Wind power plant monitoring data automatic restoration method based on signal frequency domain feature reconstruction

Abstract

The invention belongs to the technical field of wind power plant monitoring data restoration, and particularly discloses a wind power plant monitoring data automatic restoration method based on signal frequency domain feature reconstruction, comprising the following steps: acquiring to-be-restored wind power plant monitoring data, and randomly generating a noise vector; constructing a generative adversarial network model; inputting the noise vector into a generator of a generative adversarial network model to obtain generated data, and calculating loss of the generated data and real wind power plant monitoring data; adjusting a noise vector by alternately using an optimizer strategy to obtain minimum loss of generated data and real wind power plant monitoring data, and performing iterative optimization for multiple times to obtain an optimal noise vector; and inputting the optimal noise vector into the generator, and outputting the missing part of the wind power plant monitoring data to complete the restoration of the wind power plant monitoring data. According to the method, the problem that high-frequency detail information of the data is ignored in an existing wind power data reconstruction and restoration method is solved, and the accuracy of data reconstruction and filling is improved.

Description

Technical Field

[0001] The invention belongs to the technical field of wind farm monitoring data repair, and in particular relates to a design of an automatic repair method for wind farm monitoring data based on signal frequency domain feature reconstruction. Background Art

[0002] Wind power generation is intermittent, volatile and random. To ensure the safe and stable operation of wind power grid connection, it is crucial to monitor and record the operating data of various parameters of wind turbines and wind farms. The data quality directly affects the operation of power grids, planning and scheduling, and power market evaluation. However, due to measurement errors, transmission, control, unplanned maintenance, wind power curtailment and power rationing, wind farm monitoring data usually contains a large amount of abnormal data. Direct removal will cause the integrity of the time series and the abundance of data to be damaged to a certain extent. In order to reconstruct and repair complete wind power data, most existing methods combine autoencoders and generative adversarial networks, optimize and improve the network structure of the generator and discriminator, apply spatial domain loss functions, and achieve filling by mining the time domain characteristics of data. However, since neural networks will give priority to learning low-frequency components during the training process, they often ignore the high-frequency details of the data. The lack of detailed information will lead to a decrease in data quality, which makes the filling results less realistic and leads to poor filling effects. In addition, since the input low-dimensional vector of the generator is randomly sampled from a normal distribution, the generated new time series data may still be very different from the original data. That is, when using random data to generate, the false data generated by the generator is very random and volatile, resulting in poor filling effect. Summary of the invention

[0003] The purpose of the present invention is to solve the problems that the existing wind power data reconstruction and repair methods ignore the high-frequency detail information of the data, and the false data generated by the generator is very random and volatile, resulting in poor filling effect. An automatic repair method for wind farm monitoring data based on signal frequency domain feature reconstruction is proposed.

[0004] The technical solution of the present invention is as follows: In a first aspect, a method for automatically repairing wind farm monitoring data based on signal frequency domain feature reconstruction comprises the following steps:

[0005] S1. Obtain monitoring data of the wind farm to be repaired and randomly generate a noise vector;

[0006] S2. Construct a generative adversarial network and train the generative adversarial network to obtain a generative adversarial network model;

[0007] S3. Input the noise vector into the generator of the generative adversarial network model to obtain generated data, and calculate the loss between the generated data and the real wind farm monitoring data;

[0008] S4. By alternately using the optimizer strategy, the noise vector is adjusted to minimize the loss between the generated data and the real wind farm monitoring data, and the optimization is iterated multiple times to obtain the optimal noise vector;

[0009] S5. Input the optimal noise vector into the generator, output the missing part of the wind farm monitoring data, and complete the wind farm monitoring data repair.

[0010] The beneficial effects of the present invention are:

[0011] 1. The present invention adjusts the input noise vector so that the generated data becomes closer to the real data under a specific loss function, thereby providing a higher quality filling result.

