Multi-feature fusion-based generative adversarial network for ultra-short-term wind power prediction

Abstract

The application provides a multi-feature fusion generative adversarial network ultra-short-term wind power prediction method, which combines variational mode decomposition and attention mechanism and long short-term memory neural network as a generator, and uses a convolutional neural network as a discriminator; uses W distance as the objective function of the generator, and uses the Lipschitz continuity condition to limit the discriminator to improve the stability of the model; and adopts a technical route of inversely deducing the whole field power by single machine prediction to make up for the accuracy of wind power prediction when the wind tower data cannot reflect the whole field wind resource situation. Through the application, the accuracy of the ultra-short-term wind power prediction can be improved, and problems such as great fluctuation of wind power, different contribution rates of multiple factors influencing the wind power to the ultra-short-term wind power prediction, and non-convergence and instability of the original generative adversarial network model are solved.

Description

Technical Field

[0001] This invention relates to the field of wind power generation technology, specifically to a multi-feature fusion generative adversarial network method for ultra-short-term wind power prediction. Background Technology

[0002] In new power systems led by new energy sources, the installed capacity of new energy sources is increasing, and wind power is a crucial pillar of the new energy industry. However, due to the intermittent and uncertain nature of wind power output, large-scale grid connection of wind power can affect the safety and stability of the power grid. Accurate ultra-short-term wind power forecasting plays an irreplaceable role in solving grid peak-shaving problems and formulating reasonable dispatch strategies. With the increase in grid-connected wind power capacity, the power grid has placed higher demands on ultra-short-term wind power forecasting, and wind farms face substantial penalties under dual detailed assessment rules each year, with penalties for ultra-short-term wind power forecasting being the most prominent. Especially in mountainous wind farms with complex terrain, the significant microclimate phenomena and high turbulence intensity in different sections of the wind farm lead to large fluctuations in wind power, making it even more difficult to meet the accuracy requirements of ultra-short-term wind power forecasting.

[0003] Due to the nonlinear and non-stationary nature of wind power sequences, improving prediction accuracy is quite challenging. Scholars both domestically and internationally have conducted extensive research on this topic, with the most common algorithms including the Auto-Regressive and Moving Average Model (ARMA), Long Short-Term Memory (LSTM) network, and CNN. Some scientists have proposed a combined model based on Complementary Empirical Mode Decomposition (CEEMD), an improved Genetic Algorithm (IGA) optimized TS fuzzy neural network, and Markov error correction to improve the accuracy of ultra-short-term wind power prediction. Others have proposed a novel hybrid deep learning model, consisting of convolutional layers, recurrent unit layers (GRU), and fully connected neural networks. Convolutional layers can automatically learn complex features from raw data, while GRU layers can directly learn multiple parallel input data sequences, improving the accuracy of ultra-short-term wind power prediction for wind farms. Another group of scientists proposed an ultra-short-term wind power prediction method that uses a random forest algorithm to correct the prediction error of a bidirectional gated recurrent network model. This method uses a bidirectional gated recurrent unit network model and a random forest algorithm to extract and construct an error model, and uses a kernel density estimation method to fit the probability distribution of the corrected error. The effectiveness and applicability of the method are verified by predicting wind power at multiple time scales using data from a wind farm.

[0004] In recent years, Generative Adversarial Networks (GANs) have become an important branch of deep learning due to their unique adversarial approach and unsupervised learning methods. GANs also possess a strong ability to capture deep, implicit relationships between data points, leading to their increasing application in time series forecasting. Currently, some researchers use GANs to predict high-frequency stock markets, effectively improving prediction accuracy and reducing errors. Others have proposed a financial time series forecasting approach based on Empirical Mode Decomposition (EMD) GANs, utilizing EMD to analyze financial time series and improve the objective function to enhance prediction accuracy.

[0005] In summary, to achieve accuracy, universality, and overcome the randomness, nonlinearity, and stationarity of wind power prediction, scholars both domestically and internationally have proposed various advanced prediction algorithm models. However, existing prediction algorithms still have the following shortcomings: EMD is prone to mode aliasing, resulting in poor prediction performance; the original GAN ​​model suffers from low prediction accuracy, model instability, and inability to reach convergence; the influence of wind direction features on the accuracy of wind power prediction is not significant due to the real-time yaw of the wind turbine; and a good multi-feature fusion strategy is not adopted when processing multiple features that affect wind power. Summary of the Invention

[0006] The purpose of this invention is to propose a multi-feature fusion generative adversarial network (GAN) method for ultra-short-term wind power prediction. This method is based on VMD decomposition to reduce noise and non-stationarity of wind power sequences. At the same time, it uses the AM mechanism to dynamically adjust and optimize the weights of multiple features. Finally, it uses the adversarial idea of ​​WGAN and its ability to capture deep implicit relationships between data to predict wind power, thereby improving the prediction accuracy of the model. Furthermore, through field data analysis, the impact of yaw error features on wind power prediction is proposed.

