A high-precision wind speed soft measurement method for wind power prediction of a wind farm

Abstract

The application provides a high-precision wind speed soft measurement method for wind power prediction of a wind farm. In view of the nonlinear and strong dynamic characteristics of NWP data, a feature learning network based on VAE is proposed. Meanwhile, the entropy weight method is used to give different weights to each variable loss adaptively, so as to improve the loss function of the original VAE model. Nonlinear feature extraction is performed on the basic framework of the Gaussian mixture model, and a regressor based on the BP neural network between the latent variable and the wind speed is constructed, so as to finally realize the soft measurement of the wind speed and have high prediction accuracy.

Description

Technical Field

[0001] This application belongs to the field of wind power measurement technology, specifically relating to a high-precision soft measurement method for wind speed prediction in wind farms. Background Technology

[0002] Wind energy is a clean and renewable energy source that can be converted into electricity through wind power generation and then transmitted to the power grid. The stability of wind power output is crucial for the safe and stable operation of the power grid. Accurate wind speed prediction provides strong support for wind resource assessment, a basis for wind turbine selection and site selection, and a guarantee for power grid operation, maintenance, and dispatch. However, wind speed exhibits nonlinearity and random fluctuations, which can lead to fluctuations in power generation, affecting the safe and stable operation of the power grid, resulting in low wind power utilization and ultimately impacting power grid quality.

[0003] In related technologies, the main method for measuring wind speed is the direct measurement method based on the anemometer in the wind turbine nacelle. This method is easily affected by the measurement environment, making it difficult to measure the real-time wind speed accurately.

[0004] Therefore, it is necessary to provide a high-precision soft measurement method for wind power prediction in wind farms to solve the above problems. Summary of the Invention

[0005] This application provides a high-precision soft measurement method for wind speed in wind farm wind power prediction. Addressing the nonlinear and highly dynamic characteristics of NWP data, a feature learning network based on VAE is proposed. Nonlinear feature extraction is performed on the basic framework of a Gaussian mixture model. Simultaneously, a BP neural network-based regressor is constructed between latent variables and wind speed, ultimately achieving soft measurement of wind speed with high prediction accuracy.

[0006] To solve the above-mentioned technical problems, this application is implemented as follows:

[0007] A high-precision soft measurement method for wind power prediction in wind farms includes the following steps:

[0008] S1: Collect numerical weather forecast data related to wind farms and wind speed to form an auxiliary variable dataset, and preprocess the auxiliary variable dataset;

[0009] S2: Construct a VAE feature learning network, which includes an encoder and a decoder. The auxiliary variable dataset is fed into the encoder as raw data to extract the latent features of the raw data. The latent features are then fed into the decoder as latent variables for reconstruction to obtain the generated data.

[0010] S3: Construct the original loss function of the VAE feature learning network. The original loss function includes two parts: KL divergence and reconstruction loss. Calculate the entropy weight of each sample in the original data based on the entropy weight method and assign it to the reconstruction loss part of the original loss function. At the same time, assign different weights to the KL divergence part and the reconstruction loss part to obtain the improved loss function. Train the VAE feature learning network by minimizing the improved loss function and save the trained model parameters and latent variables.

[0011] S4: Use the wind speed corresponding to the auxiliary variable dataset as the label of the latent variable, and use it together with the latent variable to form a training set. Use the training set as the input of the regressor, and use the stochastic gradient descent algorithm to update the regressor model parameters to complete the training process.

[0012] S5: For any given soft measurement of wind speed, repeat steps S1-S4 and output the predicted wind speed value.

[0013] Preferably, the "numerical weather forecast data related to wind speed" in step S1 includes short-term weather type, rainfall type, air temperature, perceived temperature, humidity, wind direction, visibility, air pressure, and the weather type of the day.

[0014] Preferably, the "preprocessing" process in step S1 includes: deleting missing values, serializing text data, removing the decimal part of the data, and normalization processing.

[0015] Preferably, in step S2, the extraction process of latent features of the original data specifically involves: calculating the mean and variance of the distribution that the original data follows in the latent space, and using the reparameterization technique to sample the latent features of the original data using the mean and variance of the previous layer.

