Label new energy scene generation method based on improved style generative adversarial network

Abstract

The invention provides a label new energy scene generation method based on an improved style generative adversarial network, and the method comprises the seven steps: 1-3, generating a decoupled style control parameter through introducing a mapping network, and injecting the decoupled style control parameter into a synthesis network layer by layer, thereby effectively solving the problem of feature entanglement, remarkably improving the accuracy and controllability of scene generation, and improving the scene generation efficiency. And step 4 to step 7, through cyclic training of the generator and the discriminator, diversity and authenticity of the generated scene are further ensured.

Description

Technical Field

[0001] This invention relates to the field of new energy scene generation technology, and in particular to a method for generating labeled new energy scenes based on an improved style generative adversarial network. Background Technology

[0002] With the development of new power systems, high-quality and targeted renewable energy scenario data are playing an increasingly important role in current power system research. Existing controllable generation methods for renewable energy scenarios mainly adopt traditional conditional generative adversarial network methods, such as the C-WGAN-GP model. This model directly concatenates data labels with historical power output data and feeds them into the generator network. Renewable energy output scenarios are generated through adversarial training between the generator and the discriminator. This is generally used as a common technique for controllable generation of renewable energy scenarios. Although it is simple in structure and easy to implement, it suffers from significant feature entanglement, leading to unstable training, pattern collapse, and limited generalization ability. Furthermore, the generated scenarios have a large root mean square error (RMSE) and a large interval width at the same confidence level, limiting its ability to characterize the seasonal characteristics of renewable energy and the spatial correlation of power plants. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art and provide a label new energy scene generation method based on an improved style generative adversarial network, which can solve the problems of unstable training and poor scene accuracy caused by feature entanglement in traditional methods.

[0004] Therefore, the present invention adopts the following technical solution: A method for generating labeled new energy scenarios based on improved style generative adversarial networks includes the following steps: Step 1: Obtain the condition vector c and the real data x of the new energy power output scenario. The real data x of the new energy power output scenario includes historical power output data p and predicted power output data f. The condition vector c includes meteorological data, time information and new energy power station location data. Step 2: Construct the CStyleGAN model, which includes a neural network generator G and a discriminator D. The neural network generator G includes a mapping network G. map With the synthetic network G syn ; Step 3: Input the Gaussian noise z and the conditional vector c into the mapping network G. map Through the mapping network G map Generate decoupled style control parameters yi; Step 4: Inject the style control parameter yi layer by layer into the synthesis network G. syn Generate sample data x' for new energy power output scenarios; Step 5: Input the real data x of the new energy power output scenario, the sample data x' of the new energy power output scenario, and the condition vector c into the discriminator D, and output the discriminant values ​​D(x) and D(x') through the discriminator D. Step 6: Calculate the loss function based on the discriminant value and update the parameters of the discriminator D and the parameters of the neural network generator G; Step 7: Repeat steps 3 to 6. As training continues, the distribution difference between the generated samples and the real data will gradually decrease. When Nash equilibrium is reached, the repetition will terminate. Extract the generator G after training is completed, which can then be used to generate the specified new energy power output scenario.

[0005] Based on the above technical solutions, the present invention may also employ the following further technical solutions, or combine these further technical solutions: To control various indicators in the generated wind and light scene, the historical power output data, including month (m), weather conditions (w), temperature (t), humidity (h), wind speed (v), solar radiation intensity (r), and station number (num), are used to form a condition vector c, which is specifically composed of: (1) Where m is a 12-bit unique hot code representing the month, with each bit representing a month; w is a 4-bit unique hot code representing the weather, with each bit representing one of the four weather conditions: sunny, partly cloudy, cloudy, or rain / snow; t max , t min , t ave These represent the maximum, minimum, and average temperatures of the day, respectively, h ave r represents the average humidity of the day. ave v represents the solar radiation intensity of the day. ave Represents the average wind speed of the day, and num is a series of unique hot codes, representing the current serial number of the new energy power station.

