Unified spatiotemporal wind power forecasting method based on fourier generative adversarial network

Abstract

The application provides a unified space-time wind power prediction method based on a Fourier generative adversarial network, comprising: obtaining a hyper-variable graph based on unified space-time, the hyper-variable graph being constructed based on historical power sequence data sets of multiple wind generators; constructing a multi-wind farm power prediction model; the multi-wind farm power prediction model is a deep regret analysis generative adversarial network (DRAGAN), which is an adversarial network with a Fourier GNN as a generator and a convolutional neural network (CNN) as a discriminator; the hyper-variable graph is used to iteratively train the multi-wind farm power prediction model to obtain a multi-wind farm power prediction model meeting the prediction requirements. The method provided by the application optimizes the parameters of the generator through the mutual game training of the generator Fourier GNN and the discriminator CNN, improves the quality of the generated samples, and further improves the prediction accuracy.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of power prediction, in particular to a unified space-time wind power prediction method based on a Fourier generative adversarial network. BACKGROUND

[0002] Wind power, as a green and low-carbon renewable energy, has developed rapidly and is a key to environmental protection and sustainable development. However, the randomness and nonlinearity of wind power change bring challenges to the dispatching and balancing of power grids, greatly affecting the safe and stable operation of power grids and hindering the effective utilization and management of wind energy. Therefore, accurate wind power prediction is of great significance for improving the efficiency of wind power generation, optimizing the operation of power systems, and promoting the development of renewable energy.

[0003] Wind power prediction plays a crucial role in the power system, aiming to accurately predict the future power output of wind power generation to help grid operators effectively plan and manage system operation and ensure the balance and reliability of the power system. According to the prediction method, it can be divided into physical method, statistical method and artificial intelligence method. The physical method relies heavily on high-precision weather forecast data, resulting in limited prediction accuracy. The statistical method has limited non-linear modeling capability, and the prediction effect of wind power with strong volatility is poor. In order to overcome the limitations of physical and statistical models, artificial intelligence is an effective method to achieve accurate wind power prediction.

[0004] Wind data, as a special type of space-time data, may exhibit rich correlations between different time points and stations due to the influence of atmospheric flow. Therefore, when predicting the power of multiple wind farms in multiple regions, the space-time correlation between different stations needs to be considered. Traditional space-time modeling methods often use graph neural network-based methods to extract spatial correlation and then use time series models to extract temporal correlation. These graph neural network-based methods rely on pre-defined graph structures to specify spatial correlation and cannot dynamically capture time-varying space-time correlation. On the other hand, space-time separation modeling violates the unified space-time interdependence relationship in the real world.

[0005] Existing space-time prediction models cannot reflect the unified space-time autocorrelation and cross-correlation of wind power sequences in the real world, lack understanding of complex interwoven space-time features, and have limited prediction accuracy. SUMMARY

[0006] The application provides a unified space-time wind power prediction method based on a Fourier generative adversarial network. To solve the above technical problems, the application adopts the following technical methods:

[0007] In a first aspect, the application provides a unified space-time wind power prediction method based on a Fourier generative adversarial network, comprising:

[0008] Obtain a hypervariable graph based on a unified spatiotemporal framework, which is constructed based on historical power sequence datasets of multiple wind turbines;

[0009] A multi-wind farm power prediction model is constructed; the multi-wind farm power prediction model is a deep regret analysis generative adversarial network DRAGAN, which is an adversarial network with Fourier GNN as the generator and convolutional neural network CNN as the discriminator.

[0010] The multi-wind farm power prediction model is iteratively trained using the hypervariable graph to obtain a multi-wind farm power prediction model that meets the prediction requirements.

[0011] Optionally, the hypervariable graph is constructed based on historical power sequence datasets of multiple wind turbines; including:

[0012] Obtain historical power sequence sample data from multiple wind turbines;

[0013] After cleaning and normalizing the historical power sequence sample data, a historical power sequence dataset is obtained.

