A wind farm wind speed prediction method and system based on multi-model integration

Abstract

The present invention relates to the field of wind speed prediction, and discloses a wind speed prediction method and system for a wind farm based on multi-model integration. The method comprises: obtaining historical wind speed w 0 (t) , obtained by interpolation and elimination of drift wind speed w(t) , and decomposed using DVNCMD w(t) ; Calculate the fusion feature entropy of each sequence obtained by decomposition, and divide it into low, medium and high entropy value sequences according to the size of the entropy value; Establish BP, LSTM, CNN‑Bi‑LSTM‑Attention models corresponding to low, medium and high entropy value sequences, and use NRBO to optimize the hyperparameters of LSTM and CNN‑Bi‑LSTM‑Attention models; Finally, superimpose all sequence prediction results to obtain the final predicted wind speed. The system uses a data volume threshold judgment module to update model parameters to adapt to wind speed changes caused by climate change, and displays predicted data in real time. The present invention uses DVNCMD decomposition to reduce the amount of calculation and improve decomposition efficiency, while establishing BP, LSTM, and CNN‑Bi‑LSTM‑Attention models to optimize calculation efficiency and reduce calculation time, and adds NRBO optimization algorithm to further improve prediction accuracy and efficiency, providing new ideas and methods for wind speed power generation.

Description

Technical Field

[0001] The present invention relates to the field of wind speed prediction and wind power generation, and aims to provide a method and system for predicting wind speed in a wind farm based on multi-model integration. Background Art

[0002] As a clean and renewable energy source, wind power generation has been widely used around the world. The power generation efficiency of wind farms is closely related to wind speed changes. Therefore, accurate wind speed prediction is crucial to optimizing the operation of wind farms and increasing power output. Accurate wind speed prediction not only helps power grid operators to reasonably dispatch power generation equipment, but also effectively reduces the risks brought by wind speed fluctuations.

[0003] Although a variety of wind speed prediction technologies have been developed, wind speed prediction methods contain some limitations, including statistical methods, physical models, and machine learning techniques. Statistical methods usually require a large amount of historical data; physical models rely on accurate geographical and climate information; machine learning techniques, although they perform well in processing large data sets and discovering potential patterns, often face problems of insufficient prediction accuracy or poor model generalization when faced with complex meteorological data.

[0004] In order to improve the accuracy and reliability of wind speed prediction, overcome the limitations of existing methods and better adapt to changing meteorological conditions. The present invention proposes a multi-model integrated wind farm wind speed prediction method and system, which improves the overall prediction accuracy and shortens the calculation time by constructing corresponding prediction models for different sequence complexities. In addition, the system is also provided with a data volume threshold judgment module, which can update the model parameters to adapt to wind speed changes, thereby ensuring the accuracy and timeliness of the prediction results. Summary of the invention

[0005] Purpose of the invention: In response to the problems pointed out in the background technology, the present invention discloses a wind speed prediction method and system for a wind farm based on multi-model integration. By integrating multiple prediction models, the limitations of existing wind speed prediction methods are solved and the accuracy and reliability of the prediction are improved. It can not only optimize the operation and power generation efficiency of wind farms, but also provide important prediction information for power grid operators, help them reasonably dispatch power generation equipment, reduce the risks brought by wind speed fluctuations, and thus promote the development of the clean energy field.

[0006] Technical solution: The present invention discloses a wind speed prediction method for a wind farm based on multi-model integration, comprising the following steps:

[0007] Step 1: Obtain historical wind speed data w0(t);

[0008] Step 2: Use the generative adversarial network GAN to fill the missing values ​​in w0(t) to obtain the complete wind speed data w1(t). Then use the short-time Fourier transform STFT with the spectral kurtosis algorithm to accurately identify and remove the drift data in the data to obtain the wind speed data w(t), and then perform normalization processing;

[0009] Step 3: Introduce reduced-order variational nonlinear chirp modal decomposition DVNCMD to decompose the data into K modal functions IMF, calculate the information entropy H, sample entropy S, and permutation entropy P of each modal component, and obtain the fused feature entropy E of each component through feature fusion comprehensive entropy value;

[0010] Step 4: Use the classifier model to divide the sequence into low entropy, medium entropy and high entropy sequences, and establish BP neural network, LSTM and CNN-Bi-LSTM-Attention prediction models for prediction;

[0011] Step 5: Use the Newton-Raphson optimization algorithm (NRBO) to optimize the hyperparameters of LSTM and CNN-Bi-LSTM-Attention to obtain the optimal parameter combination;

[0012] Step 6: Denormalize the data, add up the prediction results of each component, and get the final wind speed prediction data;

[0013] Step 7: Evaluate the prediction results, and the evaluation indicators cited are not limited to Kling-Gupta efficiency coefficient KGE and mean relative error MAPE.

