1 As shown, a BO-MVMD-BiGRU wind power prediction method combining a wind power converter loss model includes the following steps: Step 1: Perform preprocessing on historical operating data and meteorological data of wind turbines during operation to obtain a preprocessed dataset. Use Pearson correlation analysis to perform feature correlation analysis on the measured meteorological data to reduce the feature dimensionality of the data. Step 2: Use the multivariate variational mode decomposition algorithm to perform multimode decomposition on the wind power and wind speed time series signals in the preprocessed dataset of Step 1, decompose the wind power sequence and wind speed variable into several intrinsic mode components, so as to achieve separation of characteristics at different time scales and frequency bands. Step 3: Construct a bidirectional gated recurrent unit (BiGRU) neural network prediction sub-model for each intrinsic mode component in Step 2. Use the BiGRU neural network to extract temporal features simultaneously along the forward and backward time directions to obtain the predicted values ​​of each mode component. Step 4: Jointly optimize the decomposition parameters of the multivariate variational mode decomposition algorithm in Step 2, as well as the structural parameters and training hyperparameters of the BiGRU neural network in Step 3, based on the Bayesian optimization algorithm; Step 5: Utilize the MVMD optimized by the Bayesian optimization algorithm in Step 4 The BiGRU model predicts the preprocessed data to obtain the prediction results of each modal component, and reconstructs the prediction results of each modal component to obtain the predicted wind power of the wind turbine. Step 6: Build a simulation circuit model of the wind power converter based on the MATLAB platform; construct a BP neural network model for network training to establish a nonlinear mapping model between converter power and efficiency; Step 7: Use the wind power prediction value from Step 5 as the input to the BP neural network in Step 6. By performing real-time efficiency correction on the predicted power, a wind power value closer to the actual output of the unit can be obtained, improving the accuracy of the wind power prediction model and providing more accurate results for subsequent system optimization configuration.

[0022] It also includes step 8: calculating prediction error indices to evaluate model performance, including mean square error (MSE), root mean square error (RMSE), and coefficient of determination (R²).

[0023] In step 1, the operational data includes wind speed, wind direction, ambient temperature, and atmospheric pressure; preprocessing operations include data cleaning, outlier removal, missing value imputation, time-series alignment, and normalization / standardization; step 1 includes the following steps: Step 1-1) Clean the raw wind power data to remove extreme outliers caused by equipment malfunctions or sudden behavior. Apply the interquartile range (IQR) clipping method to detect and remove outliers, effectively ensuring the rationality of data distribution. Steps 1-2) Divide the preprocessed dataset into training, validation, and test sets according to a certain ratio; Steps 1-3) perform data type unification processing, including standardization of time fields, encoding conversion of categorical variables, and simplification of data types for numerical variables.

[0024] Step 2 includes the following steps: Step 2-1) Using a multivariable variational mode decomposition method, the wind power and wind speed sequences are combined and decomposed into several intrinsic mode functions with finite bandwidth within the same framework, such as... Figure 2 , Figure 3 As shown; Step 2-2) Randomly initialize each modal component, center frequency, and Lagrange multiplier; determine the center frequency and bandwidth of each component through continuous iteration, and introduce Lagrange multipliers. and secondary penalty items Solve the problem; Steps 2-3) Calculate the error of the modal components in adjacent iterations, determine whether the termination iteration condition has been met, and output the modal components with finite bandwidth and their corresponding center frequencies.

[0025] In step 2-1), the model obtained after decomposition is: (1); In the formula: For the decomposition of the first k One modal component; The corresponding center frequency; For Dirac functions; This refers to the initial wind power signal or wind speed. K To preset the number of modal components, For the constructed analytic signal, For time derivative.

[0026] In step 2-2), Lagrange multipliers are introduced. and secondary penalty items Solve the equation as shown in equation (2): (2); in, Let Lagrange multipliers be the functions of the Lagrange multipliers. It represents the amplitude change of a certain frequency component.