[0014] Based on the historical power sequence dataset, a hypervariable graph based on unified spatiotemporal conditions is constructed.

[0015] Optionally, the loss function of the generator is constructed as follows:

[0016] Construct the first loss function for the generator;

[0017] Construct feature matching expressions;

[0018] The first loss function and the feature matching expression are added together to determine the generator's loss function.

[0019] Optionally, the loss function of the discriminator is constructed as follows:

[0020] The loss function of the discriminator is determined by combining the DRAGAN penalty term and the Wasserstein distance.

[0021] Optionally, the step of using the hypervariable graph to iteratively train the multi-wind farm power prediction model to obtain a multi-wind farm power prediction model that meets the prediction requirements includes:

[0022] Based on the hypervariable graph, determine the label data corresponding to the hypervariable graph;

[0023] The hypervariable graph is input into the generator, and the power prediction results of multiple wind farms are output.

[0024] Based on the power prediction results of the multiple wind farms and the labeled data, the internal parameters of the generator and discriminator are optimized through adversarial training to obtain a power prediction model of the multiple wind farms that meets the prediction requirements.

[0025] Optionally, the step of optimizing the internal parameters of the generator and discriminator based on the multi-wind farm power prediction results and the label data to obtain a multi-wind farm power prediction model that meets the prediction requirements includes:

[0026] Substitute the multi-wind farm power prediction results and the tag data into the loss function of the generator and the loss function of the discriminator to calculate the generator loss value and the discriminator loss value.

[0027] The generator loss value and the discriminator loss value are input into the model optimizer to optimize the internal parameters of the generator and discriminator, thereby obtaining a multi-wind farm power prediction model that meets the prediction requirements.

[0028] Optionally, the optimization algorithm used by the model optimizer is an adaptive matrix estimation algorithm.

[0029] Secondly, this application also provides a computer system, comprising:

[0030] Memory is used to store instructions that can be executed by the processor;

[0031] A processor for executing the instructions to implement the method as described in any of the first aspects.

[0032] Thirdly, this application also provides a computer-readable medium storing computer program code that, when executed by a processor, implements the method as described in any of the first aspects.

[0033] This application has the following beneficial effects:

[0034] The method proposed in this application optimizes the parameters of the generator by training the generator FourierGNN and the discriminator CNN through mutual game, thereby improving the quality of generated samples and further enhancing prediction accuracy. Attached Figure Description

[0035] Figure 1 A flowchart illustrating the unified spatiotemporal wind power prediction method based on Fourier generative adversarial networks provided in this application embodiment;

[0036] Figure 2 A hypervariable plot with three time series provided for embodiments of this application;

[0037] Figure 3 This is a schematic diagram of the FourierGNN network structure provided in the embodiments of this application. Detailed Implementation

[0038] To facilitate understanding by those skilled in the art, the present application will be further described below in conjunction with embodiments and accompanying drawings. The content mentioned in the embodiments is not intended to limit the present application.

[0039] To solve the above technical problems, such as Figure 1 As shown, this application proposes a unified spatiotemporal wind power prediction method based on Fourier generative adversarial networks, including:

[0040] Step S101: Obtain a hypervariable graph based on a unified spatiotemporal model, wherein the hypervariable graph is constructed based on historical power sequence datasets of multiple wind turbines;

[0041] In most multivariate scenarios, the graph structure of the sequences is usually unknown. Furthermore, due to time lag effects, elements within and between sequences exhibit spatial and temporal interdependencies. Hypervariable graphs can learn these spatiotemporal dependencies from a unified perspective; therefore, in this step, a hypervariable graph is constructed based on historical power sequence sample data from multiple wind turbines.

