A wind power prediction method

Abstract

The present invention provides a wind power prediction method, comprising: obtaining an actual wind power data set, and then obtaining a processed data set; performing complete empirical mode decomposition of adaptive noise on the processed data set, and decomposing the data into a number of intrinsic mode function components; using a Bayesian optimization algorithm to perform a global search on the hyperparameters of a Mamba model based on physical constraints, and obtaining an optimal hyperparameter; using the intrinsic mode function components as input features of the Mamba model based on physical constraints, and combining the optimal hyperparameters, integrating them into the Mamba model based on physical constraints for training, and obtaining a trained Mamba model based on physical constraints; and using the trained Mamba model based on physical constraints to predict wind power. The present invention can improve the accuracy and stability of wind power prediction, thereby reducing the frequent adjustment problem of the MPPT system caused by rapid changes in wind speed.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a wind power prediction method. Background Art

[0002] With the rapid development of wind power generation, the accuracy of wind power forecasting is crucial to grid dispatching and system stability. Wind power is affected by meteorological conditions and equipment status, and has significant volatility and uncertainty. Relying only on meteorological factors extracted from the mathematical formula of wind power output as forecast input features often leads to reduced interpretability of the model, affecting the accuracy and stability of wind power forecasting.

[0003] In addition, one of the goals of wind power forecasting is to reduce unnecessary energy losses and improve the output power stability of wind farms. Accurate forecast values ​​can be used as input to help optimize the operation of the Maximum Power Point Tracking (MPPT) controller. The predicted value of wind power reflects the power trend of the future wind power generation system and can provide forward-looking data for the control strategy, thereby improving the accuracy and stability of MPPT. However, power fluctuations can cause the MPPT algorithm to frequently adjust its operating point, especially in the case of large wind speed fluctuations. If the power fluctuations are frequent and large, the MPPT may over-respond, resulting in reduced system efficiency or even damage to the equipment. Summary of the invention

[0004] The object of the present invention is to provide a wind power prediction method to improve the accuracy and stability of wind power prediction, thereby reducing the frequent adjustment problem of the MPPT system caused by rapid changes in wind speed.

[0005] A wind power prediction method comprises the following steps:

[0006] Step S1, obtaining an actual wind power data set, using the Pearson coefficient method to calculate the correlation coefficient between each input feature in the actual wind power data set and the wind power, and then obtaining the target input feature, and then performing Z-score normalization processing on the target input feature to obtain a processed data set;

[0007] Step S2, performing complete empirical mode decomposition of adaptive noise on the processed data set to decompose the data into several intrinsic mode function components;

[0008] Step S3, using a Bayesian optimization algorithm to perform a global search on the hyperparameters of the Mamba model based on physical constraints to obtain optimal hyperparameters, wherein the Mamba model based on physical constraints is based on the Mamba model and uses a second-order Jensen wake model as a physical constraint;

[0009] Step S4, using the intrinsic mode function components obtained in step S2 as input features of the Mamba model based on physical constraints, and integrating them with the optimal hyperparameters obtained in step S3 into the Mamba model based on physical constraints for training, thereby obtaining a trained Mamba model based on physical constraints;

[0010] Step S5: Use the trained Mamba model based on physical constraints to predict wind power.

[0011] The wind power prediction method provided by the present invention has the following beneficial effects:

[0012] (1) The present invention uses the complete empirical mode decomposition with adaptive noise (CEEMDAN) to decompose the wind power data set to obtain several intrinsic mode function (IMF) components, thereby effectively removing the noise in the data in the frequency domain and enhancing the regularity of the input features, so that the model can still maintain a high prediction accuracy on the data set with large noise;

[0013] (2) The present invention uses the Bayesian optimization (BO) algorithm to perform a global search for the hyperparameters of the Mamba model, uses the Gaussian process (GP) as a proxy model to approximate the objective function, and combines the expected improvement (EI) acquisition function to select the optimal hyperparameters, which can effectively improve the efficiency and accuracy of model training. The Bayesian optimization algorithm can dynamically adjust the search direction according to the update of the proxy model, so as to quickly find the optimal hyperparameters and improve the adaptability of the Mamba model to different wind farm data. In addition, the Mamba model adopts a state space model mechanism, which has higher computational efficiency and feature extraction capabilities than the traditional Transformer model. The Mamba model uses a selective state propagation and information forgetting mechanism to enable it to more effectively capture key information in long time series prediction, reduce information redundancy, and thus improve the overall prediction performance of the model;