[0016] Preferably, step S3 specifically includes the following steps:

[0017] S31: Construct the marginal likelihood function of the data generated by the VAE feature learning network, and transform the lower bound of the evidence of the marginal likelihood function into the loss function of the VAE feature learning network to obtain the original loss function:

[0018]

[0019] In the formula, Loss represents the original loss function; x represents the original data; This represents the generated data after reconstruction; Let denote the posterior distribution; p(z) denotes the prior distribution; Let p(z) represent the prior distribution and the posterior distribution. The KL divergence is used to measure the difference between the prior distribution p(z) and the posterior distribution p(z). The degree of approximation; Indicates the reconstruction loss;

[0020] S32: Calculate the i-th sample x under the j-th indicator in the original data. ij The proportion it accounts for, and it is regarded as the probability p used in the calculation of relative entropy. ij The information entropy and information utility value of each indicator are calculated sequentially, and the information utility value is normalized to obtain the entropy weight of each indicator. The calculation process is expressed as follows:

[0021]

[0022] In the formula, e j The information entropy of the j-th indicator is represented by d; the larger the information entropy, the smaller the corresponding amount of information. J represents the dimension of the original data. j W represents the information utility value of the j-th indicator; j This represents the entropy weight of the j-th index;

[0023] S33: Assign the entropy weights of each indicator to the reconstruction loss part of the loss function, and assign weights α and β to the KL divergence and reconstruction loss, respectively, to obtain the improved loss function Loss′, expressed as:

[0024]

[0025] In the formula, α and β are both hyperparameters, and α+β=1.

[0026] The beneficial effects of this application are:

[0027] To address the nonlinear and highly dynamic characteristics of NWP data, a feature learning network based on VAE is proposed. Entropy weighting is used to adaptively assign different weights to each variable loss, thereby improving the loss function of the original VAE model. Nonlinear feature extraction is performed within the basic framework of Gaussian mixture models. Furthermore, a BP neural network-based regressor is constructed to connect latent variables with wind speed, ultimately achieving soft measurement of wind speed with high prediction accuracy. Attached Figure Description

[0028] Figure 1 This is a flowchart illustrating a high-precision soft measurement method for wind power prediction in wind farms provided in this application.

[0029] Figure 2 This diagram illustrates the architecture of the VAE feature learning network provided in this application.

[0030] Figure 3 This graph shows the prediction results of the BP model in Example 1.

[0031] Figure 4This graph shows the prediction results of the PCA_BP model in Example 1.

[0032] Figure 5 This figure shows the prediction results of the VAE_BP model in Example 1. Detailed Implementation

[0033] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0034] Please refer to the following: Figure 1-5 This invention provides a high-precision soft measurement method for wind speed prediction in wind farms, comprising the following steps:

[0035] S1: Collect numerical weather prediction (NWP) data related to wind farms and wind speed to form an auxiliary variable dataset, and preprocess the auxiliary variable dataset.

[0036] "Numerical weather forecast data related to wind speed" includes short-term weather type, precipitation type, temperature, wind chill, humidity, wind direction, visibility, air pressure, and the weather type for the day.

[0037] The "preprocessing" process includes: removing missing values, serializing text data, removing decimal parts of the data, and normalization.

[0038] S2: Construct a VAE feature learning network, which includes an encoder and a decoder. The auxiliary variable dataset is fed into the encoder as raw data to extract latent features from the raw data. The latent features are then fed into the decoder as latent variables for reconstruction to obtain the final generated data.

[0039] Please see Figure 1 , Figure 1 This diagram illustrates the architecture of the VAE (Variational Autoencoder) feature learning network. In this VAE feature learning network, the encoder... It is an inference model, parameterized to have one parameter. The neural network maps data to a low-dimensional space to obtain its latent features. For the decoder p... θ(x|z) is parameterized into another neural network with parameter θ, and the latent features are fed into the encoder as latent variables z. The encoder reconstructs the latent variables z into the original data.

[0040] In step S2, the extraction process of latent features of the original data specifically involves: calculating the mean of the distribution that the original data follows in the latent space. and variance By using reparameterization techniques, the latent features of the original data are obtained by sampling the mean and variance of the previous layer.

[0041] S3: Construct the original loss function of the VAE feature learning network. The original loss function includes two parts: KL divergence and reconstruction loss. Calculate the entropy weight of each sample in the original data based on the entropy weight method and assign it to the reconstruction loss part of the original loss function. At the same time, assign different weights to the KL divergence part and the reconstruction loss part to obtain the improved loss function. Train the VAE feature learning network by minimizing the improved loss function and save the trained model parameters and latent variables.