[0014] Furthermore, the specific steps of step 2 are as follows:

[0015] Step 2.1: Import historical wind speed data w0(t), and use GAN to fill the missing values ​​in w0(t) to obtain complete wind speed data w1(t). The steps are as follows:

[0016] a: First, build a GAN, which includes generating missing wind speed data with a distribution similar to w0(t) The generator G and the discriminator D whose output data w is the probability of w0(t) are as follows:

[0017] Generator:

[0018] Discriminator: D(w,θ D )∈[0,1]

[0019] Where: z is the random noise of the generator, w is w0(t) or θ G ,θ D are the parameters of the generator and the discriminator respectively;

[0020] b: GAN is then trained by minimizing a specific objective function V(D, G), as follows:

[0021]

[0022] Where: P data(w) represents the data distribution of w0(t), P z(z) represents the prior distribution of random noise z, represents the expected log-likelihood of the historical wind speed data w0(t), Indicates the generated data The expected log-likelihood of ;

[0023] c: Construct the complete wind speed data w1(t), detect w0(t), if it is the original value, keep it, if it is a missing value, use the data generated by the generator G to fill it, and finally get the complete wind speed data 0, as follows:

[0024]

[0025] Step 2.2: Use STFT with spectral kurtosis algorithm to accurately identify w1(t) and remove the drift wind speed in the data. Define W(f, t) as the STFT result of w1(t) at frequency f and time t; Φ(f, t) is the spectral kurtosis, which is used to identify the drift wind speed of w1(t). The expression is as follows:

[0026]

[0027] Where: j is an imaginary unit, Θ is the threshold set by the adaptive threshold method; R(f,t) is an indicator variable, when it is 1, it means there is drift wind speed; when its value is 0, there is no drift.

[0028] Step 2.3: The calculation formula of wind speed w(t) is as follows:

[0029] w(t)=IFFT(∑ f [W(f,t)·(1-R(f,t))])

[0030] Where: IFFT is the inverse short-time Fourier transform.

[0031] Furthermore, step 3 introduces reduced-order variational nonlinear frequency modulation mode decomposition DVNCMD to decompose the wind speed data w(t) and calculate the fusion characteristic entropy E of each component, specifically including:

[0032] Step 3.1: Determine the window size T and construct the wind speed prediction input feature matrix W(t) for the wind speed data w(t) in the time window [tT, t];

[0033] Step 3.2: Then select the nonlinear chirp basis function ψ(t), which is as follows:

[0034]

[0035] Where: j is an imaginary unit, a i is the amplitude, b i (t) is the instantaneous frequency, θ i is the phase, ψ i (t) the i-th nonlinear chirp basis function;

[0036] Step 3.3: Construct an energy function F to measure the wind speed w(t) and the wind speed w approximated by the nonlinear chirp basis function approx The differences between (t) are as follows:

[0037]

[0038] Where: c i is the coefficient, I is the number of basis functions;

[0039] Step 3.4: Construct the IMF components of wind speed w(t) as follows:

[0040] a: Select a set of initial basis function parameters a i , b i (t),θ i ;

[0041] b: Use the gradient descent method to minimize the energy function F, find the optimal basis function parameters and construct the approximate wind speed w approx (t);

[0042] c: Approximate wind speed w approx (t) is subtracted from the wind speed w(t) to obtain the residual w r (t);

[0043] e: The residual w r (t) is used as a new input, and steps b to c are repeated to obtain K IMF components of wind speed w(t);

[0044] Step 3.5: Calculate the information entropy H, sample entropy S and permutation entropy P of the IMF components, determine the weights by the coefficient of variation method, and calculate the fusion entropy E to quantify the complexity of the K IMF components. A high entropy value indicates a high complexity. The formula of E is as follows:

[0045] E=ω1×H+ω2×S+ω3×P

[0046] Where: ω1, ω2, ω3 are the weights of information entropy H, sample entropy S, and permutation entropy P respectively.