[0042] First, historical power sequence sample data from multiple wind turbines are obtained. Since the original data may contain missing data, data cleaning is necessary. Furthermore, given the varying time spans of historical data from different sites, a max-min normalization operation is performed on the experimental dataset to eliminate discrepancies, accelerate model convergence, and reduce training time. The set of historical power sequences from N sites with a length of M is denoted as... After normalization, new time series data are obtained. The formula is as follows:

[0043] (1)

[0044] In equation (1), min represents the minimum value in the sequence, and max represents the maximum value in the sequence. This represents the time series data set of N stations after normalization.

[0045] The data after the above normalization process For example, suppose the input data for the prediction model consists of data from T consecutive time points. The model outputs the predicted power value at time point (T+1). ,but The corresponding tag data can be recorded as We obtain (MT) groups of samples from the original data, denoted as... The last S sets of samples in the (MT) group are used as the test set for historical power sequence data, and all previous sample data are used as the training set for historical power sequence data.

[0046] Then, based on historical power sequence datasets To construct a hypervariable graph based on a unified spacetime. ,in , Given a multivariate time series window Where N and T are the number of variables and the time length, respectively. Each element in the graph is treated as a node, thus constructing a hypervariable graph with NT nodes. ,in Representing node characteristics, Denotes the initial adjacency matrix, such that It becomes a fully connected graph. Figure 2 The hypervariable plots for three time series are shown. Using this data structure, the multivariate time series forecasting task can be reformulated as a prediction of the hypervariable plots, as follows:

[0047] (2)

[0048] in, Represents the network parameters of the hypervariable graph. This is achieved through the construction of a fully connected hypervariable graph structure. We can learn about the interdependence of space and time from a unified perspective.

[0049] Step S102: Construct a multi-wind farm power prediction model; the multi-wind farm power prediction model is a deep regret analysis generative adversarial network DRAGAN, which is an adversarial network with Fourier GNN as the generator and convolutional neural network CNN as the discriminator;

[0050] This section constructs a multi-wind farm power prediction model, primarily using a Deep Regret Analytic GAN (DRAGAN). This adversarial network mainly employs a Fourier Graph Neural Network (DRAGAN) as the generator and a Convolutional Neural Network (CNN) as the discriminator. The network structure of the Fourier GAN is as follows: Figure 3 As shown. The specific process is as follows:

[0051] The constructed hypervariable graph By embedding matrix Assign to each node dimensional vector, Projection to node embedding middle.

[0052] To capture spatiotemporal features simultaneously, the embedding is first performed... Perform a discrete Fourier transform on each discrete spatiotemporal dimension and obtain the frequency output. Then, it is executed in Fourier space and FGO. and Perform recursive multiplication between the two operations and output the result:

[0053] (3)

[0054] Then, to Perform an inverse Fourier transform to transform it to the time domain. Finally, project it onto the time domain using a feedforward network. The prediction task is completed in a future step, thereby outputting the power prediction results of multiple wind farms.

[0055] The aforementioned adversarial networks are prone to mode collapse during training. Therefore, this application defines the cost function of the discriminator by combining the DRAGAN penalty term and the Wasserstein distance. These two terms are then combined with the weight parameters. The sum is as follows:

[0056] (4)

[0057] in, The weights of the gradient penalty term, This represents a noisy actual sample, where From Gaussian distribution The selected random number. Hyperparameters The allowed range of gradient norms is specified. This represents the optimized parameters of the discriminator D. express Generate data samples under the distribution Indicates the distribution of generated data. For actual data, This indicates the actual data distribution.

[0058] The discriminator's loss function mainly consists of the mean difference between the real data output and the generated data output, plus a penalty term, which is used to improve the discriminator's training stability or performance to help train the generator.

[0059] The first loss function for constructing the generator is as follows:

[0060] (5)

[0061] To enhance GAN optimization, a feature matching expression is constructed as shown below. This expression is then added to the loss function of the first generator to improve training stability. This operation is called feature matching.

[0062] (6)

[0063] in, This indicates that the discriminator-specific intermediate layer is related to the input. The output of .

[0064] The generator's first loss function is added to the feature matching expression to determine the generator's loss function as follows:

[0065] (7)

[0066] In the formula, The weights of expression (6) are expressed.