[0006] To eliminate sparse and discontinuous one-hot codes in the condition vector c, an embedding layer is added after the condition vector c to map it into a low-dimensional and continuous embedded condition vector c'. Then, the embedded condition vector c' is concatenated with Gaussian noise z, normalized, and fed into a multilayer perceptron module. Decoupled feature vector w is generated through multiple fully connected layers. The process is described as follows: (2) Furthermore, the eigenvector w undergoes an affine transformation matrix A i Decomposed into multiple style control parameters y i : (3) Among them, ys,i y is the scaling parameter. b,i For bias parameters, Ultimately, the mapping network G map Output the decoupled style control parameters y i y i It includes new energy output style information generated based on various tag information such as meteorology, geography, and date, and will y i By inputting the synthesis network, different styles of new energy power output scenarios can be generated.

[0007] Synthetic Network G syn A progressive growth model is adopted, using multiple upsampling modules to gradually expand the constant tensor x0' into sample data x' of new energy power output scenarios with corresponding styles, and the style control parameter y i The process involves injecting upsampling modules at different levels to correct parameters and gradually control scene generation. This process can be described as follows: (4) Wherein, the noise vector n i It provides detail diversity and randomness to the feature maps of the generated scene.

[0008] Synthetic Network G syn Each upsampling module in the code consists of an adaptive instance normalization layer, a convolutional layer, and an upsampling layer. The convolutional layer uses ordinary 2D convolution, and the upsampling part uses bilinear interpolation to double the image resolution. AdaIN, as the core layer of the upsampling module, is used to apply the style control parameter y. i In the injected feature map, the main structure of this layer can be described as follows: (5) Where, μ(x) i ), σ(x) i ) are the mean and standard deviation of the feature map, respectively; y s,i y b,i It is the style vector y i The scaling and bias components.

[0009] The discriminator D consists of multiple downsampling modules. Each downsampling module consists of a downsampling layer and a convolutional layer. The downsampling layer uses average pooling to halve the length of the feature map. The convolutional module uses a convolutional layer with residual structure to enhance the feature extraction capability, enabling the convolutional layer to learn more complex features of the sample and avoid the gradient vanishing problem. At the end of the downsampling module, the feature map passes through two fully connected layers, and finally outputs a discriminant value D(x) and D(x') that characterizes the authenticity of the sample.

[0010] R1 regularization is introduced into the loss function of the discriminator D to prevent overfitting by penalizing the gradient of the discriminator D with respect to the real image. The loss function of the discriminator can be expressed as follows: (6) Where ∇xD(x) is the gradient of the discriminator with respect to the real image x, and λ is the regularization coefficient, which controls the weight of the regularization term; For the neural network generator G, path length regularization is introduced. By constraining the generator's mapping network, the changes in the noise space z and the changes in the image space are made smoother, thereby improving the quality and diversity of the generated images. The loss function is as follows: (7) Where γ is the path length regularization coefficient, which controls the weight of the regularization term; Jz is the Jacobian matrix of the generator with respect to the noise vector z; y is the style vector; and a is the target value, which is usually set to 1.

[0011] The quality of the specified new energy output scenario generated by generator G is verified by evaluation indicators, which include time characteristic indicators and spatial characteristic indicators.

[0012] To verify that the specified renewable energy output scenario generated by generator G can reflect the temporal characteristics of the real-world scenario, a comparative analysis was conducted using two methods: daily generation and seasonal generation. Evaluation metrics such as root mean square error, coverage rate, interval width, and probability distribution were selected for assessment. The calculation methods for each metric are as follows: Root mean square error: (8) In the formula, N is the total number of sampling points in the generated scene; For the value of the i-th generated scene, Values ​​for real-world scenarios; Coverage and Range Width: (9) (10) Where N is the total number of generated scene sampling points; N(1-α) is the total number of sampling points falling within the confidence interval at a confidence level of 1-α; , Let represent the maximum and minimum values ​​of the data at the i-th sampling point of the generated scene with a confidence level of 1-α, respectively; Probability density function: (11) in, is the density estimate at point x; K(.) is the kernel function; h w is the bandwidth parameter; n is the number of samples.