[0007] This invention provides a multi-feature fusion generative adversarial network method for ultra-short-term wind power prediction, comprising:

[0008] Obtain historical power data and corresponding weather data of the wind turbine to be predicted, perform data preprocessing, and use the preprocessed data as training data.

[0009] A short-term wind power prediction model for wind turbines to be predicted is constructed. The model is trained using training data. The model uses variational mode decomposition combined with attention mechanism and long short-term memory neural network as generator, convolutional neural network as discriminator, and W distance as objective function of generator.

[0010] The real-time power data and weather data of the wind turbine to be predicted are input into the trained ultra-short-term wind power prediction model of the wind turbine to be predicted, and the model output is used as the real-time ultra-short-term wind power prediction result.

[0011] The data preprocessing steps include:

[0012] Outlier removal and imputation are performed on the historical power data of the wind turbines to be predicted and the corresponding weather data to form a standardized original dataset, which is used as training data.

[0013] In the step of performing variational mode decomposition on the power data column in the original dataset, a variational model is constructed through iterative search, decomposing the original wind power time series P(t) in the training data into components u of different frequencies. k (t), corresponding to the center frequency ω k The optimal solution of the variational model is found by updating iteratively in an alternating manner; where,

[0014] The variational mode decomposition constraint expression is:

[0015]

[0016] In the formula, K is the number of modes in the decomposition, {u k}、{ω k} represents the k-th modal component and center frequency, respectively, δ(t) is the Dirac function, and * is the convolution operator.

[0017] In this process, the Lagrange multiplication operator λ and the quadratic penalty factor α are introduced to transform the constrained variational problem of equation (1) into an unconstrained variational problem for solution, resulting in the constrained expression of the wind power output signal VMD as follows:

[0018]

[0019] Each component and the corresponding center frequency ω k and multiplication operators The updates are performed using formulas (3), (4), and (5) respectively.

[0020]

[0021]

[0022]

[0023] In the formula: and The Fourier transforms of f(t), uk(t), and λ(t) are respectively; ω is the frequency; n is the number of iterations.

[0024] The precision convergence criterion is set to ε > 0; if the following conditions are met:

[0025] And n < N (6)

[0026] Then the iteration is complete, and the output is given. and ω k Otherwise, repeat the iteration.

[0027] For each power component that has been decomposed Combining wind speed V, yaw error ER, temperature T, humidity H, and pressure p 空 Features form a new dataset X k All datasets X k The set formed is called {X} k};

[0028] {x k Each dataset X in} k The inputs are fed into the corresponding AM-LSTM network structures in the generator G. Each network structure contains 6 input features, and the output h is the LSTM hidden unit at the current time step. t-1 The system calculates the weight of each feature at the current moment using the memory information, quantifies the weights, and adjusts the feature weights in real time through an attention mechanism to achieve orderly fusion of multiple features.

[0029] The weight calculation formula is as shown in formula (7):

[0030]

[0031] Among them, V e W e and U e b is the weight matrix of the multilayer perceptron for the attention mechanism; e For bias terms;

[0032] Using the normalized exponential function according to formula (8) Normalization is performed to ensure that the sum of the feature attention weights is 1;

[0033]

[0034] Then the hidden layer state h of LSTM t The update method is as follows:

[0035]

[0036] In the formula, f1 is an LSTM network unit.

[0037] In the step of constructing the ultra-short-term wind power prediction model of the wind turbine to be predicted, K AM-LSTM network models are superimposed and combined into a generator G of the generative adversarial network. The outputs are superimposed in a weighted manner to obtain the power prediction output value of the generator.

[0038] The model uses the RMSProp optimization algorithm. The generator consists of VMD decomposition and four AM-LSTM neural networks. The LSTM input layer has six neurons, the hidden layer contains ten neurons, and the output layer contains one neuron. The activation function is tanh, and the learning rate is 0.0001.