[0042] Step S3 specifically includes the following steps:

[0043] S31: Construct the marginal likelihood function of the data generated by the VAE feature learning network, and transform the lower bound of the evidence of the marginal likelihood function into the loss function of the VAE feature learning network to obtain the original loss function:

[0044]

[0045] In the formula, Loss represents the original loss function; x represents the original data; This represents the generated data after reconstruction; Let denote the posterior distribution; p(z) denotes the prior distribution; Let p(z) represent the prior distribution and the posterior distribution. The KL divergence is used to measure the difference between the prior distribution p(z) and the posterior distribution p(z). The degree of approximation; This represents the reconstruction loss.

[0046] The training process of the VAE feature learning network involves minimizing its negative variational evidence lower bound, thereby optimizing the parameters of the model network. And θ, finally achieving the goal of feature learning, therefore, after transforming the lower bound of the evidence, the original loss function can be directly obtained.

[0047] The derivation of the original loss function is as follows:

[0048] The marginal likelihood function of the data generated in the VAE feature learning network is expressed as:

[0049]

[0050] Maximizing the marginal likelihood is equivalent to solving:

[0051] Maximum L = ∑ x logp(x);

[0052] in:

[0053]

[0054] Based on the property of the KL divergence in Jensen's inequality that it is always greater than or equal to zero, we can obtain the lower bound of evidence for the likelihood function logp(x).

[0055]

[0056] As can be seen from the above equation, the lower bound of evidence for the log-likelihood function logp(x) includes two terms. The first term is regularization, which guides the approximate posterior distribution to be as close as possible to the prior distribution. The prior distribution p(z) is compared with the variational distribution. The KL divergence is used to represent the prior distribution p(z), which is a standard normal distribution N(0,1); the second term is the reconstruction loss, which is used to ensure the degree of matching between the reconstructed data and the original data.

[0057] Maximize the lower bound of evidence This is equivalent to minimizing the sum of KL divergences and minimizing the reconstruction loss, and it relates to the lower bound of the evidence. The original loss function can be obtained by performing a transformation.

[0058] S32: Calculate the i-th sample x under the j-th indicator in the original data. ij The proportion it accounts for, and it is regarded as the probability p used in the calculation of relative entropy. ij The information entropy and information utility value of each indicator are calculated sequentially, and the information utility value is normalized to obtain the entropy weight of each indicator. The calculation process is expressed as follows:

[0059]

[0060] In the formula, e j The information entropy of the j-th indicator is represented by d; the larger the information entropy, the smaller the corresponding amount of information. J represents the dimension of the original data. j W represents the information utility value of the j-th indicator; j This represents the entropy weight of the j-th index;

[0061] S33: Assign the entropy weights of each indicator to the reconstruction loss part of the loss function, and assign weights α and β to the KL divergence and reconstruction loss, respectively, to obtain the improved loss function Loss', expressed as:

[0062]

[0063] In the formula, α and β are both hyperparameters, and α + β = 1; x represents the original data. This represents the reconstructed original data.

[0064] The degree of variation of various meteorological features in NWP data from different regions varies. For example, some meteorological features remain almost unchanged throughout the year, thus their importance for predicting wind speed is relatively low. If the VAE feature learning network treats all indicators in the NWP data equally, it will be unable to accurately learn the latent feature representations of the original data. Therefore, in the technical solution of this application, different weights are adaptively assigned to the loss of each input variable using the entropy weight method, which enables the VAE feature learning network to better learn the latent features of the original data. Furthermore, since the VAE model was originally used for generating and extracting features from image datasets, considering the differences between the NWP dataset and image datasets, the importance of KL divergence and reconstruction loss for network training is different. Therefore, this application introduces two hyperparameters, α and β, to assign different weights to KL divergence and reconstruction loss, facilitating better network training and improving prediction accuracy.

[0065] S4: Use the wind speed corresponding to the auxiliary variable dataset as the label of the latent variable, and use it together with the latent variable to form a training set. Use the training set as the input of the regressor, and use algorithms such as stochastic gradient descent to update the regressor model parameters to complete the training process.

[0066] S5: For any given soft measurement of wind speed, repeat steps S1-S4 and output the predicted wind speed value.

[0067] Once the predicted wind speed is obtained, the predicted wind power can be calculated. The formula for calculating wind power P is:

[0068]

[0069] In the formula, ρ represents air density; A represents the circular area formed by the rotation of the wind turbine blades; and v represents wind speed.