[0012] 2. When designing the loss function, the present invention starts from the frequency domain perspective, avoiding the problem of loss of high-frequency details caused by optimization only in the spatial domain, and improving the accuracy of data reconstruction and filling.

[0013] 3. The present invention searches for the optimal noise variable in the latent variable space through gradient descent, so that the data generated by the generator can best match the known part of the missing data, efficiently explores the latent variable space, and improves the quality and accuracy of the generated data.

[0014] Preferably, the step S2 specifically includes the following sub-steps:

[0015] S21. Obtaining original data samples of wind farm operation, generating random noise vectors, and dividing the original data samples into a training set and a test set;

[0016] S22. Input the random noise vector into the generator of the generative adversarial network for training, and output the generated data;

[0017] S23. Input the generated data and the training set data into the discriminator of the generative adversarial network and output the discrimination result;

[0018] S24. Feedback the discrimination result to the generator and the discriminator, update the parameters of the generator and the parameters of the discriminator, and execute step S25;

[0019] S25. Repeat steps S22 to S24 until the generative adversarial network reaches a Nash equilibrium, fix the parameters of the generator and the parameters of the discriminator, and obtain a trained generative adversarial network;

[0020] S26. Input the test set data into the trained generative adversarial network for testing and optimization to obtain a generative adversarial network model.

[0021] The beneficial effects of the above preferred solution are:

[0022] Through the above-mentioned preferred scheme, the data generated by the constructed generative adversarial network model can best match the known part of the missing data, efficiently explore the latent variable space, and improve the quality and accuracy of the generated data.

[0023] Preferably, the generator in step S22 is constructed using a one-dimensional convolutional neural network; the one-dimensional convolutional neural network includes a deconvolution layer, a batch normalization layer, a ReLU activation function layer, four custom layers and a convolution layer connected in sequence;

[0024] The four custom layers each include a deconvolution layer, a batch normalization layer, and a ReLU activation function layer that are connected in sequence.

[0025] The beneficial effects of the above preferred solution are:

[0026] The generator constructed by the one-dimensional convolutional neural network can better capture the temporal relationship in the data. The batch normalization layer can speed up the training and improve the stability. The nonlinearity introduced by the ReLU activation function layer enables the network to learn complex features. The transposed convolution layer further improves the performance of the network, and the generated samples are more detailed and realistic.

[0027] Preferably, the discriminator in step S23 comprises: four fully connected layers connected in sequence; the first three layers of the four fully connected layers use the LeakyReLU function as the activation function, and the last fully connected layer uses the Sigmoid function as the activation function.

[0028] Preferably, the method further comprises:

[0029] The four fully connected layers are spectrally normalized, specifically, by calculating the spectral norm of the weight matrix of each layer, and dividing the weight matrix by its spectral norm.

[0030] The beneficial effects of the above preferred solution are:

[0031] By using the LeakyReLU function and the Sigmoid function as the activation functions of the fully connected layer and performing spectral normalization on the four fully connected layers, the Lipschitz continuity of the network layer can be maintained, reducing the possibility of gradient disappearance or explosion, thereby ensuring that the gradient and training of the network are more stable.

[0032] Preferably, in step S24, the parameters of the generator are updated based on the loss function and gradient descent algorithm of the generator; the parameters of the discriminator are updated based on the loss function and gradient descent algorithm of the discriminator.

[0033] The beneficial effects of the above preferred solution are:

[0034] The gradient descent algorithm can minimize the loss function of the generator and maximize the loss function of the discriminator, and can efficiently and accurately iterate and solve the parameters of the generator and discriminator.