[0039] The discriminator is composed of a CNN, which includes one convolutional layer, one pooling layer and one fully connected layer. The number of input neurons is 16, the number of convolutional kernels is 32, the activation function is LRELU, the learning rate is 0.0001, and the pooling method is max pooling.

[0040] To address the issues of vanishing gradients, non-convergence during training, and model crashes in the original GAN, a W-distance is introduced. The discriminator estimates the W-distance between the generated distribution and the true distribution, and the generator narrows the W-distance to achieve the goal of generating predicted power that approximates the true power. The gradient of the discriminator relative to the input X is restricted to a certain range so that the discriminator satisfies the Lipschitz condition, i.e., a gradient penalty term is added to the discriminator loss function.

[0041] The loss function of the generator is:

[0042]

[0043] In the formula: J(G) represents the objective function of the WGAN discriminator, G represents the generator network, z is the ultra-short-term power prediction sample generated by the generator, w is the network weight, and f w p represents a neural network z Let G(z) represent the sample distribution for ultra-short-term power prediction, and let G(z) represent the ultra-short-term power dataset generated by the generator. The mathematical expectation of the ultra-short-term power samples generated by the generator.

[0044] The loss function of the discriminator is:

[0045]

[0046] In the formula: J(D) represents the objective function of the WGAN discriminator, r represents the real sample data, x is the input data column, and p r The sample distribution of the original power data, Let represent the gradient, where for The sample distribution This represents the mathematical expectation of the actual power data sampling. Representing data Data expectations.

[0047] The training steps for the ultra-short-term wind power prediction model of the wind turbine to be predicted include:

[0048] Initialize the learning rate of the generator and discriminator and the generator parameter θ d Discriminator parameters θ ω The original dataset is divided into sliding window groups, and the generator takes m data points X as input for each group: (x (1) x (2) , ..., x (m) The discriminator inputs n data points Y for each group: (y (1) y (2) , ..., y (n) The wind power features from dataset X are extracted to form a separate dataset X. p :(x (1)p x (2)p , ..., x (m)p );

[0049] For dataset X p Perform variational mode decomposition, and combine each component with other features to form a new dataset X: (x (1) x (2) , ..., x (m) The inputs are fed into the generator of the attention mechanism, and the prediction results of each component are summed to obtain the generated dataset G(z): (G(z)) (1) , G(z) (2) ,…,G(z) (n) The discriminator D is fed with G(z) and the real dataset Y, and its parameters θ are updated using the stochastic gradient descent algorithm. d ;

[0050] Update the parameters θ of the discriminator G using the stochastic gradient descent algorithm. ω ;

[0051] After training the discriminator 5 times, the generator is trained 1 time. When the discriminator's parameters θ ω If convergence occurs, the training ends.

[0052] Unlike existing technologies, this invention's multi-feature fusion generative adversarial network (GAN) method for ultra-short-term wind power prediction constructs a variational mode decomposition-based GAN model. This model uses a combination of variational mode decomposition, attention mechanisms, and long short-term memory neural networks as the generator, and a convolutional neural network as the discriminator. The W-distance is used as the objective function for the generator, and the Lipshitz continuity condition is used to constrain the discriminator to improve model stability. Furthermore, it employs a technique of using single-machine prediction to infer the overall field power, compensating for the accuracy of wind power prediction when anemometer data cannot reflect the overall wind resource situation. This invention improves the accuracy of ultra-short-term wind power prediction and solves problems such as large fluctuations in wind power, varying contribution rates of multiple factors affecting wind power to ultra-short-term wind power prediction, and the non-convergence and instability of the original GAN ​​model. Attached Figure Description

[0053] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings:

[0054] Figure 1 This is a flowchart illustrating a multi-feature fusion generative adversarial network method for ultra-short-term wind power prediction provided by the present invention.

[0055] Figure 2 This is a schematic diagram of the network model structure of the ultra-short-term wind power prediction model of the wind turbine to be predicted in the multi-feature fusion generative adversarial network ultra-short-term wind power prediction method provided by the present invention.

[0056] Figure 3 This is a schematic diagram of the AM-LSTM coding network structure in the multi-feature fusion generative adversarial network ultra-short-term wind power prediction method provided by the present invention.