[0070] In this application's technical solution, a VAE-based feature learning network is established for NWP data. A Gaussian mixture model is used to describe the latent feature distribution of the VAE, and the original data is mapped to the latent space through nonlinear mapping, thereby extracting nonlinear latent variables. Simultaneously, the entropy weight method is used to adaptively assign different weights to each variable's loss, thus improving the loss function of the original VAE model. A backpropagation neural network is used to establish a regression model between the latent variables and the wind speed variable to achieve soft measurement of wind speed. Compared to traditional direct measurement, it is less affected by the measurement environment and can accurately measure real-time wind speed.

[0071] During the training process of the VAE feature learning network, in order to intuitively evaluate the measurement performance of the soft measurement model, this application uses the root mean square error (RMSE) and the squared correlation R between the actual value and the predicted value. 2 Two metrics are used to quantify the model's predictive performance: RMSE and R. 2 The definition is as follows:

[0072]

[0073] In the formula, I represents the number of samples; y i This represents the actual output value of the i-th sample; It is the predicted output value of the i-th sample; It is the average of the actual output values ​​of all samples; the smaller the RMSE, the better the R. 2 The closer a value is to 1, the better the soft measurement performance of the model.

[0074] As a meteorological dataset, NWP data contains meteorological variables with nonlinear and highly dynamic characteristics. Directly using it for wind speed regression prediction yields poor results for soft measurement. However, VAE feature learning networks can effectively capture the latent features affecting wind speed changes, and utilizing these latent features for regression prediction can significantly improve the accuracy of soft measurement.

[0075] Furthermore, considering that the degree of variation of various meteorological characteristics in NWP data varies across different regions—for example, some meteorological characteristics remain almost unchanged throughout the year—the importance of these characteristics for predicting wind speed is relatively low. If the VAE feature learning network treats all NWP indicators equally, it will be unable to learn effective latent feature representations of the original data. The entropy weight method, by adaptively assigning different weights to the loss of each input variable and improving the loss function of the original VAE model, can effectively solve this problem.

[0076] S5: For any given soft measurement of wind speed, repeat steps S1-S4 and output the predicted wind speed value.

[0077] Example 1

[0078] This paper uses numerical weather prediction (NWP) data and measured wind speed data obtained hourly from the Segede wind farm in Belgium from 2006 to 2016 as the dataset for simulation testing. The NWP dataset is shown in Table 1.

[0079] Table 1. Auxiliary and dominant variables of NWP data

[0080]

[0081]

[0082] The Szeged wind farm in Belgium is located in a temperate maritime climate zone, characterized by highly variable weather, sometimes experiencing all four seasons in a single day. Therefore, the NWP data from this wind farm is a typical example of NWP data with nonlinear and strong dynamic characteristics, which is helpful in verifying the reliability of this application.

[0083] After data preprocessing, the auxiliary variables selected in this application are easy to measure and predict in the short term, and are related to the changes in the dominant variable (wind speed).

[0084] 5000 samples were used to construct the soft sensor model, with 175 samples serving as the test set and the remaining samples as the training set. To demonstrate the reliability of the proposed algorithm (VAE_BP), a BP neural network model (BP) and a soft sensor model based on principal component analysis (PCA_BP) were constructed and their performance compared. Through trial and error parameter tuning, the encoder of VAE_BP was ultimately set to a single-layer neural network with 32 hidden layer neurons and a latent variable dimension of 7. The decoder structure was symmetrical to the encoder. The regressor was set to a single-layer BP neural network with 32 hidden layer neurons. The parameters of the BP neural network model were consistent with those of the regressor in the VAE_BP model. The latent variable dimension in the PCA_BP model was also set to 7, and the regressor structure and parameters were consistent with those of the VAE_BP regressor.

[0085] The data samples were fed into the VAE_BP, BP, and PCA_BP models for training, and the performance evaluation metrics of the three models were obtained, as shown in Table 2:

[0086] Table 2 Model Performance Evaluation Indicators

[0087] <![CDATA[R 2 ]]> 0.55 -0.03 0.73 RMSE 40.50 71.19 33.72

[0088] Comparing Table 2, it can be seen that the VAE_BP model improves the accuracy of wind speed soft measurement based on the BP neural network model, while the PCA_BP model, with its linear feature extraction method, actually reduces the accuracy of wind speed soft measurement. Please refer again. Figure 3-5 , Figure 1 A graph showing the prediction results of the BP model; Figure 2 The graph represents the prediction results of the PCA_BP model; Figure 3 The image shows the prediction results of the VAE_BP model. A comparison with reference 3-5 shows that the prediction accuracy of the VAE_BP model is significantly improved compared to the BP model. Although the original BP neural network method also has good prediction capabilities, its predictions fluctuate greatly and cannot accurately capture the changing trends and characteristics of wind speed. Furthermore, the PCA_BP model, a soft measurement method based on principal component analysis feature extraction, not only failed to improve the prediction accuracy of the original BP neural network method but actually reduced it, with its prediction results deviating from the true values ​​overall. This indicates that linear feature extraction methods like principal component analysis cannot accurately extract data with nonlinear and highly dynamic characteristics, such as NWP. The VAE_BP model provided in this application, however, can effectively extract the latent features of NWP data, thus achieving relatively high soft measurement accuracy even with a low regressor model complexity. This not only ensures the model's soft measurement accuracy but also its generalization ability.

[0089] The soft measurement method provided in this application can be applied to ultra-short-term wind speed prediction, short-term wind speed prediction, and medium-term wind speed prediction, simply by replacing the input NWP data with ultra-short-term, short-term, and medium-term data in sequence.

[0090] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A high-precision soft measurement method for wind speed in wind farm power prediction, characterized in that, Includes the following steps: S1: Collect numerical weather forecast data related to wind farms and wind speed to form an auxiliary variable dataset, and preprocess the auxiliary variable dataset; S2: Construct a VAE feature learning network, which includes an encoder and a decoder. The auxiliary variable dataset is fed into the encoder as raw data to extract the latent features of the raw data. The latent features are then fed into the decoder as latent variables for reconstruction to obtain the generated data. S3: Construct the original loss function of the VAE feature learning network. The original loss function includes two parts: KL divergence and reconstruction loss. Calculate the entropy weight of each sample in the original data based on the entropy weight method and assign it to the reconstruction loss part of the original loss function. At the same time, assign different weights to the KL divergence part and the reconstruction loss part to obtain the improved loss function. Train the VAE feature learning network by minimizing the improved loss function and save the trained model parameters and latent variables. S4: Use the wind speed corresponding to the auxiliary variable dataset as the label of the latent variable, and use it together with the latent variable to form a training set. Use the training set as the input of the regressor, and use the stochastic gradient descent algorithm to update the regressor model parameters to complete the training process. S5: For any given soft measurement of wind speed, repeat steps S1-S4 and output the predicted wind speed value.

2. The high-precision soft-measurement method for wind power prediction in wind farms according to claim 1, characterized in that, The "numerical weather forecast data related to wind speed" in step S1 includes short-term weather type, rainfall type, temperature, perceived temperature, humidity, wind direction, visibility, air pressure, and the weather type of the day.

3. The high-precision soft measurement method for wind power prediction in wind farms according to claim 1, characterized in that, The "preprocessing" process in step S1 includes: deleting missing values, serializing text data, removing the decimal part of the data, and normalization.

4. The high-precision soft measurement method for wind power prediction in wind farms according to claim 1, characterized in that, In step S2, the extraction process of latent features of the original data is as follows: calculate the mean and variance of the distribution that the original data follows in the latent space, and obtain the latent features of the original data by sampling using the mean and variance of the previous layer through the reparameterization technique.

5. The high-precision soft-measurement method for wind power prediction in wind farms according to claim 1, characterized in that, Step S3 specifically includes the following steps: S31: Construct the marginal likelihood function of the data generated by the VAE feature learning network, and transform the lower bound of the evidence of the marginal likelihood function into the loss function of the VAE feature learning network to obtain the original loss function: ； In the formula, Represents the original loss function; Represents the original data; This represents the generated data after reconstruction; Indicate the posterior distribution; Represent the prior distribution; Represent the prior distribution With posterior distribution The KL divergence is used to measure the prior distribution. and posterior distribution The degree of approximation; Indicates the reconstruction loss; S32: Calculate the first... in the original data. j The first item under the indicator i Sample The proportion it accounts for, and it is regarded as the probability used in the calculation of relative entropy. The information entropy and information utility value of each indicator are calculated sequentially, and the information utility value is normalized to obtain the entropy weight of each indicator. The calculation process is expressed as follows: ； In the formula, Indicates the first j The information entropy of each indicator is such that the higher the information entropy, the smaller the corresponding amount of information. J represents the dimension of the original data; Indicates the first j The information utility value of each indicator; Indicates the first j Entropy weight of each indicator; S33: Assign entropy weights to the reconstruction loss component of the loss function for each metric, and simultaneously assign weights to the KL divergence and reconstruction loss. and The improved loss function is obtained. , is represented as: ； In the formula, and All are hyperparameters, and .

Images