[0047] Furthermore, step 4 introduces a classifier model, which specifically includes:

[0048] Step 4.1: According to the fusion entropy E obtained in step 3, first calculate the mean value, the formula is as follows:

[0049]

[0050] Where: E s is the fusion feature entropy of the sth IMF component;

[0051] Step 4.2: Set high and low entropy thresholds, and divide the sequence into stable sequence, relatively stable sequence and complex sequence according to low entropy value, medium entropy value and high entropy value. The threshold calculation formula is as follows:

[0052]

[0053] Where: H E , L E It represents the high threshold of entropy and the low threshold of entropy, and σ1 and σ2 represent the proportion of high and low thresholds.

[0054] Furthermore, step 5 uses the Newton-Raphson optimization algorithm NRBO to optimize the number of hidden layer neurons, learning rate and batch size of LSTM, and optimizes the learning rate, number of hidden layer nodes, number, size and step size of convolution kernels of CNN layer, number of LSTM units of Bi-LSTM layer, query, key and value of Attention of CNN-Bi-LSTM-Attention.

[0055] Furthermore, the specific steps of using the Newton-Raphson optimization algorithm NRBO to optimize the hyperparameters of LSTM and CNN-Bi-LSTM-Attention are as follows:

[0056] Step 5.1: Initialize the hyperparameters of LSTM and CNN-Bi-LSTM-Attention:

[0057] ξ k =ξ k,min +rand·(ξ k,max -ξ k,min )

[0058] Where: k represents the kth parameter, ξ k,min ,ξ k,max Respectively represent the minimum and maximum values ​​of the parameters, rand is a random number between (0, 1);

[0059] Step 5.2: Introduce the NRBO algorithm to optimize hyperparameters. The specific steps are as follows:

[0060] a: Use the Newton-Raphson search rule NRSR to iterate and obtain the updated value of the hyperparameter. The iteration formula is as follows:

[0061]

[0062] Where: h is the number of iterations;

[0063] b: Use the Trap Avoidance Operator (TAO) to help the algorithm avoid falling into the local optimal solution, where TAO is calculated by optimizing the hyperparameter ξ best and k,h To generate an enhanced hyperparameter The details are as follows:

[0064]

[0065] ξ=θ1×(μ1×ξ best -μ2×ξ k,h )+θ2×δ×(μ1×Mean(ξ k,h )-μ2×ξ k,h

[0066] μ1=3×β×rand+(1-β)

[0067] μ2=β×rand+(1-β)

[0068] Where: β is a binary number, θ1∈(-1,1),θ2∈(-0.5,0.5), Mean(ξ k,h ) is expressed as mean;

[0069] c: Use the greedy criterion to determine whether to use the hyperparameters generated by TAO

[0070] Step 5.3: Use mean square error to perform fitness analysis, and then determine whether the maximum number of iterations has been reached:

[0071] Step 5.4: Output the optimal hyperparameter combination.

[0072] The present invention also discloses a wind farm wind speed prediction system based on multi-model integration, the system comprising: a data acquisition module, a gateway node, an Alibaba Cloud Internet of Things platform, a data volume threshold judgment module, an intelligent prediction system module and a client display module;

[0073] The data acquisition module is connected to the Alibaba Cloud Internet of Things platform through a gateway node, and the intelligent prediction system module is connected to the Alibaba Cloud Internet of Things platform. It predicts the wind speed of the wind farm based on the historical wind speed data w0(t) collected by the data acquisition module. The above-mentioned wind farm wind speed prediction method based on multi-model integration is executed between the intelligent prediction system module and the data acquisition module. The client display module is connected to the Alibaba Cloud Internet of Things platform through the intelligent prediction system module to visually display the real-time detected wind speed data and wind speed prediction results.

[0074] Preferably, it also includes a data volume threshold judgment module, which is arranged between the intelligent prediction system module and the Alibaba Cloud Internet of Things platform, and is used to determine when to trigger the data volume threshold for model retraining in the intelligent prediction system module.

[0075] Beneficial effects:

[0076] The prediction method proposed in the present invention uses short-time Fourier transform (STFT) in combination with spectral kurtosis algorithm, which can effectively identify drift data in wind speed signals and remove drift, significantly improve the quality of wind speed data, reduce the influence of noise and outliers, and thus improve the accuracy of wind speed prediction. In addition, the reduced-order variational nonlinear chirp mode decomposition (DVNCMD) is used to effectively reduce the amount of calculation and improve the decomposition efficiency through order reduction processing, while ensuring the accurate capture of key wind speed features.