[0013] Using the Pearson correlation coefficient, a Pearson correlation matrix is ​​constructed to characterize spatial properties. (12) Where Xi and Yi represent the i-th output values ​​of power stations X and Y, respectively, X(—) and Y(—) represent the average output values ​​of power stations X and Y, respectively, and n represents the total number of new energy power stations.

[0014] Compared with the prior art, the present invention has the following advantages and beneficial effects: by introducing a mapping network to generate decoupled style control parameters and injecting them into the synthesis network layer by layer, the feature entanglement problem is effectively solved, significantly improving the accuracy and controllability of the generated scene. Through the cyclic training of the generator and discriminator, the diversity and realism of the generated scene are further guaranteed. Attached Figure Description

[0015] Figure 1 This is a basic schematic diagram of the CStyleGAN principle of the present invention.

[0016] Figure 2 This is a flowchart illustrating the training process of the CStyleGAN model of the present invention. Figure 3 This is a basic structural diagram of the mapping network of the present invention.

[0017] Figure 4 This is a basic structural diagram of the synthetic network of the present invention.

[0018] Figure 5 This is a basic structural diagram of the discriminator of the present invention.

[0019] Figure 6 Generate result images for landscape scenes using CStyleGAN and C-WGAN-GP.

[0020] Figure 7(a) is a historical wind power frequency distribution diagram of the present invention.

[0021] Figure 7(b) is a frequency distribution diagram of the C-WGAN-GP wind power according to the present invention.

[0022] Figure 7(c) is a wind power frequency distribution diagram of the CStyleGAN of the present invention.

[0023] Figure 8 This is a Pearson correlation matrix diagram of photovoltaic power output data in this invention. Detailed Implementation

[0024] To enable those skilled in the art to better understand the technical solutions of the present invention, preferred embodiments of the present invention are described below in conjunction with specific examples. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote elements with the same or similar functions throughout. However, it should be understood that the drawings are for illustrative purposes only and should not be construed as limiting the present invention. To better illustrate this embodiment, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product size. It is understandable for those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings. The positional relationships described in the drawings are for illustrative purposes only and should not be construed as limiting the present invention.

[0025] The present invention will be further described below with reference to the accompanying drawings and embodiments, but this should not be construed as limiting the present invention.

[0026] The basic principle of the labeling new energy scene generation method based on improved style generative adversarial network of this invention is as follows: Figure 1 As shown.

[0027] To generate more detailed and accurate power output scenarios than conditional generation networks (CRNs) based on known historical power output data, day-ahead predicted power output data, meteorological conditions, time information, and station location data of renewable energy power plants, this technology introduces a mapping network and a progressive growth model into the traditional conditional generation network to achieve conditional decoupling and style control, known as the CStyleGAN model. CStyleGAN is trained using historical power output data, day-ahead predicted power output data, and corresponding data labels from wind and solar power plants to learn the power output distribution characteristics of renewable energy. The model consists of two adversarial neural network generators G (including a mapping network Gmap and a synthesis network Gsyn) and a discriminator D. Its inputs are Gaussian noise z, a conditional vector c composed of meteorological, geographical, and date information, historical power output data p, and predicted power output data f. The final outputs are generated samples x' and the confidence scores used by the discriminator to distinguish between true and false samples. The model training process is as follows: Figure 2 As shown. During training, the generator and discriminator continuously compete against each other. The generator constantly adjusts its output to try to fool the discriminator, while the discriminator continuously updates its parameters to improve its ability to distinguish between real and fake samples. As training progresses, the distribution difference between the generated samples and the real data gradually decreases, and the process terminates when Nash equilibrium is reached. Then, the distribution P' generated by the generator... x It is considered to already represent the true distribution of renewable energy P x Extract the trained generator G, and input specified weather, time, geographical location, and other conditions to generate output scenarios for specific situations.

[0028] Technical Point 1: Feature-Decoupled Mapping Network Structure In order to generate new energy output data with specified labels based on historical output data, a network structure is needed to process the input conditional label data and associate the data labels with various features of historical output.