[0014] (3) The present invention uses the second-order Jensen wake model as a physical constraint and models the physical wake effect of the wind farm through the second-order Jensen wake model. It can effectively describe the mutual influence between the turbines in the wind farm and calculate the power prediction value of the second-order Jensen wake physical model, thus realizing prediction based on real physical phenomena. Moreover, this physical constraint is embedded into the total loss function through the regularization term, which enhances the generalization ability in complex scenarios, reduces the deviation between the Mamba model prediction value and the physical reality, makes the prediction result more in line with the physical law, effectively improves the interpretability and reliability of the model, and improves the accuracy, stability and robustness of the prediction.

[0015] (4) By adopting the method of the present invention, it is possible to make adjustments in advance according to the predicted power changes and determine the timing and amplitude of the maximum power point tracking controller adjustment, thereby effectively reducing the frequent adjustment problem of the MPPT system caused by severe wind speed fluctuations and avoiding efficiency losses due to excessive adjustment or delayed adjustment. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a schematic diagram of the flow chart of the wind power prediction method of the present invention. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0018] See also Figure 1 The wind power prediction method provided by the embodiment of the present invention includes steps S1 to S5:

[0019] Step S1, obtaining an actual wind power data set, using the Pearson coefficient method to calculate the correlation coefficient between each input feature in the actual wind power data set and wind power, and then obtaining the target input feature, and then performing Z-score normalization processing on the target input feature to obtain a processed data set.

[0020] In this embodiment, the experimental object is an actual wind power data set of a wind farm in Xinjiang, China.

[0021] In step S1, the following formula is used to calculate the correlation coefficient between each input feature and wind power in the actual wind power data set:

[0022] ;

[0023] in, m 1 represents the sample value of the first input feature in the actual wind power dataset, represents the sample mean of the first input feature, m 2 represents the wind power corresponding to the first input feature, represents the mean wind power corresponding to the first input feature, Represents the correlation coefficient corresponding to the first input feature;

[0024] The input features with correlation coefficients greater than the threshold are taken as target input features.

[0025] To eliminate the impact of different dimensions on data analysis results and prevent features with larger dimensions from dominating the analysis results. The present invention adopts Z-score normalization. The calculation process of Z-score normalization is relatively simple. It only needs to subtract the mean and divide by the standard deviation for each data point in the data set to obtain a standardized result. This calculation method is often suitable for processing large-scale data sets such as wind power, making computer operation convenient and easy to implement, and the screened strong correlation features are normalized by Z-score. Specifically, the following formula is used to perform Z-score normalization on the target input features:

[0026] ;

[0027] in, o is the original value, is the standardized value, is the mean, is the standard deviation.

[0028] Step S2, performing complete empirical mode decomposition of adaptive noise on the processed data set, and decomposing the data into a number of intrinsic mode function components.

[0029] Among them, in step S2, the complete empirical mode decomposition of adaptive noise (CEEMDAN) is to add white noise to the original signal through multiple iterations, and decompose it successively to obtain the intrinsic mode function (IMF). In wind power prediction, the CEEMDAN algorithm can be used to decompose the classified data in the frequency domain into sub-modes of different frequency bands, reduce the volatility between data in different frequency bands, enhance the regularity between wind power data, and help improve the accuracy of prediction.

[0030] Step S2 specifically includes:

[0031] Step S201, adding the wind power data sequence in the processed data set H Substandard normal white noise, we get H A wind power data sequence with white noise added;

[0032] The purpose of adding white noise is to make it easier to identify different frequency components during the decomposition process. The signal-to-noise ratio (SNR) of the white noise is set, and the white noise signal is superimposed on the original wind power data to generate a new input sequence, thereby avoiding the frequency aliasing problem (i.e., modal aliasing) in the signal.