[0077] The present invention proposes a fusion entropy (E) strategy, which provides a multi-angle and multi-dimensional data complexity quantification method by fusing information entropy (H), sample entropy (S) and permutation entropy (P). This strategy can not only accurately evaluate the uncertainty and complexity of wind speed sequences, but also effectively divide wind speed sequences into low entropy, medium entropy and high entropy sequences according to the evaluation results, providing a scientific basis for selecting a suitable model.

[0078] The present invention establishes BP neural network, LSTM and CNN-Bi-LSTM-Attention prediction models for low entropy, medium entropy and high entropy sequences respectively. This strategy optimizes the use of computing resources and avoids excessive calculation of simple sequences; enhances the generalization ability of the model for data of different complexity; and reduces the calculation time through targeted model selection. In addition, NRBO is used to optimize the LSTM and CNN-Bi-LSTM-Attention prediction models to obtain the optimal hyperparameter combination, further improving the prediction accuracy.

[0079] The system proposed in the present invention establishes a data volume threshold judgment module to timely update the model parameters. At the same time, the system has real-time monitoring and feedback functions, which can help users obtain the latest forecast information at any time. The system timely updates the model parameters to enable the model to adapt to climate change, thereby maximizing the model prediction accuracy and thus improving the power generation efficiency of the wind farm. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] Figure 1 Shown is a flow chart of the content proposed by the present invention;

[0081] Figure 2 Shown is a system diagram proposed by the present invention;

[0082] Figure 3 The figure shows the intelligent prediction flow chart proposed by the present invention;

[0083] Figure 4 The figure shows the optimization flow chart of the improved algorithm;

[0084] Figure 5 The figure shows the result of data preprocessing for the Sotavento wind farm;

[0085] Figure 6 The figure shows the result of DVNCMD decomposition data;

[0086] Figure 7 , Figure 8 They respectively represent the prediction result graph and the prediction result indicator graph. DETAILED DESCRIPTION

[0087] The technical solution of the present invention is further described in detail below with reference to examples.

[0088] The present invention uses a wind speed data sequence collected from the Sotavento wind farm in Galicia Province, Spain to perform an example simulation to verify the effect of the present invention.

[0089] Figure 1 A flow chart of the content of the present invention, Figure 2 This is a system block diagram proposed in the present invention. The system includes: a data acquisition module, a gateway node, an Alibaba Cloud Internet of Things platform, a data volume threshold judgment module, an intelligent prediction system module, and a client display module.

[0090] The data acquisition module is connected to the Alibaba Cloud IoT platform through a gateway node. During operation, the data acquisition module determines the input wind speed and uploads it to the Alibaba Cloud IoT platform through a gateway node, providing data for the intelligent prediction module while ensuring data storage. The intelligent prediction system module is connected to the Alibaba Cloud IoT platform, and performs wind speed prediction for the wind farm based on the historical wind speed data w0(t) collected by the data acquisition module. The client display module is connected to the Alibaba Cloud IoT platform through the intelligent prediction system module, and visually displays the real-time detected wind speed data and wind speed prediction results. The system updates the model parameters through the data volume threshold judgment module, which is set between the intelligent prediction system module and the Alibaba Cloud IoT platform to determine when to trigger the data volume threshold for retraining the model in the intelligent prediction system module to adapt to the wind speed changes in different seasons. Finally, the client display module can observe the predicted data in real time. The data volume threshold judgment module is designed to trigger the model retraining once the threshold is reached, and the model parameters are updated in real time to adapt to the wind speed changes caused by climate change.

[0091] Figure 3 The intelligent prediction flow chart proposed by the present invention is as follows:

[0092] Step 1: Get historical wind speed data w0(t).

[0093] Step 2: Use the generative adversarial network (GAN) to fill the missing values ​​in w0(t) to obtain the complete wind speed data w1(t). Then use the short-time Fourier transform (STFT) combined with the spectral kurtosis algorithm to accurately identify and remove the drift data in the data, obtain the wind speed data w(t), and normalize it.