[0057] Figure 4 This is a schematic diagram of the CNN network structure in the generative adversarial network ultra-short-term wind power prediction method with multi-feature fusion provided by the present invention. Detailed Implementation

[0058] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0059] See Figure 1 This invention provides a multi-feature fusion generative adversarial network method for ultra-short-term wind power prediction, comprising:

[0060] Obtain historical power data and corresponding weather data of the wind turbine to be predicted, perform data preprocessing, and use the preprocessed data as training data.

[0061] A short-term wind power prediction model for the wind turbine to be predicted is constructed. This model is then trained using training data. The model utilizes a combination of variational mode decomposition and attention mechanism with a long short-term memory neural network as the generator, a convolutional neural network as the discriminator, and W distance as the objective function. The network structure of the short-term wind power prediction model is as follows: Figure 2 As shown.

[0062] The real-time power data and weather data of the wind turbine to be predicted are input into the trained ultra-short-term wind power prediction model of the wind turbine to be predicted, and the model output is used as the real-time ultra-short-term wind power prediction result.

[0063] The data preprocessing steps include:

[0064] Outlier removal and imputation are performed on the historical power data of the wind turbines to be predicted and the corresponding weather data to form a standardized original dataset, which is used as training data.

[0065] In the step of performing variational mode decomposition on the power data column in the original dataset, a variational model is constructed through iterative search, decomposing the original wind power time series P(t) in the training data into components u of different frequencies. k (t), corresponding to the center frequency ω k The optimal solution of the variational model is found by updating iteratively in an alternating manner; where,

[0066] The variational mode decomposition constraint expression is:

[0067]

[0068] In the formula, K is the number of modes in the decomposition, {u k}、{ω k} represents the k-th modal component and center frequency, respectively, δ(t) is the Dirac function, and * is the convolution operator.

[0069] In this process, the Lagrange multiplication operator λ and the quadratic penalty factor α are introduced to transform the constrained variational problem of equation (1) into an unconstrained variational problem for solution, resulting in the constrained expression of the wind power output signal VMD as follows:

[0070]

[0071] Each component and the corresponding center frequency ω k and multiplication operators The updates are performed using formulas (3), (4), and (5) respectively.

[0072]

[0073]

[0074]

[0075] In the formula: and f(t) and u are respectively k The Fourier transforms of λ(t) and λ(t) are given by ω, where ω is the frequency and n is the number of iterations.

[0076] The precision convergence criterion is set to ε > 0; if the following conditions are met:

[0077] And n < N (6)

[0078] Then the iteration is complete, and the output is given. and ω k Otherwise, repeat the iteration.

[0079] For each power component that has been decomposed Combining wind speed V, yaw error ER, temperature T, humidity H, and pressure p 空 Features form a new dataset X k All datasets X k The set formed is called {X} k};

[0080] {x k Each dataset X in} k The inputs are fed into the corresponding AM-LSTM network structures in the generator G. Each network structure contains 6 input features, and the output h is the LSTM hidden unit at the current time step. t-1 The system calculates the weight of each feature at the current moment using the memory information, quantifies the weights, and adjusts the feature weights in real time through an attention mechanism to achieve orderly fusion of multiple features.

[0081] The weight calculation formula is as shown in formula (7):

[0082]

[0083] Among them, V e W e and U e b is the weight matrix of the multilayer perceptron for the attention mechanism; e For bias terms;

[0084] Using the normalized exponential function according to formula (8) Normalization is performed to ensure that the sum of the feature attention weights is 1;

[0085]

[0086] Then the hidden layer state h of LSTM t The update method is as follows:

[0087]

[0088] In the formula, f l This is an LSTM network unit. The AM-LSTM coding network structure is as follows: Figure 3 As shown.

[0089] In the step of constructing the ultra-short-term wind power prediction model of the wind turbine to be predicted, K AM-LSTM network models are superimposed and combined into a generator G of the generative adversarial network. The outputs are superimposed in a weighted manner to obtain the power prediction output value of the generator.

[0090] The optimization algorithm used in the model is RMSProp. The generator consists of VMD decomposition and K AM-LSTM neural networks. The LSTM input layer has 6 neurons, the hidden layer contains 10 neurons, and the output layer contains 1 neuron. The activation function is tanh, and the learning rate is 0.0001.

[0091] The discriminator is constructed using a CNN, consisting of one convolutional layer, one pooling layer, and one fully connected layer. It has 16 input neurons, 32 convolutional kernels, LRELU activation, a learning rate of 0.0001, and max pooling. The CNN network structure is as follows: Figure 4 As shown.