[0027] In steps 2-3), the iteration uses the alternating direction multiplier method, as shown in equation (3): (3); In the formula: , and They are respectively , and Fourier transform; For frequency, n This represents the number of iterations.

[0028] In step 5, the specific steps are as follows: Variational mode decomposition is applied to the wind speed sequence and wind power sequence respectively to decompose the non-stationary signal into several intrinsic mode functions to characterize the power fluctuation characteristics at different time scales; on this basis, BO is introduced to jointly optimize the key hyperparameters of MVMD and BiGRU, and BiGRU prediction sub-models are constructed for each decomposed mode component, and the optimal parameters obtained by BO are used to complete the training; finally, the prediction results of each mode component are superimposed point by point on the power scale to reconstruct the power prediction value of the wind turbine, thereby achieving high-precision prediction of wind power generation.

[0029] like Figure 6 As shown, step 6 includes the following steps: Step 6-1: Build a simulation circuit model of the wind power converter based on the MATLAB platform, and set the characteristic parameters of each power device according to the above loss analysis results to accurately simulate the actual loss characteristics of the circuit. Step 6-2: Configure the simulation solver parameters and iteration step size, perform simulation calculations under different input power conditions, and obtain the input power and output efficiency dataset of the converter; Step 6-3: Construct a BP neural network model, using power data obtained from simulation as input and efficiency data as output for network training, thereby establishing a nonlinear mapping model between converter power and efficiency.

[0030] Step 7 specifically includes: the BP neural network architecture uses five hidden layers, with the number of neurons in each hidden layer being 80, 60, 40, 20, and 10 respectively, forming a deep feedforward network that "goes from wide to narrow." The first three layers are wider to fully characterize the complex nonlinear mapping relationship between input wind power and efficiency, while the last two layers gradually shrink to achieve feature dimensionality reduction, thereby improving the model's generalization ability. Figure 7 As shown, by coupling the BO-MVMD-BiGRU wind power prediction model with the wind power converter efficiency model, the predicted power can be corrected in real time to obtain wind power that is closer to the actual output of the unit.

[0031] Step 3 specifically includes: The Bidirectional Gated Recurrent Unit (BiGRU) builds upon this by introducing a reverse recurrent structure, consisting of two GRU layers: a forward layer and a backward layer. These layers encode the input sequence from the forward and reverse directions respectively, and concatenate the hidden state vectors from both directions as the final output, thus capturing the global temporal features of the sequence more fully. For example... Figure 4 As shown, the principle expression of BiGRU is: (4); In the formula: w t and v t They are respectively t The output weights of the front and back hidden layers of the time series at each time step; and They are t- The output states of the forward and backward GRUs at time 1.

[0032] Step 4 specifically includes: The core of the Bayesian optimization algorithm consists of two parts: Gaussian process regression and a data acquisition function. A surrogate model is constructed using randomly acquired data points, and the local optimum of the model is found based on the data acquisition function. The surrogate model is essentially an approximation of the actual function. The algorithm uses minimizing the mean absolute percentage error as the objective function, automatically searches for hyperparameter combinations, and selects the optimal hyperparameters based on comprehensive evaluation metrics on the validation set. Figure 5 As shown.

[0033] Step 8 specifically includes: To verify that the power and wind speed sequences decomposed by MVMD can effectively aid model training and improve prediction accuracy, the prediction results are compared with those of... Figure 8 As shown, and by using a converter efficiency model, the results can be made more accurate, such as... Figure 9As shown. This paper uses mean absolute percent error (MAPE), root mean square error (RMSE), and coefficient of determination. R 2 To evaluate the model's performance, the expression is as follows: (5); (6); (7); In the formula: n The total number of samples, This is the actual value. For predicted values, This is the average of the actual values.

[0034] Example: The prediction models all employ mean absolute percentage error (MAPE), root mean square error (RMSE), and coefficient of determination. R 2 The model's performance was evaluated. Table 1 shows a comparison of the model results, serving as metrics for evaluating model accuracy and performance.