[0094] Step 2.1: Import historical wind speed data w0(t), and use GAN to fill the missing values ​​in w0(t) to obtain complete wind speed data w1(t). The steps are as follows:

[0095] a: First, build a GAN, which includes generating missing wind speed data with a distribution similar to w0(t) The generator G and the discriminator D whose output data w is the probability of w0(t) are as follows:

[0096] Generator:

[0097] Discriminator: D(w,θ D )∈[0,1]

[0098] Where: z is the random noise of the generator, w is w0(t) or θ G ,θ Dare the parameters of the generator and the discriminator respectively.

[0099] b: GAN is then trained by minimizing a specific objective function V(D, G), as follows:

[0100]

[0101] Where: P data(w) represents the data distribution of w0(t), P z(z) represents the prior distribution of random noise z, represents the expected log-likelihood of the historical wind speed data w0(t), Indicates the generated data The expected log-likelihood of .

[0102] c: Construct the complete wind speed data w1(t), detect w0(t), if it is the original value, keep it, if it is a missing value, use the data generated by the generator G to fill it, and finally get the complete wind speed data w1(t), as follows:

[0103]

[0104] Step 2.2: Use STFT with spectral kurtosis algorithm to accurately identify w1(t) and remove the drift wind speed in the data. Define W(f, t) as the STFT result of w1(t) at frequency f and time t; Φ(f, t) is the spectral kurtosis, which is used to identify the drift wind speed of w1(t). R(f, t) is used to determine whether there is drift wind speed. The expression is as follows:

[0105]

[0106] Where: j is the imaginary unit, Θ is the threshold set by the adaptive threshold method.

[0107] Step 2.3: The calculation formula of wind speed w(t) is as follows:

[0108] w(t)=IFFT(Σ f [W(f,t)·(1-R(f,t))])

[0109] Where: IFFT is the inverse short-time Fourier transform.

[0110] Step 3: Reduced-order variational nonlinear chirp modal decomposition (DVNCMD) is introduced to decompose the data into K modal functions (IMFs), calculate the information entropy (H), sample entropy (S), and permutation entropy (P) of each modal component, and integrate these entropy values ​​through feature fusion to obtain the fused feature entropy (E) of each component.

[0111] Step 3.1: Determine the window size T and construct the wind speed prediction input feature matrix W(t) for the wind speed data w(t) in the time window [tT, t].

[0112] Step 3.2: Then select the nonlinear chirp basis function ψ(t), which is as follows:

[0113]

[0114] Where: j is an imaginary unit, a i is the amplitude, b i (t) is the instantaneous frequency, θ i is the phase, ψ i (t) The i-th nonlinear chirp basis function.

[0115] Step 3.3: Construct an energy function F to measure the wind speed w(t) and the wind speed w approximated by the nonlinear chirp basis function approx The differences between (t) are as follows:

[0116]

[0117] Where: c i is the coefficient and I is the number of basis functions.

[0118] Step 3.4: Construct the IMF components of wind speed w(t) as follows:

[0119] a: Select a set of initial basis function parameters a i , b i (t),θ i ;

[0120] b: Use the gradient descent method to minimize the energy function F, find the optimal basis function parameters and construct the approximate wind speed w approx (t);

[0121] c: Approximate wind speed w approx (t) is subtracted from the wind speed w(t) to obtain the residual w r (t);

[0122] e: The residual w r (t) is used as the new input and steps b to c are repeated to obtain the K IMF components of wind speed w(t).

[0123] Step 3.5: Calculate the information entropy (H), sample entropy (S) and permutation entropy (P) of the IMF components, determine the weights by the coefficient of variation method, and calculate the fusion entropy (E) to quantify the complexity of the K IMF components. A high entropy value indicates a high complexity. The formula for E is as follows:

[0124] E=ω1×H+ω2×S+ω3×P

[0125] Where: ω1, ω2, ω3 are the weights of information entropy (H), sample entropy (S), and permutation entropy (P), respectively.

[0126] Step 4: Through the classifier model, the sequence is divided into low entropy, medium entropy and high entropy sequences, and the BP neural network, LSTM and CNN-Bi-LSTM-Attention prediction models are established for prediction.

[0127] Step 4.1: According to the fusion entropy (E) obtained in step 3, first calculate the mean value, the formula is as follows:

[0128]

[0129] Where: E s is the fusion feature entropy of the sth IMF component.

[0130] Step 4.2: Set high and low entropy thresholds, and divide the sequence into stable sequence, relatively stable sequence and complex sequence according to low entropy value, medium entropy value and high entropy value. The threshold calculation formula is as follows:

[0131]

[0132] Where: H E , L E It represents the high threshold of entropy and the low threshold of entropy, and σ1 and σ2 represent the proportion of high and low thresholds.