[0035] Preferably, the loss function of the generator includes the original loss function of the generator and a focus frequency loss algorithm;

[0036] The formula of the original loss function Loss_G_adv is:

[0037]

[0038] in, represents the expectation of the noise distribution, D(·) represents the probability that the discriminator determines that the generated data is real, G(·) represents the generated data of the generator, and z represents the random noise vector;

[0039] The formula of the focus frequency loss algorithm FFL is:

[0040]

[0041] Where M represents the number of rows of samples, N represents the number of columns of samples, (u, v) represents the spectral coordinates, w(u, v) represents the weight matrix of the spatial frequency at (u, v), and F r (u,v) represents the spatial frequency value at the spectrum coordinate (u,v) of the real sample, and its complex expression is F r (u,v)=a r +b r i, a r Indicates F r The real part of (u,v), b r Indicates F r The imaginary part of (u,v), i represents the imaginary unit, F f (u,v) represents the spatial frequency value at the spectral coordinate (u,v) of the generated pseudo sample, and its complex expression is F f (u,v)=a f +b f i means, a f Indicates F f The real part of (u,v), b f Indicates F f The imaginary part of (u,v), Represents the frequency distance between the real sample and the generated fake sample;

[0042] The formula of the generator's loss function Loss_G is:

[0043] Loss_G=Loss_G_adv+FFL.

[0044] The beneficial effects of the above preferred solution are:

[0045] The above preferred scheme combines the focused frequency loss algorithm into the original generator loss function. The loss algorithm dynamically adjusts the loss weights in the frequency domain through the spectrum weight matrix, so that the generated padded data is more balanced in each frequency component.

[0046] As a preference,

[0047] in, represents the expectation of the probability distribution of the real sample, represents the expectation of the probability distribution of generated samples, x represents the real sample, Represents the generated data of a generator.

[0048] Preferably, the loss between the generated data and the real wind farm monitoring data in step S3 is calculated by a filling loss function;

[0049] The formula of the filling loss function Loss_z is:

[0050]

[0051] Among them, N represents the number of sample columns; n represents the number of samples; j represents the jth sample; x represents the real sample, Represents the generated data of a generator.

[0052] The beneficial effects of the above preferred solution are:

[0053] By filling the loss function, the mean square error between the generated samples and the real samples is calculated, the loss is back-propagated, and the noise vector is adjusted by the optimizer to find the optimal noise vector in the latent variable space, so that the data generated by the generator can best match the known part of the missing data.

[0054] Preferably, the strategy of alternating the optimizer in step S4 is specifically: alternating the use of Adam and RMSprop optimizers in odd and even iterations.

[0055] The beneficial effects of the above preferred solution are:

[0056] An alternating optimizer strategy is adopted. Adam and RMSprop optimizers are used alternately through odd and even iterations to adaptively adjust the learning rate of each parameter, helping the model to converge faster and avoid falling into the local optimal solution, thereby finding the optimal noise vector in the latent variable space. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 Shown is a flow chart of an automatic repair method for wind farm monitoring data based on signal frequency domain feature reconstruction.

[0058] Figure 2 Shown is the basic structure diagram of the generative adversarial network. DETAILED DESCRIPTION

[0059] Now, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings. It should be understood that the embodiments shown and described in the accompanying drawings are only exemplary and are intended to explain the principles and spirit of the present invention, rather than to limit the scope of the present invention.

[0060] Embodiment 1, as Figure 1 As shown, a method for automatically repairing wind farm monitoring data based on signal frequency domain feature reconstruction includes the following steps:

[0061] S1. Obtain monitoring data of the wind farm to be repaired and randomly generate a noise vector;

[0062] S2. Construct a generative adversarial network and train the generative adversarial network to obtain a generative adversarial network model;

[0063] S3. Input the noise vector into the generator of the generative adversarial network model to obtain generated data, and calculate the loss between the generated data and the real wind farm monitoring data;

[0064] S4. By alternately using the optimizer strategy, the noise vector is adjusted to minimize the loss between the generated data and the real wind farm monitoring data, and the optimization is iterated multiple times to obtain the optimal noise vector;

[0065] S5. Input the optimal noise vector into the generator, output the missing part of the wind farm monitoring data, and complete the wind farm monitoring data repair.