[0029] Traditional conditional generative adversarial networks (GANs) directly concatenate data labels and historical output data before feeding them into the generator network. This leads to a certain feature entanglement problem, where the mapping relationships between different conditional information and the generated data exhibit some overlap and coupling, causing the generated results to fail to accurately reflect the specified conditions. To improve the feature entanglement problem and achieve accurate and controllable generation, the CStyleGAN model uses a mapping network Gmap to decouple the input features. Gmap's role is to map Gaussian noise z and the conditional vector c into a disentangled latent space through a series of neural network layers, ultimately generating the decoupled style control parameters yi.

[0030] To control various indicators in the generated wind and light scene, the historical power output data, including month (m), weather conditions (w), temperature (t), humidity (h), wind speed (v), solar radiation intensity (r), and station number (num), are used to form a condition vector c, which is specifically composed of: (1) Where m is a 12-bit unique thermal code representing the month, with each bit representing one month. w is a 4-bit unique thermal code representing the weather, with each bit representing one of the four weather conditions: sunny, partly cloudy, cloudy, or rain / snow. tmax, tmin, and tave represent the maximum, minimum, and average temperatures of the day, respectively. h ave This represents the average humidity for that day. ave This represents the intensity of solar radiation on that day. ave This represents the average wind speed for the day. num is a series of unique hot codes, representing the current station's sequence number.

[0031] To adapt to the diverse needs of different research projects for scene generation and achieve flexible and controllable scene generation, the length of the condition vector c can be reduced or expanded according to specific scene requirements. For example, in a scene requiring power output under specific weather conditions, the encoding information representing meteorological conditions in the condition vector c is retained, while other irrelevant encodings are deleted. To eliminate sparse and discontinuous one-hot codes in the condition vector c, this technique sets an embedding layer after the condition vector c, mapping it to a low-dimensional and continuous condition vector c'. Subsequently, c' is concatenated with Gaussian noise z, normalized, and then fed into a multilayer perceptron module, generating a decoupled feature vector w through multiple fully connected layers.

[0032] The process is described as follows: (2) Furthermore, the eigenvector w undergoes an affine transformation matrix Ai Decomposed into multiple style control parameters y i : (3) Among them, y s,i y is the scaling parameter. b,i This is the bias parameter.

[0033] Finally, the Gmap mapping network outputs the decoupled style control parameters y. i y i It includes new energy output style information generated based on various tag information such as meteorology, geography, and date, and will y i By inputting the synthesis network, different styles of new energy power output scenarios can be generated.

[0034] Technical Point 2: Style Synthesis and Discriminant Networks Based on Progressive Structure The purpose of the synthetic network Gsyn is to adjust the style control parameter y. i The corresponding new energy output scenario x' is generated. The purpose of the discriminant network D is to determine the authenticity of the generated scenario x' and update the parameters of the generator and discriminator based on the determination result. Traditional conditional generative networks use a simple concatenated convolutional layer pattern, which suffers from training instability and mode collapse. This technique effectively improves this problem by using a progressive generative structure in Gsyn and D.

[0035] 2.1 Synthetic Network Gsyn The main structure of the synthetic network Gsyn is as follows Figure 4 As shown. Gsyn employs a progressive growth model, using multiple upsampling modules to gradually expand the constant tensor x0' into sample data x' of the corresponding style of new energy power output scenario. Style control parameter y i The process involves injecting upsampling modules at different levels to correct parameters and gradually control scene generation. This process can be described as follows: (4) Wherein, the noise vector n i To provide detail diversity and randomness to the feature maps of the generated scene, y i Style features ranging from coarse to fine are injected into different upsampling levels. Specifically, in the generation of new energy scenes, the style control variable y from the low-level upsampling module is injected. i This determines the overall characteristics of the generated scenario, such as the type of new energy output and overall trends. The intermediate-level style variable y... i This determines more refined characteristics, such as the amplitude of output fluctuations in landscape. And the higher-level style variable y... i This determines the output disturbance situation with a small amplitude over a short period of time.