[0035] Table 1 Comparison of Model Prediction Results Indicators

[0036] This invention proposes a BO model that combines a wind power converter loss model. MVMD The BiGRU wind power prediction method first cleans, removes outliers, and completes missing data on wind turbine power, wind speed, and related meteorological and operational data. A high-quality sample set is then constructed through time-series alignment, normalization, and standardization to provide a reliable data foundation for subsequent modeling. Second, multivariate variational mode decomposition (MVMD) is used to decompose the wind speed and power sequences at multiple scales within the same framework. This decomposes the non-stationary signal into several intrinsic mode components with finite bandwidth, thus fully characterizing the fluctuation features at different time scales and reducing the complexity and noise interference of the original sequence. Subsequently, a BiGRU prediction sub-model is established for each mode component. A bidirectional cyclic structure is used to simultaneously extract forward and backward time-series dependencies, achieving a more complete model of global dynamic features. The prediction results of each sub-model are then superimposed and reconstructed point by point to obtain the preliminary power prediction value. Furthermore, Bayesian optimization is introduced to jointly optimize the MVMD decomposition parameters and BiGRU model hyperparameters. With the goal of minimizing MAPE, a better parameter combination is automatically searched, significantly reducing the uncertainty caused by empirical parameter tuning and improving the model's repeatability and generalization ability. Compared to existing methods, the key advantage of this patent lies in explicitly incorporating the dynamic impact of wind power converter losses: A converter simulation circuit is built using the MATLAB platform, generating an "input power-output efficiency" dataset. A BP neural network is trained to establish an efficiency equivalent model, thereby implementing real-time efficiency corrections to the initial predicted power, making the prediction results closer to the actual grid-connected output of the unit. Finally, MAPE, RMSE, and the coefficient of determination R are used... 2 The method is evaluated comprehensively based on indicators such as [list of indicators] to verify its advantages in improving prediction accuracy, enhancing engineering applicability, and supporting power grid dispatching decisions.

Claims

1. A BO-MVMD-BiGRU wind power prediction method combining a wind power converter loss model, characterized in that, Includes the following steps: Step 1: Perform preprocessing operations on historical operating data and meteorological data of wind turbines during operation to obtain a preprocessed dataset. Then, use Pearson correlation analysis to perform feature correlation analysis on the measured meteorological data to reduce the feature dimensionality of the data. Step 2: Use the multivariate variational mode decomposition algorithm to perform multimode decomposition on the wind power and wind speed time series signals in the preprocessed dataset of Step 1, and decompose the wind power sequence and wind speed variables into several intrinsic mode components. Step 3: Construct a bidirectional gated recurrent unit (BiGRU) neural network prediction sub-model for each intrinsic mode component in Step 2. Use the BiGRU neural network to extract temporal features simultaneously along the forward and backward time directions to obtain the predicted values ​​of each mode component. Step 4: Jointly optimize the decomposition parameters of the multivariate variational mode decomposition algorithm in Step 2, as well as the structural parameters and training hyperparameters of the BiGRU neural network in Step 3, based on the Bayesian optimization algorithm; Step 5: Utilize the MVMD optimized by the Bayesian optimization algorithm in Step 4 The BiGRU model predicts the preprocessed data to obtain the prediction results of each modal component, and reconstructs the prediction results of each modal component to obtain the predicted wind power of the wind turbine. Step 6: Build a simulation circuit model of the wind power converter based on the MATLAB platform; construct a BP neural network model for network training to establish a nonlinear mapping model between converter power and efficiency; Step 7: Use the wind power prediction value from Step 5 as the input to the BP neural network in Step 6. By performing real-time efficiency correction on the predicted power, a wind power value closer to the actual output of the unit can be obtained, providing a more accurate result for subsequent system optimization configuration.

2. The method according to claim 1, characterized in that, It also includes step 8: calculating prediction error indices to evaluate model performance, including mean square error (MSE), root mean square error (RMSE), and coefficient of determination (R²).