[0133] Step 5: Use the Newton-Raphson optimization algorithm (NRBO) to optimize the hyperparameters of LSTM and CNN-Bi-LSTM-Attention to obtain the optimal parameter combination.

[0134] The Newton-Raphson optimization algorithm (NRBO) is used to optimize the number of hidden layer neurons, learning rate, and batch size of LSTM, the learning rate, number of hidden layer nodes, parameters of CNN layer (number, size, and step size of convolution kernels), parameters of Bi-LSTM layer (number of LSTM units), and parameters of Attention (query, key, and value) of CNN-Bi-LSTM-Attention:

[0135] Step 5.1: Initialize the hyperparameters of LSTM and CNN-Bi-LSTM-Attention:

[0136] ξ k =ξ k,min +rand·(ξ k,max -ξ k,min )

[0137] Where: k represents the kth parameter, ξ k,min ,ξ k,max They represent the minimum and maximum values ​​of the parameters respectively, and rand is a random number between (0, 1).

[0138] Step 5.2: Introduce the NRBO algorithm to optimize hyperparameters. The specific steps are as follows:

[0139] a: Use the Newton-Raphson search rule (NRSR) to iterate and obtain the updated value of the hyperparameter. The iteration formula is as follows:

[0140]

[0141] Where: h is the number of iterations.

[0142] b: Use the Trap Avoidance Operator (TAO) to help the algorithm avoid falling into the local optimal solution, where TAO is calculated by optimizing the hyperparameter ξ best and k,h To generate an enhanced hyperparameter The details are as follows:

[0143]

[0144] ξ=θ1×(μ1×ξ best -μ2×ξ k,h )+θ2×δ×(μ1×Mean(ξ k,h )-μ2×ξ k,h

[0145] μ1=3×β×rand+(1-β)

[0146] μ2=β×rand+(1-β)

[0147] Where: β is a binary number, θ1∈(-1,1),θ2∈(-0.5,0.5), Mean(ξ k,h ) is expressed as the mean.

[0148] c: Use the greedy criterion to determine whether to use the hyperparameters generated by TAO

[0149] Step 5.3: Use mean square error to perform fitness analysis, and then determine whether the maximum number of iterations has been reached.

[0150] Step 5.4: Output the optimal hyperparameter combination.

[0151] Step 6: Denormalize the data, add up the prediction results of each component, and get the final wind speed prediction data.

[0152] Step 7: Evaluate the prediction results, using evaluation indicators including but not limited to Kling-Gupta efficiency coefficient (KGE) and mean relative error (MAPE).

[0153] In order to further describe the intelligent prediction module, the present invention selects wind speed data from 00:00 on April 17, 2017 to 23:50 on April 23, 2017 for model verification. The sampling interval of the data set is 10 minutes, and the total number is 1008.

[0154] like Figure 5 The figure shows the result of data preprocessing at the Sotavento wind farm, using GAN to intelligently fill in missing values ​​and STFT combined with spectral kurtosis to accurately identify and remove drift data in the data.

[0155] like Figure 6 As shown in the figure, the DVNCMD decomposition data result diagram, the entropy values ​​of each component are obtained by fusing the feature entropy as shown in Table 1:

[0156] Table 1. Calculation results of entropy values ​​for each sequence

[0157]

[0158] The classifier model is introduced, and the classification results are shown in Table 2 through calculation of each component:

[0159] Table 2 shows the calculation threshold and the classification of the sequence of entropy values.

[0160]

[0161]

[0162] In order to further prove the superiority of the model proposed in this paper, a single model LSTM and INRBO optimized LSTM were designed for comparative verification. The results are shown in the figure. Figure 7 shown.

[0163] The prediction results are evaluated, and the indicators are shown in Table 3. Figure 8 shown.

[0164] Table 3 shows the calculation results of the four indicators

[0165]

[0166] The above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable people familiar with the technology to understand the content of the present invention and implement it accordingly, and they cannot be used to limit the protection scope of the present invention. Any equivalent transformation or modification made according to the spirit of the present invention should be included in the protection scope of the present invention.