[0066] The Generative Adversarial Networks (GAN) consists of a generator (G) and a discriminator (D). Through training with a large amount of sample data, the generation ability of the generator and the discrimination ability of the discriminator are gradually improved in the confrontation. The ultimate goal is to make the data generated by the generator as consistent as possible with the characteristics of the real sample. Figure 2 As shown, the probability distribution of the real data set x is p data , the probability distribution of the random noise vector z is p z (z), when the input is a random noise vector z, the probability distribution of the data G(z) generated by the generator is p g The random noise vector z is input into the generator, and through continuous training, the probability distribution of generated data p is g Fit the probability distribution p of the real data as much as possible data, and the discriminator is used to determine whether the input sample is real or generated, and the discrimination result is used to feed back to the generator and discriminator for parameter update. The networks interact and train with each other until the entire network reaches Nash equilibrium. The training process is transformed into a mathematical model maximization and minimization problem, and the loss function of the generative adversarial network is:

[0067]

[0068] in, The expectation of the probability distribution of the real sample; D(·) represents the discriminant result; p d (x) represents the true data distribution; V(D,G) represents the optimization function of the generative adversarial network.

[0069] In this embodiment, the generator in step S22 is constructed using a one-dimensional convolutional neural network; the one-dimensional convolutional neural network includes a deconvolution layer, a batch normalization layer, a ReLU activation function layer, four custom layers and a convolution layer connected in sequence; the four custom layers each include a deconvolution layer, a batch normalization layer, and a ReLU activation function layer connected in sequence.

[0070] The generator first performs upsampling through a deconvolution layer to convert low-dimensional input features into high-dimensional features, thereby generating high-resolution data. Batch normalization layers are used in the network to speed up training and improve stability, and ReLU activation function layers are used to introduce nonlinearity, allowing the network to learn complex features. Repeated transposed convolution layers further enhance the performance of the network, and the generated samples are more detailed and realistic. Finally, convolution layers are used to map high-dimensional features to the output of the generated samples.

[0071] In this embodiment, the discriminator in step S23 includes: four fully connected layers connected in sequence; the first three layers of the four fully connected layers use the LeakyReLU function as the activation function, and the last fully connected layer uses the Sigmoid function as the activation function.

[0072] In this embodiment, the method further includes:

[0073] The four fully connected layers are spectrally normalized, specifically, by calculating the spectral norm of the weight matrix of each layer, and dividing the weight matrix by its spectral norm.

[0074] In this embodiment, the parameters of the generator in step S24 are updated based on the loss function and gradient descent algorithm of the generator; the parameters of the discriminator are updated based on the loss function and gradient descent algorithm of the discriminator.

[0075] In this embodiment, the loss function of the generator includes the original loss function of the generator and the focus frequency loss algorithm;

[0076] The original loss function Loss_G_adv is the term about the random noise vector z in the adversarial loss V(D,G), and the formula of the original loss function Loss_G_adv is:

[0077]

[0078] in, represents the expectation of the noise distribution, D(·) represents the probability that the discriminator determines that the generated data is real, G(·) represents the generated data of the generator, and z represents the random noise vector;

[0079] The focus frequency loss algorithm FFL is obtained by performing Hadamard product calculation on the spectrum weight matrix and the frequency distance matrix. The formula of the focus frequency loss algorithm FFL is:

[0080]

[0081] Where M represents the number of rows of samples, N represents the number of columns of samples, (u, v) represents the spectral coordinates, w(u, v) represents the weight matrix of the spatial frequency at (u, v), and F r (u,v) represents the spatial frequency value at the spectrum coordinate (u,v) of the real sample, and its complex expression is F r (u,v)=a r +b r i, a r Indicates F r The real part of (u,v), b r Indicates F r The imaginary part of (u,v), i represents the imaginary unit, F f (u,v) represents the spatial frequency value at the spectral coordinate (u,v) of the generated pseudo sample, and its complex expression is F f (u,v)=a f +b f i means, a f Indicates F f The real part of (u,v), b f Indicates F f The imaginary part of (u,v), Represents the frequency distance between the real sample and the generated fake sample;

[0082] The formula of the generator's loss function Loss_G is:

[0083] Loss_G=Loss_G_adv+FFL.