3. The method according to claim 1, characterized in that, In step 1, the running data includes wind speed, wind direction, ambient temperature, and atmospheric pressure; the preprocessing operations include data cleaning, outlier removal, missing value imputation, time-series alignment, and normalization and standardization. Step 1 includes the following steps: Step 1-1) Clean the raw wind power data to remove extreme outliers caused by equipment malfunctions or sudden behavior. Apply the interquartile range (IQR) clipping method to detect and remove outliers, effectively ensuring the rationality of data distribution. Steps 1-2) Divide the preprocessed dataset into training, validation, and test sets according to a certain ratio; Steps 1-3) Perform data type unification processing, including standardization of time fields, encoding conversion of categorical variables, and simplification of data types for numerical variables.

4. The method according to claim 1, characterized in that, Step 2 includes the following steps: Step 2-1) Using the multivariable variational mode decomposition method, the wind power and wind speed sequences are combined and decomposed into several intrinsic mode functions with finite bandwidth under the same framework; Step 2-2) Randomly initialize each modal component, center frequency, and Lagrange multiplier; determine the center frequency and bandwidth of each component through continuous iteration, and introduce Lagrange multipliers. and secondary penalty items Solve the problem; Steps 2-3) Calculate the error of the modal components in adjacent iterations, determine whether the termination condition has been met, and output the modal components with finite bandwidth and their corresponding center frequencies.

5. The method according to claim 4, characterized in that, In step 2-1), the model obtained after decomposition is: （1）； In the formula: For the decomposition of the first k One modal component; The corresponding center frequency; For Dirac functions; This is the initial wind power or wind speed signal. K To preset the number of modal components, For the constructed analytic signal, For time derivative.

6. The method according to claim 5, characterized in that, In step 2-2), Lagrange multipliers are introduced. and secondary penalty items Solve the equation as shown in equation (2): （2）； in, Let Lagrange multipliers be the functions of the Lagrange multipliers. It represents the amplitude change of a certain frequency component.

7. The method according to claim 5, characterized in that, In steps 2-3), the iteration uses the alternating direction multiplier method, as shown in equation (3): （3）； In the formula: , and They are respectively , and Fourier transform; For frequency, n This represents the number of iterations.

8. The method according to any one of claims 1 to 7, characterized in that, In step 5, the specific steps are as follows: Variational mode decomposition is applied to the wind speed sequence and wind power sequence respectively to decompose the non-stationary signal into several intrinsic mode functions to characterize the power fluctuation characteristics at different time scales; on this basis, BO is introduced to jointly optimize the key hyperparameters of MVMD-BiGRU, and BiGRU prediction sub-models are constructed for each decomposed mode component, and the optimal parameters obtained by BO are used to complete the training; finally, the prediction results of each mode component are superimposed point by point on the power scale to reconstruct the power prediction value of the wind turbine, thereby achieving high-precision prediction of wind power generation.

9. The method according to claim 1, characterized in that, Step 6 includes the following steps: Step 6-1: Build a simulation circuit model of the wind power converter based on the MATLAB platform, and set the characteristic parameters of each power device according to the above loss analysis results to accurately simulate the actual loss characteristics of the circuit. Step 6-2: Configure the simulation solver parameters and iteration step size, perform simulation calculations under different input power conditions, and obtain the input power and output efficiency dataset of the converter; Step 6-3: Construct a BP neural network model, using power data obtained from simulation as input and efficiency data as output for network training, thereby establishing a nonlinear mapping model between converter power and efficiency.

10. The method according to claim 1, characterized in that, Step 7 specifically includes: the BP neural network architecture adopts five hidden layers, with the number of neurons in the hidden layers being 80, 60, 40, 20, and 10 respectively, forming a deep feedforward network that is "wide to narrow". The first three layers are relatively wide to fully characterize the complex nonlinear mapping relationship between the input wind power and efficiency, while the last two layers gradually shrink to achieve feature dimensionality reduction, thereby improving the generalization ability of the model; by coupling the BO-MVMD-BiGRU wind power prediction model with the wind power converter efficiency model, the predicted power is corrected in real time for efficiency, which can obtain wind power that is closer to the actual output of the unit.

Images