[0067] The generator's loss function calculates the average of the squared differences between the real and generated data and the intermediate layer features of the discriminator, thus encouraging the generator to produce features that more closely resemble the real data. By minimizing this loss function, the generator is encouraged to produce outputs that are as deceptive as possible to the discriminator, and the generated data should be similar to the real data in terms of intermediate layer features, thereby improving the generator's performance.

[0068] Step S103: Using the hypervariable graph, iteratively train the multi-wind farm power prediction model to obtain a multi-wind farm power prediction model that meets the prediction requirements.

[0069] As explained above, the hypervariable plot is constructed based on a training set of historical power sequence data, while the label data is obtained from normalized data. Therefore, there is a correspondence between the hypervariable plot and the label data. The hypervariable plot is then input into the generator for training, and the output is the power prediction result for multiple wind farms. Then, based on the power prediction results of multiple wind farms and the label data, the internal parameters of the generator and discriminator are optimized through adversarial training to obtain a power prediction model for multiple wind farms that meets the prediction requirements.

[0070] Based on the power prediction results and label data from multiple wind farms, the internal parameters of the generator and discriminator are optimized through adversarial training to obtain a power prediction model for multiple wind farms that meets the prediction requirements.

[0071] Substitute the multi-wind farm power prediction results and the labeled data into the generator loss function (7) and the discriminator loss function (4) to calculate the generator loss value and the discriminator loss value. Input the generator loss value and the discriminator loss value into the model optimizer to optimize the internal parameters of the generator and the discriminator to obtain a multi-wind farm power prediction model that meets the prediction requirements.

[0072] The optimization algorithm used by the above model optimizer is the Adaptive Moment Estimation (Adam) algorithm as the optimization algorithm for wind power prediction model. Through iterative training, the network parameters are updated, so that the loss function values ​​of the generator and discriminator converge quickly.

[0073] After iterative training, the prediction model, having learned from the training set samples, has acquired the ability to extract features from wind power data at multiple sites. Therefore, S sets of test set samples are sequentially fed into the generator to obtain prediction results. These prediction results are then analyzed and evaluated to understand the model's predictive performance.

[0074] To verify the predictive performance of the proposed model, regression evaluation metrics were used to assess the accuracy of the prediction results. Four error metrics were used to evaluate the model from multiple perspectives: mean squared error (MSE), mean absolute error (MAE), root mean squared error (RMSE), and coefficient of determination (R²).

[0075] Mean squared error (MSE) is used to evaluate the difference between a model's predictions and the actual values. A smaller MSE indicates more accurate predictions from the model.

[0076] (8)

[0077] Mean Absolute Error (MAE) is the average of absolute errors and can better reflect the actual situation of the error.

[0078] (9)

[0079] The root mean square error (RMSE) is the sum of squares of the deviations between predicted and actual values, used to measure the discrepancy between the predicted and actual values.

[0080] (10)

[0081] Coefficient of determination This reflects the degree of fit between the actual data and the model's predicted data. Its value ranges from 0 to 1, with a closer value to 1 indicating a better fit between the model and the data.

[0082] (11)

[0083] In the formula, n is the number of samples; These are measured values; This is a predicted value; This is the summation and average of the actual values.

[0084] This invention selects data from the NREL dataset, using the historical wind power of six nearby wind turbines as the research object for the wind power prediction model. Data collection spanned from March 1, 2010 to February 28, 2011, a total of one year of wind power data, with a data resolution of 5 minutes. In this wind power prediction, the original data was resampled every 15 minutes, collecting a total of 96 data points per day. Since wind power generation is affected by seasonal factors, wind speed, wind direction, and other meteorological conditions may vary significantly in different seasons, thus affecting the power generation of wind turbines. Therefore, this paper conducts experiments according to the four seasons: spring, summer, autumn, and winter. For each season, the dataset is divided into training and test sets in a 7:3 ratio. The training set is used to train and adjust the hyperparameters of the prediction model, while the test set is used to evaluate the performance of the prediction model. The overall prediction performance of the proposed method is verified in one-step and multi-step predictions.