[0033] Step S202: Perform empirical mode decomposition on each wind power data sequence after adding white noise to obtain H The decomposition results are then taken HThe average value of all the first intrinsic mode function components in the group decomposition results is used as the first intrinsic mode function component of the complete empirical mode decomposition of the adaptive noise;

[0034] in, H The first intrinsic mode function component in the group decomposition result reflects the highest oscillation frequency in the current frequency component. Since CEEMDAN is a complete set decomposition method, the process is repeated many times, and different white noise is added each time to generate a series of IMF components with the same frequency. Then take H The average value of all the first intrinsic mode function components in the group decomposition results is used as the first intrinsic mode function component of the complete empirical mode decomposition of adaptive noise. The high-frequency oscillation component is effectively extracted by taking the average value, and the random error introduced by white noise is reduced, thereby improving the stability of the IMF component. The first intrinsic mode function component of the complete empirical mode decomposition of adaptive noise finally obtained represents the highest frequency component of wind power data.

[0035] Step S203, subtracting the first intrinsic mode function component of the complete empirical mode decomposition of the adaptive noise from the processed data set to obtain a residual signal, which will be used as an input for the next decomposition;

[0036] Step S204, re-adding white noise to the residual signal and performing empirical mode decomposition to generate a second intrinsic mode function component;

[0037] Step S205, repeatedly executing step S203 and step S204 until the residual signal can no longer be decomposed, and obtaining a plurality of intrinsic mode function components.

[0038] In each iteration, white noise is added to the new residual signal, and the frequency of the decomposed IMF component is gradually reduced until the frequency of the residual signal is the lowest and cannot be further decomposed. Through multiple iterations, the CEEMDAN algorithm decomposes the original wind power data into a set of IMF components, which cover all frequency components from high frequency to low frequency. Each IMF component represents a different time scale characteristic of wind power fluctuations. Generally, high-frequency IMF components reflect short-period random fluctuations, while low-frequency IMF components may correspond to longer-period trends. After all IMF components are decomposed, the remaining low-frequency residual part is used as the final residual term, representing the smooth part or long-term trend of the wind power data.

[0039] Step S3, using a Bayesian optimization algorithm to perform a global search on the hyperparameters of the Mamba model based on physical constraints to obtain optimal hyperparameters, wherein the Mamba model based on physical constraints is based on the Mamba model and uses a second-order Jensen wake model as a physical constraint.

[0040] Wherein, step S3 specifically includes:

[0041] Step S301, using the Gaussian Process model (GP) as a proxy model, the Gaussian Process model is used to calculate the objective function of the hyperparameter combination g ( w ) to make predictions and get the predicted mean v ( w ) and the prediction variance r 2 ( w ), the construction expression of the Gaussian process model is:

[0042] ;

[0043] ;

[0044] in, w represents a hyperparameter combination, m is the kernel function, W n represents the currently evaluated hyperparameter combination, z n Representation and W n The corresponding objective function value is is the variance of the observation noise, I is the identity matrix in the Gaussian process model;

[0045] Step S302, based on the predicted mean and predicted variance, a Bayesian optimization algorithm is used to determine the hyperparameter combination to be evaluated in the next step. In this embodiment, an expected improvement (EI) acquisition function is used, and the expression is:

[0046] ;

[0047] ;

[0048] in, In order to improve the acquisition function, is the current optimal objective function value, is the cumulative distribution function of the standard normal distribution, is the probability density function of the standard normal distribution, Y is the standardized improvement amount;

[0049] Step S303, selecting the next hyperparameter combination by the expected improvement acquisition function W n+1 , and apply the hyperparameter combination in the Mamba model Wn+1 , and calculate the corresponding prediction error, combining the hyperparameters W n+1 The corresponding objective function value g ( W n+1 ) Add the existing data set to update the Gaussian process model;

[0050] Each time new data is added, the predicted mean and predicted variance are updated, so that the Gaussian process model gradually approaches the objective function.

[0051] Step S304: After multiple iterations, the Bayesian optimization algorithm evaluates the hyperparameters based on the proxy model and the expected improved acquisition function, and finally determines the optimal hyperparameters. w + , the expression is:

[0052] .