[0084] The training goal of the generator is to generate samples as close to real samples as possible, that is, to minimize its loss function and continuously update the parameters of the generator through the gradient descent algorithm.

[0085] In this embodiment, the loss function of the discriminator is the term in the adversarial loss V(D, G) with respect to the real sample x. The formula of the loss function of the discriminator is:

[0086]

[0087] in, represents the expectation of the probability distribution of the real sample, represents the expectation of the probability distribution of generated samples, x represents the real sample, Represents the generated data of the generator; the goal of the discriminator is to better distinguish the generated samples from the real samples, that is, to maximize the loss function and continuously update the parameters of the discriminator through the gradient descent algorithm.

[0088] In this embodiment, step S7 specifically includes the following steps:

[0089] S71. Obtain monitoring data of the wind farm to be repaired and randomly generate a noise vector;

[0090] S72. Input the noise vector into the generator to obtain generated data, and calculate the loss between the generated data and the real wind farm monitoring data;

[0091] S73. By alternately using the optimizer strategy, adjusting the noise vector, minimizing the loss between the generated data and the real wind farm monitoring data, and iterating the optimization multiple times to obtain the optimal noise vector;

[0092] S74. Input the optimal noise vector into the generator, output the missing part of the wind farm monitoring data, and complete the wind farm monitoring data repair.

[0093] In this embodiment, the loss between the generated data and the real wind farm monitoring data in step S3 is calculated by filling in the loss function;

[0094] The formula of the filling loss function Loss_z is:

[0095]

[0096] Among them, N represents the number of sample columns; n represents the number of samples; j represents the jth sample; x represents the real sample, Represents the generated data of a generator.

[0097] In this embodiment, the strategy of alternately using the optimizer in step S4 is specifically: alternately using the Adam and RMSprop optimizers in odd and even iterations.

[0098] The present invention combines the focused frequency loss algorithm into the original generator loss function, and dynamically adjusts the loss weight of the frequency domain through the spectrum weight matrix, so that the generated filling data is more balanced in each frequency component, better captures the detailed information in the data, and improves the quality of the generated data. At the same time, the gradient descent algorithm is improved by using the alternate optimizer strategy, and the input noise vector is adjusted through the algorithm, so that the generated data becomes closer to the real data under a specific loss function, providing a higher quality filling result.

[0099] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A method for automatically repairing wind farm monitoring data based on signal frequency domain feature reconstruction, characterized in that: The method comprises the following steps: S1. Obtain monitoring data of the wind farm to be repaired and randomly generate a noise vector; S2. Construct a generative adversarial network and train the generative adversarial network to obtain a generative adversarial network model; S3. Input the noise vector into the generator of the generative adversarial network model to obtain generated data, and calculate the loss between the generated data and the real wind farm monitoring data; S4. By alternately using the optimizer strategy, the noise vector is adjusted to minimize the loss between the generated data and the real wind farm monitoring data, and the optimization is iterated multiple times to obtain the optimal noise vector; S5. Input the optimal noise vector into the generator, output the missing part of the wind farm monitoring data, and complete the wind farm monitoring data repair.

2. The method for automatically repairing wind farm monitoring data based on signal frequency domain feature reconstruction according to claim 1 is characterized in that: The step S2 specifically includes the following sub-steps: S21. Obtaining original data samples of wind farm operation, generating random noise vectors, and dividing the original data samples into a training set and a test set; S22. Input the random noise vector into the generator of the generative adversarial network for training, and output the generated data; S23. Input the generated data and the training set data into the discriminator of the generative adversarial network and output the discrimination result; S24. Feedback the discrimination result to the generator and the discriminator, update the parameters of the generator and the parameters of the discriminator, and execute step S25; S25. Repeat steps S22 to S24 until the generative adversarial network reaches a Nash equilibrium, fix the parameters of the generator and the parameters of the discriminator, and obtain a trained generative adversarial network; S26. Input the test set data into the trained generative adversarial network for testing and optimization to obtain a generative adversarial network model.