Claims

1. A wind farm wind speed prediction method based on multi-model integration, characterized in that: The following steps are involved: Step 1: Obtain historical wind speed data w0(t); Step 2: Use the generative adversarial network GAN to fill the missing values ​​in w0(t) to obtain the complete wind speed data w1(t). Then use the short-time Fourier transform STFT with the spectral kurtosis algorithm to accurately identify and remove the drift data in the data to obtain the wind speed data w(t), and then perform normalization processing; Step 3: Introduce reduced-order variational nonlinear chirp modal decomposition DVNCMD to decompose the data into K modal functions IMF, calculate the information entropy H, sample entropy S, and permutation entropy P of each modal component, and obtain the fused feature entropy E of each component through feature fusion comprehensive entropy value; Step 4: Through the classifier model, each modal component is divided into stable sequence, relatively stable sequence and complex sequence according to low entropy value, medium entropy value and high entropy value, and BP neural network, LSTM and CNN-Bi-LSTM-Attention prediction models are established based on the stable sequence, relatively stable sequence and complex sequence for prediction; Step 4.1: According to the fusion feature entropy E obtained in step 3, first calculate the mean value, the formula is as follows: Where: E s is the fusion feature entropy of the sth IMF component; Step 4.2: Set high and low entropy thresholds, and divide each modal component into stable sequence, relatively stable sequence and complex sequence according to low entropy value, medium entropy value and high entropy value. The threshold calculation formula is as follows: Where: H E , L E It represents the high threshold of entropy and the low threshold of entropy, σ1 and σ2 represent the proportion of high and low thresholds; Step 5: Use the Newton-Raphson optimization algorithm (NRBO) to optimize the hyperparameters of LSTM and CNN-Bi-LSTM-Attention to obtain the optimal parameter combination; Step 6: Denormalize the data, add up the prediction results of each component, and get the final wind speed prediction data; Step 7: Evaluate the prediction results, using the evaluation indicators Kling-Gupta efficiency coefficient KGE and mean relative error MAPE.

2. The wind farm wind speed prediction method based on multi-model integration according to claim 1 is characterized in that: The specific steps of step 2 are as follows: Step 2.1: Import historical wind speed data w0(t), and use GAN to fill the missing values ​​in w0(t) to obtain complete wind speed data w1(t). The steps are as follows: a: First, build a GAN, which includes generating missing wind speed data with a distribution similar to w0(t) The generator G and the discriminator D whose output data is probability are as follows: Generator: Discriminator: D(w,θ D )∈[0,1] Where: z is the random noise of the generator, w is w0(t) or θ G ,θ D are the parameters of the generator and the discriminator respectively; b: GAN is then trained by minimizing a specific objective function V(D, G), as follows: Where: P data(w) represents the data distribution of w0(t), P z(z) represents the prior distribution of random noise z, represents the expected log-likelihood of the historical wind speed data w0(t), express ; c: Construct the complete wind speed data w1(t), detect w0(t), if it is the original value, keep it, if it is a missing value, use the data generated by the generator G to fill it, and finally get the complete wind speed data, as follows: Step 2.2: Use STFT with spectral kurtosis algorithm to accurately identify w1(t) and remove the drift wind speed in the data. Define W(f, t) as the STFT result of w1(t) at frequency f and time t; Φ(f, t) is the spectral kurtosis, which is used to identify the drift wind speed of w1(t). The expression is as follows: Where: j is an imaginary unit, Θ is the threshold set by the adaptive threshold method; R(f,t) is an indicator variable, when it is 1, it means there is drift wind speed; when its value is 0, there is no drift; Step 2.3: The calculation formula of wind speed w(t) is as follows: w(t)=IFFT(Σ f [W(f,t)·(1-R(f,t))]) Where: IFFT is the inverse short-time Fourier transform.