[0053] Step S4, using the intrinsic mode function components obtained in step S2 as input features of the Mamba model based on physical constraints, and integrating them with the optimal hyperparameters obtained in step S3 into the Mamba model based on physical constraints for training, thereby obtaining a trained Mamba model based on physical constraints.

[0054] Among them, the core of the Mamba model is a state space model (SSM) mechanism, which can represent any cyclic process with potential states, use first-order differential equations to represent the evolution of the internal state of the system, and use another set of differential equations to describe the relationship between the potential state and the output sequence. The structured state space model (SSM) is derived from SSM, which can effectively capture the dependencies of long time series through its internal state storage and update mechanism, making up for the shortcomings of traditional models in long sequence modeling. The Mamba model introduces a selection mechanism in the structured state space model, which selects relevant information and ignores irrelevant content through dynamic adjustment of input dependencies, thereby achieving better memory management and more efficient sequence modeling. Therefore, not only can relevant information be captured more efficiently and effectively in long sequences, but there is also a mechanism for the model to selectively propagate or forget information based on the input content, thereby enhancing the model's feature extraction ability and maintaining linear complexity when processing long sequence data, thereby improving the accuracy and efficiency of wind power prediction.

[0055] In this embodiment, the state space model expression of the Mamba model is:

[0056] ;

[0057] ;

[0058] in, q ( t ) is the hidden state sequence, express q ( t ), p ( t ) is the input sequence, f ( t ) is the output sequence, A , B , C is a learnable matrix;

[0059] The zero-order hold method is used to discretize the state space model to obtain a structured state space model. The structured state space model obtained by discretization is expressed as:

[0060] ;

[0061] ;

[0062] ;

[0063] ;

[0064] in, q t is a discrete time step t The hidden state sequence of q t-1 is a discrete time step t -1 hidden state sequence, and is the discretized learnable matrix, p t is a discrete time step t The input sequence is f t is a discrete time step t The output sequence of E is the identity matrix in the structured state-space model.

[0065] The Mamba model is based on the structured state space model by introducing input-dependent dynamic parameters. , C and As an input dependency function, the model's selectivity mechanism is enhanced so that it can dynamically adjust state updates based on the input content. Through this mechanism, information can be selectively memorized or forgotten, thereby improving adaptability and processing capabilities for long sequences.

[0066] In step S4, the optimal hyperparameters searched by the Bayesian optimization algorithm are input into the Mamba model, and the data after the complete empirical mode decomposition of the adaptive noise is substituted into the Mamba model as input features for training and prediction, and finally the predicted value of the Mamba model is obtained.

[0067] Step S5: Use the trained Mamba model based on physical constraints to predict wind power.

[0068] In this embodiment, by substituting the wind farm data into the second-order Jensen wake physical model, the power prediction value of the second-order Jensen wake physical model can be calculated. Then, the residual between the prediction value of the Mamba model and the power prediction value of the second-order Jensen wake physical model is calculated, also known as the regularization term, and this is embedded into the loss function of the deep learning network as a physical constraint term, and gradient descent is performed to update the weight parameters of the deep learning network. The prediction results of the second-order Jensen wake physical model and the Mamba model are integrated, and the wind power prediction value is finally output, which can make the prediction results more interpretable, more accurate, and avoid overfitting.

[0069] Among them, the total loss function of the trained physical constraint-based Mamba model is L The expression is:

[0070] ;

[0071] ;

[0072] ;

[0073] ;

[0074] in, L 1 represents the loss function of the Mamba model, L 2 represents the regularization term, β represents the weight of the regularization term, N represents the number of samples, y i Indicates i The true value of the sample, The Mamba model is i The predicted value of the sample, P i The second-order Jensen wake physical model is i The predicted value of the sample, represents the wind field density,D represents the swept area of ​​the wind turbine rotor, K represents the power conversion efficiency of the wind turbine, Represents the inflow wind speed to the wind turbine.