3. The method for automatically repairing wind farm monitoring data based on signal frequency domain feature reconstruction according to claim 2 is characterized in that: The generator in step S22 is constructed using a one-dimensional convolutional neural network; the one-dimensional convolutional neural network includes a deconvolution layer, a batch normalization layer, a ReLU activation function layer, four custom layers and a convolution layer connected in sequence; The four custom layers each include a deconvolution layer, a batch normalization layer, and a ReLU activation function layer that are connected in sequence.

4. The method for automatically repairing wind farm monitoring data based on signal frequency domain feature reconstruction according to claim 2 is characterized in that: The discriminator in step S23 includes: four fully connected layers connected in sequence; the first three layers of the four fully connected layers use the LeakyReLU function as the activation function, and the last fully connected layer uses the Sigmoid function as the activation function.

5. The method for automatically repairing wind farm monitoring data based on signal frequency domain feature reconstruction according to claim 4 is characterized in that: The method further comprises: The four fully connected layers are spectrally normalized, specifically, by calculating the spectral norm of the weight matrix of each layer, and dividing the weight matrix by its spectral norm.

6. The method for automatically repairing wind farm monitoring data based on signal frequency domain feature reconstruction according to claim 2 is characterized in that: In step S24, the parameters of the generator are updated based on the loss function and gradient descent algorithm of the generator; the parameters of the discriminator are updated based on the loss function and gradient descent algorithm of the discriminator.

7. The method for automatically repairing wind farm monitoring data based on signal frequency domain feature reconstruction according to claim 6 is characterized in that: The loss function of the generator includes the original loss function of the generator and the focus frequency loss algorithm; The formula of the original loss function Loss_G_adv is: in, represents the expectation of the noise distribution, D(·) represents the probability that the discriminator determines that the generated data is real, G(·) represents the generated data of the generator, and z represents the random noise vector; The formula of the focus frequency loss algorithm FFL is: Where M represents the number of rows of samples, N represents the number of columns of samples, (u, v) represents the spectral coordinates, w(u, v) represents the weight matrix of the spatial frequency at (u, v), and F r (u,v) represents the spatial frequency value at the spectrum coordinate (u,v) of the real sample, and its complex expression is F r (u,v)=a r +b r i, a r Indicates F r The real part of (u,v), b r Indicates F r The imaginary part of (u,v), i represents the imaginary unit, F f (u,v) represents the spatial frequency value at the spectral coordinate (u,v) of the generated pseudo sample, and its complex number expression is F f (u,v)=a f +b f i means, a f Indicates F f The real part of (u,v), b f Indicates F f The imaginary part of (u,v), Represents the frequency distance between the real sample and the generated fake sample; The formula of the generator's loss function Loss_G is: Loss_G=Loss_G_adv+FFL.

8. The method for automatically repairing wind farm monitoring data based on signal frequency domain feature reconstruction according to claim 6 is characterized in that: The formula of the loss function of the discriminator is: in, represents the expectation of the probability distribution of the real sample, represents the expectation of the probability distribution of generated samples, x represents the real sample, Represents the generated data of a generator.

9. The method for automatically repairing wind farm monitoring data based on signal frequency domain feature reconstruction according to claim 8, characterized in that: The loss between the generated data and the real wind farm monitoring data in step S3 is calculated by a filling loss function; The formula of the filling loss function Loss_z is: Among them, N represents the number of columns of samples; n represents the number of samples; j represents the jth sample; x represents the real sample, Represents the generated data of a generator.

10. The method for automatically repairing wind farm monitoring data based on signal frequency domain feature reconstruction according to claim 8, characterized in that: The strategy of alternating the optimizer in step S4 is specifically: alternating between using the Adam and RMSprop optimizers in odd and even iterations.

Images