[0085] In the experiment, the parameter settings of the prediction model based on Fourier GNN and DRAGAN were as follows: the generator had 3 layers in the Fourier neural network, with hidden layer feature dimensions of 64 and 32; the discriminator had 3 layers in the CNN network, with hidden layer feature dimensions of 220 and 110; a linear layer was used as the output layer, and the output feature dimension was the predicted value of the prediction step size. The model was trained for 200 iterations with a learning rate of 0.0001.

[0086] To verify the effectiveness of the proposed model in wind power prediction tasks, this application compares the prediction performance of GAN, WGAN-GP, CNN-GRU, and LSTM models. LSTM and GAN are based on deep learning methods; LSTM is a classic time series prediction model, while GAN is a classic generative model that learns data distribution through adversarial training. WGAN-GP addresses the mode collapse and pattern reversal problems in traditional GANs, improving training stability. CNN-GRU is a spatiotemporal prediction model that uses CNN to remove spatial features from the data and then uses GRU to capture temporal features. To ensure the validity of the experimental results, all models were trained and tested using the same dataset. To visually demonstrate the superiority of the proposed model, Table 1 shows the prediction error metrics of the proposed model, WGAN-GP, CNN-GRU, GAN, and LSTM for 15 minutes of wind power generation prediction for six wind turbines.

[0087] Table 1. Single-step WPF error indices using different prediction methods

[0088] ;

[0089] As shown in Table 1, the accuracy metrics of this application demonstrate that the proposed model outperforms other models. Specifically, taking spring as an example, compared with WGAN-GP, CNN-GRU, GAN, and LSTM models, the MSE is reduced by 1.0621, 0.1096, 0.0851, and 0.0212, respectively. The MAE is reduced by 0.718, 0.0545, 0.0218, and -0.0045, respectively. The RMSE is reduced by 0.7829, 0.1512, 0.1165, and 0.0348, respectively. 2 The value is the largest among all models. Overall, all four evaluation metrics confirm that the proposed model has good predictive performance.

[0090] To verify the performance of the proposed method in multi-step prediction, the evaluation index of the spring wind power prediction results obtained by comparing the proposed method with four comparison models, namely WGAN-GP, CNN-GRU, GAN and LSTM, is shown in Table 2.

[0091] Table 2. Error Measurements of Different Forecasting Methods in Spring

[0092] ;

[0093] It can be observed that as the prediction step size increases, the superiority of the proposed model becomes more pronounced compared to other models. Taking the 4-hour prediction in spring as an example, compared to WGAN-GP, MSE decreased by 48.70%, MAE decreased by 46.46%, RMSE decreased by 28.37%, and R2 increased by 28.88%; compared to CNN-GRU, MSE decreased by 25.82%, MAE decreased by 17.92%, RMSE decreased by 13.87%, and R2 increased by 2.10%; compared to GAN, MSE decreased by 30.00%, MAE decreased by 26.85%, RMSE decreased by 16.33%, and R2 increased by 28.88%. 2 The performance improved by 8.16%; compared to LSTM, MSE decreased by 16.38%, MAE decreased by 9.20%, RMSE decreased by 0.04%, and R2 increased by 1.93%. This indicates that compared to other models, the proposed model can more accurately capture the changing trends of data, learn more hidden features, and provide more reliable multi-scale prediction results.

[0094] In summary, the proposed method achieves unified spatiotemporal modeling by combining DRAGAN and Fourier GNN. Fourier GNN serves as the generator in the generative adversarial network (GAN), and CNN serves as the discriminator. In this model, historical values ​​of each wind turbine are used as conditional inputs to both the generator and discriminator, providing more precise control over the generated output features. A feature matching mechanism is added to the generator to enhance training stability. Through the interplay between the Fourier GNN generator and the CNN discriminator during training, the generator parameters are optimized, improving the quality of generated samples and further enhancing prediction accuracy.