[0075] The method of the present invention and the comparative scheme are compared and analyzed by evaluation indicators, and the evaluation indicators include root mean square error (RMSE) and goodness of fit R 2 RMSE reflects the prediction error level of the model. The smaller the value, the better the prediction effect of the model. R 2 The closer it is to 1, the stronger the explanatory power of the model. In the comparison scheme, except that the second-order Jensen wake model is not used as the physical constraint, the rest is the same as the present invention. The comparison results are shown in Table 1:

[0076] Table 1

[0077]

[0078] It can be seen from Table 1 that the method of the present invention shows significant advantages in the wind power prediction task. R 2 The goodness of fit is 0.99 compared with the comparison scheme. R 2 The improvement was 0.02, and the RMSE decreased by 0.42. It can be seen that the method of the present invention not only improves the learning ability, but also conforms to the physical laws of the data set on this basis. Therefore, the present invention can provide more accurate prediction results and better model fitting.

[0079] In summary, the wind power prediction method provided by the present invention has the following beneficial effects:

[0080] (1) The present invention uses the complete empirical mode decomposition with adaptive noise (CEEMDAN) to decompose the wind power data set to obtain several intrinsic mode function (IMF) components, thereby effectively removing the noise in the data in the frequency domain and enhancing the regularity of the input features, so that the model can still maintain a high prediction accuracy on the data set with large noise;

[0081] (2) The present invention uses the Bayesian optimization (BO) algorithm to perform a global search for the hyperparameters of the Mamba model, uses the Gaussian process (GP) as a proxy model to approximate the objective function, and combines the expected improvement (EI) acquisition function to select the optimal hyperparameters, which can effectively improve the efficiency and accuracy of model training. The Bayesian optimization algorithm can dynamically adjust the search direction according to the update of the proxy model, so as to quickly find the optimal hyperparameters and improve the adaptability of the Mamba model to different wind farm data. In addition, the Mamba model adopts a state space model mechanism, which has higher computational efficiency and feature extraction capabilities than the traditional Transformer model. The Mamba model uses a selective state propagation and information forgetting mechanism to enable it to more effectively capture key information in long time series prediction, reduce information redundancy, and thus improve the overall prediction performance of the model;

[0082] (3) The present invention uses the second-order Jensen wake model as a physical constraint and models the physical wake effect of the wind farm through the second-order Jensen wake model. It can effectively describe the mutual influence between the turbines in the wind farm and calculate the power prediction value of the second-order Jensen wake physical model, thus realizing prediction based on real physical phenomena. Moreover, this physical constraint is embedded into the total loss function through the regularization term, which enhances the generalization ability in complex scenarios, reduces the deviation between the Mamba model prediction value and the physical reality, makes the prediction result more in line with the physical law, effectively improves the interpretability and reliability of the model, and improves the accuracy, stability and robustness of the prediction.

[0083] (4) By adopting the method of the present invention, it is possible to make adjustments in advance according to the predicted power changes and determine the timing and amplitude of the maximum power point tracking controller adjustment, thereby effectively reducing the frequent adjustment problem of the MPPT system caused by severe wind speed fluctuations and avoiding efficiency losses due to excessive adjustment or delayed adjustment.

[0084] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0085] Although the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the claims and their equivalents.