[0092] To address the issues of vanishing gradients, non-convergence during training, and model crashes in the original GAN, a W-distance is introduced. The discriminator estimates the W-distance between the generated distribution and the true distribution, and the generator narrows the W-distance to achieve the goal of generating predicted power that approximates the true power. The gradient of the discriminator relative to the input X is restricted to a certain range so that the discriminator satisfies the Lipschitz condition, i.e., a gradient penalty term is added to the discriminator loss function.

[0093] The loss function of the generator is:

[0094]

[0095] In the formula: J(G) represents the objective function of the WGAN discriminator, G represents the generator network, z is the ultra-short-term power prediction sample generated by the generator, w is the network weight, and f w p represents a neural network z Let G(z) represent the sample distribution for ultra-short-term power prediction, and let G(z) represent the ultra-short-term power dataset generated by the generator. The mathematical expectation of the ultra-short-term power samples generated by the generator.

[0096] The loss function of the discriminator is:

[0097]

[0098] In the formula: J(D) represents the objective function of the WGAN discriminator, r represents the real sample data, x is the input data column, and p r The sample distribution of the original power data, Let represent the gradient, where for The sample distribution This represents the mathematical expectation of the actual power data sampling. Representing data Data expectations

[0099] The training steps for the ultra-short-term wind power prediction model of the wind turbine to be predicted include:

[0100] Initialize the learning rate of the generator and discriminator and the generator parameter θ d Discriminator parameters θ ω The original dataset is divided into sliding window groups, and the generator takes m data points X as input for each group: (x (1) x (2) , ..., x (m) The discriminator inputs n data points Y for each group: (y (1) y (2) , ..., y (n) The wind power features from dataset X are extracted to form a separate dataset X. p :(x (1)p x (2)p , ..., x (m)p );

[0101] For dataset X p Perform variational mode decomposition, and combine each component with other features to form a new dataset X: (x (1) x (2) , ..., x (m) The inputs are fed into the generator of the attention mechanism, and the prediction results of each component are summed to obtain the generated dataset G(z): (G(z)) (1) , G(z) (2) ,…,G(z) (n) The discriminator D is fed with G(z) and the real dataset Y, and its parameters θ are updated using the stochastic gradient descent algorithm. d ;

[0102] Update the parameters θ of the discriminator G using the stochastic gradient descent algorithm. ω ;

[0103] After training the discriminator 5 times, the generator is trained 1 time. When the discriminator's parameters θ ω If convergence occurs, the training ends.

[0104] Finally, the actual data sequence X of each wind turbine is input into the network model to obtain the predicted power sequence p of each turbine. 单机 The power sequences of each individual unit are summed to obtain the overall predicted power sequence p. 预测 .

[0105] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention 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 the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A multi-feature fusion generative adversarial network method for ultra-short-term wind power prediction, characterized in that, include: Obtain historical power data and corresponding weather data of the wind turbine to be predicted, perform data preprocessing, and use the preprocessed data as training data. A short-term wind power prediction model for the wind turbine to be predicted is constructed; the constructed short-term wind power prediction model for the wind turbine to be predicted is trained using the training data; wherein, the short-term wind power prediction model for the wind turbine to be predicted uses variational mode decomposition combined with attention mechanism and long short-term memory neural network as generator, convolutional neural network as discriminator, and W distance as objective function of generator. The real-time power data and weather data of the wind turbine to be predicted are input into the trained ultra-short-term wind power prediction model of the wind turbine to be predicted, and the model output is used as the real-time ultra-short-term wind power prediction result. To address the issues of vanishing gradients, non-convergence during training, and easy model collapse in the original GAN, a W-distance is introduced. The discriminator estimates the W-distance between the generated distribution and the true distribution, and the generator narrows the W-distance to achieve the goal of generating predicted power that approximates the true power. The gradient of the discriminator relative to the input X is restricted within a certain range to make the discriminator satisfy the Lipschitz condition, which is a gradient penalty term. The loss function of the generator is: （10） The loss function of the discriminator is: （11） In the formula: J(D) represents the objective function of the WGAN discriminator, r represents the real sample data, and x is the input data column. The sample distribution of the original power data, Let represent the gradient, where , for The sample distribution This represents the mathematical expectation of the actual power data sampling. Representing data Data expectations.