[0095] In some embodiments, this application also provides a computer system including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0096] This application also provides a computer-readable storage medium for storing a computer program. This computer-readable storage medium can be applied to a computer device, and the computer program causes the computer device to execute the corresponding processes in the methods described above in the embodiments of this application; for brevity, further details are omitted here.

[0097] The above embodiments are preferred implementations of this application. In addition, this application can be implemented in other ways. Any obvious substitutions without departing from the concept of this technical solution are within the protection scope of this application.

[0098] To facilitate understanding by those skilled in the art of the improvements made by this application compared to the prior art, some of the accompanying drawings and descriptions have been simplified, and for clarity, some other elements have been omitted from this application. Those skilled in the art should realize that these omitted elements may also constitute the content of this application.

Claims

1. A unified spatiotemporal wind power prediction method based on Fourier generative adversarial networks, characterized in that, include: Obtain a hypervariable graph based on a unified spatiotemporal framework; the hypervariable graph is constructed based on historical power sequence datasets of multiple wind turbines; A multi-wind farm power prediction model is constructed; the multi-wind farm power prediction model is a deep regret analysis generative adversarial network DRAGAN, which is an adversarial network with a Fourier graph neural network (Fourier GNN) as the generator and a convolutional neural network (CNN) as the discriminator. The multi-wind farm power prediction model is iteratively trained using the hypervariable graph to obtain a multi-wind farm power prediction model that meets the prediction requirements.

2. The method according to claim 1, characterized in that, The hypervariable graph is constructed based on historical power sequence datasets of multiple wind turbines; including: Obtain historical power sequence sample data from multiple wind turbines; After cleaning and normalizing the historical power sequence sample data, a historical power sequence dataset is obtained. Based on the historical power sequence dataset, a hypervariable graph based on unified spatiotemporal conditions is constructed.

3. The method according to claim 1, characterized in that, The loss function of the generator is constructed as follows: Construct the first loss function for the generator; Construct feature matching expressions; The first loss function and the feature matching expression are added together to determine the generator's loss function.

4. The method according to claim 3, characterized in that, The loss function of the discriminator is constructed as follows: The loss function of the discriminator is determined by combining the DRAGAN penalty term and the Wasserstein distance.

5. The method according to claim 4, characterized in that, The process of using the hypervariable graph to iteratively train the multi-wind farm power prediction model to obtain a multi-wind farm power prediction model that meets the prediction requirements includes: Based on the hypervariable graph, determine the label data corresponding to the hypervariable graph; The hypervariable graph is input into the generator, and the power prediction results of multiple wind farms are output. Based on the power prediction results of the multiple wind farms and the labeled data, the internal parameters of the generator and discriminator are optimized through adversarial training to obtain a power prediction model of the multiple wind farms that meets the prediction requirements.

6. The method according to claim 5, characterized in that, The process involves optimizing the internal parameters of the generator and discriminator based on the multi-wind farm power prediction results and the tag data to obtain a multi-wind farm power prediction model that meets the prediction requirements, including: Substitute the multi-wind farm power prediction results and the tag data into the loss function of the generator and the loss function of the discriminator to calculate the generator loss value and the discriminator loss value. The generator loss value and the discriminator loss value are input into the model optimizer to optimize the internal parameters of the generator and discriminator, thereby obtaining a multi-wind farm power prediction model that meets the prediction requirements.

7. The method according to claim 6, characterized in that, The optimization algorithm used by the model optimizer is an adaptive matrix estimation algorithm.

8. A computer system, characterized in that, include: Memory is used to store instructions that can be executed by the processor; A processor for executing the instructions to implement the method as described in any one of claims 1 to 7.

9. A computer-readable medium, characterized in that, The system contains computer program code that, when executed by a processor, implements the method as described in any one of claims 1 to 7.

Images