Claims

1. A wind power prediction method, characterized in that: The following steps are involved: Step S1, obtaining an actual wind power data set, using the Pearson coefficient method to calculate the correlation coefficient between each input feature in the actual wind power data set and the wind power, and then obtaining the target input feature, and then performing Z-score normalization processing on the target input feature to obtain a processed data set; Step S2, performing complete empirical mode decomposition of adaptive noise on the processed data set to decompose the data into several intrinsic mode function components; Step S3, using a Bayesian optimization algorithm to perform a global search on the hyperparameters of the Mamba model based on physical constraints to obtain optimal hyperparameters, wherein the Mamba model based on physical constraints is based on the Mamba model and uses a second-order Jensen wake model as a physical constraint; Step S4, using the intrinsic mode function components obtained in step S2 as input features of the Mamba model based on physical constraints, and integrating them with the optimal hyperparameters obtained in step S3 into the Mamba model based on physical constraints for training, thereby obtaining a trained Mamba model based on physical constraints; Step S5, using the trained Mamba model based on physical constraints to predict wind power; Step S2 specifically include: Step S201, adding the wind power data sequence in the processed data set H Substandard normal white noise, we get H A wind power data sequence with white noise added; Step S202: Perform empirical mode decomposition on each wind power data sequence after adding white noise to obtain H The decomposition results are then taken H The average value of all the first intrinsic mode function components in the group decomposition results is used as the first intrinsic mode function component of the complete empirical mode decomposition of the adaptive noise; Step S203, subtracting the first intrinsic mode function component of the complete empirical mode decomposition of the adaptive noise from the processed data set to obtain a residual signal, which will be used as an input for the next decomposition; Step S204, re-adding white noise to the residual signal and performing empirical mode decomposition to generate a second intrinsic mode function component; Step S205, repeatedly executing step S203 and step S204 until the residual signal can no longer be decomposed, and obtaining a plurality of intrinsic mode function components; Step S3 specifically includes: Step S301, using the Gaussian process model as a proxy model, the Gaussian process model is used to calculate the objective function of the hyperparameter combination g ( w ) to make predictions and get the predicted mean v ( w ) and the prediction variance r 2 ( w ), the construction expression of the Gaussian process model is: ； ； in, w represents a hyperparameter combination, m is the kernel function, W n represents the currently evaluated hyperparameter combination, z n Representation and W n The corresponding objective function value is is the variance of the observation noise, I is the identity matrix in the Gaussian process model; Step S302, based on the predicted mean and predicted variance, a Bayesian optimization algorithm is used to determine the hyperparameter combination to be evaluated in the next step, and the expression is: ； ； in, In order to improve the acquisition function, is the current optimal objective function value, is the cumulative distribution function of the standard normal distribution, is the probability density function of the standard normal distribution, Y is the standardized improvement amount; Step S303, selecting the next hyperparameter combination by the expected improvement acquisition function W n+1 , and apply the hyperparameter combination in the Mamba model W n+1 , and calculate the corresponding prediction error, combining the hyperparameters W n+1 The corresponding objective function value g ( W n+1 ) Add the existing data set to update the Gaussian process model; Step S304: After multiple iterations, the Bayesian optimization algorithm evaluates the hyperparameters based on the proxy model and the expected improved acquisition function, and finally determines the optimal hyperparameters. w + , the expression is: ； The state space model expression of the Mamba model is: ； ； in, q ( t ) is the hidden state sequence, express q ( t ), p ( t ) is the input sequence, f ( t ) is the output sequence, A , B , C is a learnable matrix; The zero-order hold method is used to discretize the state space model to obtain a structured state space model. The structured state space model obtained by discretization is expressed as: ； ； ； ； in, q t is a discrete time step t The hidden state sequence of q t-1 is a discrete time step t -1 hidden state sequence, and is the discretized learnable matrix, p t is a discrete time step t The input sequence is f t is a discrete time step t The output sequence of E is the identity matrix in the structured state-space model; In step S4, the optimal hyperparameters searched by the Bayesian optimization algorithm are input into the Mamba model, and the data after the complete empirical mode decomposition of the adaptive noise is substituted into the Mamba model as input features for training and prediction, and finally the prediction value of the Mamba model is obtained; Total loss function for the trained physics-based Mamba model L The expression is: ； ； ； ； in, L 1 represents the loss function of the Mamba model, L 2 represents the regularization term, β represents the weight of the regularization term, N represents the number of samples, y i Indicates i The true value of the sample, The Mamba model is i The predicted value of the sample, P i The second-order Jensen wake physical model is i The predicted value of the sample, represents the wind field density, D represents the swept area of ​​the wind turbine rotor, K represents the power conversion efficiency of the wind turbine, Represents the inflow wind speed to the wind turbine.

2. The wind power prediction method according to claim 1, characterized in that: In step S1, the correlation coefficient between each input feature and wind power in the actual wind power data set is calculated using the following formula: ； in, m 1 represents the sample value of the first input feature in the actual wind power dataset. represents the sample mean of the first input feature, m 2 represents the wind power corresponding to the first input feature, represents the mean wind power corresponding to the first input feature, Represents the correlation coefficient corresponding to the first input feature; The input features with correlation coefficients greater than the threshold are used as target input features; The following formula is used to normalize the target input feature by Z-score: ； in, o is the original value, is the standardized value, is the mean, is the standard deviation.

Images