2. The multi-feature fusion generative adversarial network method for ultra-short-term wind power prediction according to claim 1, characterized in that, The data preprocessing steps include: Outlier removal and interpolation are performed on the historical power data and corresponding weather data of the wind turbine to be predicted to form a standardized original dataset, which is used as the training data.

3. The multi-feature fusion generative adversarial network method for ultra-short-term wind power prediction according to claim 1, characterized in that, In the step of performing variational mode decomposition on the power data series in the original dataset, a variational model is constructed through iterative search to decompose the original wind power time series P(t) in the training data into components of different frequencies. The corresponding center frequency is The optimal solution of the variational model is found by updating iteratively in an alternating manner; where, The variational mode decomposition constraint expression is: （1） In the formula, K is the number of modes in the decomposition. These are the k-th modal component and its center frequency, respectively. For the Dirac function, This is the convolution operator.

4. The multi-feature fusion generative adversarial network method for ultra-short-term wind power prediction according to claim 3, characterized in that, By introducing the Lagrange multiplication operator λ and the quadratic penalty factor α, the constrained variational problem of equation (1) is transformed into an unconstrained variational problem for solution, and the constrained expression of the wind power output signal VMD is obtained as follows: （2） Each component and corresponding center frequency and multiplication operators The updates are performed using formulas (3), (4), and (5) respectively. In the formula: and They are respectively and Fourier transform; is the frequency; n is the number of iterations; The accuracy convergence criterion is set as follows If the following conditions are met: （6） Then the iteration is complete, and the output is given. and Otherwise, repeat the iteration.

5. The multi-feature fusion generative adversarial network method for ultra-short-term wind power prediction according to claim 4, characterized in that, For each power component after decomposition Combining wind speed V, yaw error ER, temperature T, humidity H, and pressure Features form a new dataset X k All datasets X k The set formed is called {X} k }; {x k Each dataset X in} k The inputs are fed into the corresponding AM-LSTM network structures in the generator G. Each network structure contains 6 input features, and the output is the LSTM hidden layer unit at the current time step. The system calculates the weight of each feature at the current moment using the memory information, quantifies the weights, and adjusts the feature weights in real time through an attention mechanism to achieve orderly fusion of multiple features. The weight calculation formula is as shown in formula (7): （7） in, and This represents the weight matrix of the multilayer perceptron for the attention mechanism. For bias terms; Using the normalized exponential function according to formula (8) Normalization is performed to ensure that the sum of the feature attention weights is 1; （8） Then the hidden layer state h of LSTM t The update method is as follows: （9） In the formula, f l It is an LSTM network unit.

6. The multi-feature fusion generative adversarial network method for ultra-short-term wind power prediction according to claim 5, characterized in that, In the step of constructing the ultra-short-term wind power prediction model of the wind turbine to be predicted, K AM-LSTM network models are superimposed and combined into a generator G of the generative adversarial network. The outputs are superimposed in a weighted manner to obtain the power prediction output value of the generator. The optimization algorithm used in the model is RMSProp. The generator consists of VMD decomposition and K AM-LSTM neural networks. The LSTM input layer has 6 neurons, the hidden layer contains 10 neurons, and the output layer contains 1 neuron. The activation function is tanh, and the learning rate is 0.0001. The discriminator is composed of a CNN, which includes one convolutional layer, one pooling layer and one fully connected layer. The number of input neurons is 16, the number of convolutional kernels is 32, the activation function is LRELU, the learning rate is 0.0001, and the pooling method is max pooling.

7. The multi-feature fusion generative adversarial network method for ultra-short-term wind power prediction according to claim 1, characterized in that, The training steps for the ultra-short-term wind power prediction model of the wind turbine to be predicted include: Initialize the learning rate of the generator and discriminator and the generator parameters. Discriminator parameters The original dataset is divided into sliding window groups, and the generator is input with m data points in each group. The discriminator inputs n data points per group. The wind power features in the X dataset are extracted to form a separate dataset. ; For dataset X p Perform variational mode decomposition, and combine each component with other features to form a new dataset. The data is input into the generator of the attention mechanism, and the prediction results of each component are summed to obtain the generated dataset. ,Will The discriminator D is fed along with the real dataset Y, and its parameters are updated using the stochastic gradient descent algorithm. ; Update the parameters of the discriminator G using the stochastic gradient descent algorithm. ; After training the discriminator 5 times, train the generator 1 time. When the discriminator parameters... If convergence occurs, the training ends.

Images