3. The wind speed prediction method for a wind farm based on multi-model integration according to claim 1, characterized in that: Step 3 introduces reduced-order variational nonlinear frequency modulation mode decomposition DVNCMD to decompose the wind speed data w(t) and calculate the fusion characteristic entropy E of each component, including: Step 3.1: Determine the window size T and construct the wind speed prediction input feature matrix W(t) for the wind speed data w(t) in the time window [tT, t]; Step 3.2: Then select the nonlinear chirp basis function ψ(t), which is as follows: Where: j is the imaginary unit, a i is the amplitude, b i (t) is the instantaneous frequency, θ i is the phase, ψ i (t) the i-th nonlinear chirp basis function; Step 3.3: Construct an energy function F to measure the wind speed w(t) and the wind speed w approximated by the nonlinear chirp basis function approx The differences between (t) are as follows: Where: c i is the coefficient, I is the number of basis functions; Step 3.4: Construct the IMF components of wind speed w(t) as follows: a: Select a set of initial basis function parameters a i 、b i (t),θ i ; b: Use the gradient descent method to minimize the energy function F, find the optimal basis function parameters and construct the approximate wind speed w approx (t); c: Approximate wind speed w approx (t) is subtracted from the wind speed w(t) to obtain the residual w r (t); e: The residual w r (t) is used as a new input, and steps b to c are repeated to obtain K IMF components of wind speed w(t); Step 3.5: Calculate the information entropy H, sample entropy S and permutation entropy P of the K IMF components, determine the weights by the coefficient of variation method, and calculate the fusion entropy E to quantify the complexity of the K IMF components. A high entropy value indicates a high complexity. The formula of E is as follows: E=ω1×H+ω2×S+ω3×P Where: ω1, ω2, ω3 are the weights of information entropy H, sample entropy S, and permutation entropy P respectively.

4. The wind speed prediction method for a wind farm based on multi-model integration according to claim 1, characterized in that: Step 5 uses the Newton-Raphson optimization algorithm NRBO to optimize the number of hidden layer neurons, learning rate, and batch size of LSTM, and optimizes the learning rate, number of hidden layer nodes, number, size, and step size of convolution kernels of the CNN layer, number of LSTM units of the Bi-LSTM layer, and query, key, and value of Attention of CNN-Bi-LSTM-Attention.

5. The method for predicting wind speed in a wind farm based on multi-model integration according to claim 4, characterized in that: The specific steps for optimizing the hyperparameters of LSTM and CNN-Bi-LSTM-Attention using the Newton-Raphson optimization algorithm NRBO are as follows: Step 5.1: Initialize the hyperparameters of LSTM and CNN-Bi-LSTM-Attention: x k =ξ k,min +rand·(ξ k,max -x k,min ) Where: k represents the kth parameter, ξ k,min ,ξ k,max They represent the minimum and maximum values ​​of the kth parameter respectively, and rand is a random number between (0, 1); Step 5.2: Introduce the NRBO algorithm to optimize hyperparameters. The specific steps are as follows: a: Use the Newton-Raphson search rule NRSR to iterate and obtain the updated value of the hyperparameter. The iteration formula is as follows: Where: h is the number of iterations; b: Use the Trap Avoidance Operator (TAO) to help the algorithm avoid falling into the local optimal solution, where TAO is calculated by optimizing the hyperparameter ξ best and k,h To generate an enhanced hyperparameter The details are as follows: ξ=θ1×(μ1×ξ best -μ2×ξ k,h )+θ2×δ×(μ1×Mean(ξ k,h )-μ2×ξ k,h ) μ1=3×β×rand+(1-β) μ2=β×rand+(1-β) Where: β is a binary number, θ1∈(-1,1),θ2∈(-0.5,0.5), Mean(ξ k,h ) is expressed as mean; c: Use the greedy criterion to determine whether to use the hyperparameters generated by TAO Step 5.3: Use mean square error to perform fitness analysis, and then determine whether the maximum number of iterations has been reached: Step 5.4: Output the optimal hyperparameter combination.

6. A wind farm wind speed prediction system based on multi-model integration, characterized in that: The system includes: data collection module, gateway node, Alibaba Cloud IoT platform, data volume threshold judgment module, intelligent prediction system module and client display module; The data acquisition module is connected to the Alibaba Cloud Internet of Things platform through a gateway node, and the intelligent prediction system module is connected to the Alibaba Cloud Internet of Things platform. It predicts the wind speed of the wind farm based on the historical wind speed data w0(t) collected by the data acquisition module. The wind speed prediction method for wind farms based on multi-model integration as described in any one of claims 1 to 5 is executed between the intelligent prediction system module and the data acquisition module. The client display module is connected to the Alibaba Cloud Internet of Things platform through the intelligent prediction system module to visually display the real-time detected wind speed data and wind speed prediction results.

7. A wind farm wind speed prediction system based on multi-model integration according to claim 6, characterized in that: It also includes a data volume threshold judgment module, which is set between the intelligent prediction system module and the Alibaba Cloud Internet of Things platform, and is used to determine when to trigger the data volume threshold for model retraining in the intelligent prediction system module.

Images