[0036] Each upsampling module in Gsyn consists of an Adaptive Instance Normalization (AdaIN) layer, a convolutional layer, and an upsampling layer. The convolutional layer uses ordinary 2D convolution, and the upsampling part uses bilinear interpolation to double the image resolution. AdaIN, as the core layer of the upsampling module, is used to apply the style control parameter y... i In the injected feature map, the main structure of this layer can be described as follows: (5) Where μ(xi) and σ(xi) are the mean and standard deviation of the feature map, respectively; y s,i y b,i It is the style vector y i The scaling and bias components.

[0037] 2.2 Discrimination Network D In contrast to the multiple upsampling modules of the synthetic network, the discriminator D of CStyleGAN consists of multiple downsampling modules. The discriminator D receives two types of inputs: one is the real sample x, i.e., historical output data p and current-day predicted output data f; the other is the generated sample x', i.e., the generated output p' and current-day predicted output data f'. A corresponding condition vector c is also required when inputting the samples.

[0038] Each downsampling module of the discriminator consists of a downsampling layer and a convolutional layer. The downsampling layer uses average pooling to halve the length of the feature map, while the convolutional module uses convolutional layers with residual structures to enhance feature extraction capabilities, enabling the convolutional layers to learn more complex features of the samples and avoiding the gradient vanishing problem. At the end of the downsampling module, the feature map passes through two fully connected layers, ultimately outputting a confidence value that characterizes the authenticity of the sample.

[0039] 2.3 Loss Function To make training more stable and improve the quality of generated images, this technique introduces R1 regularization into the discriminator's loss function, based on the original loss function. This prevents overfitting by penalizing the discriminator's gradient with respect to the real image. The discriminator's loss function can be expressed as follows: (6) Where ∇xD(x) is the gradient of the discriminator with respect to the real image x, and λ is the regularization coefficient, which controls the weight of the regularization term.

[0040] For the neural network generator G, path length regularization is introduced. By constraining the generator's mapping network, the changes in the noise space z and the image space become smoother, thereby improving the quality and diversity of the generated images. The loss function is as follows: (7) Where γ is the path length regularization coefficient, controlling the weight of the regularization term; J z y is the Jacobian matrix of the generator with respect to the noise vector z; y is the style vector; a is the target value, which is usually set to 1.

[0041] Technical Point 3: Controllable Generation of Evaluation Indicators and Scenario Generation for New Energy Scenarios To verify whether the scene generation model can accurately reflect the spatiotemporal distribution characteristics of new energy output, this technology constructs a controllable generation evaluation index for new energy scenarios. This index uses both temporal and spatial characteristic indicators as evaluation criteria, and analyzes the model's performance using three methods: daily generation, seasonal generation, and spatial generation. This index can be used to characterize the performance of various scene generation models and has universality.

[0042] 3.1 Time characteristic indicators To verify that the scenes generated by the model can reflect the temporal characteristics of real-world scenes, the models were compared and analyzed using both daily and seasonal generation methods. Evaluation metrics such as root mean square error (RMSE), coverage and interval width, and probability distribution were selected to assess the scenes. The calculation methods for each metric are as follows: (1) Root Mean Square Error (RMSE) Root mean square error and mean absolute error are used to evaluate the magnitude of the error between the generated scene and the real scene; the smaller the value, the more accurate it is.

[0043] (8) In the formula, N is the total number of sampling points in the generated scene; For the value of the i-th generated scene, The values ​​represent the actual scene.

[0044] (2) Coverage and interval width Coverage (CR) and Interval Width (PIW) reflect the magnitude of the statistical error between the current-day scene set and the real scene, indicating the ability of the generated scene to describe the uncertainty of historical output. A higher coverage and a smaller interval width result in a more accurate description of historical uncertainty by the model. The calculation method is as follows: (9) (10) Where N is the total number of generated scene sampling points; N(1-α) is the total number of sampling points falling within the confidence interval at a confidence level of 1-α; , These represent the maximum and minimum values ​​of the data at the i-th sampling point of the generated scenario with a confidence level of 1-α, respectively. Coverage and interval width together describe the ability of the generated scenario to describe historical output uncertainty. The larger the coverage and the smaller the interval width, the more accurately the model describes historical uncertainty.

[0045] (3) Probability distribution By comparing the probability distribution histograms and probability density functions of the generated scene and the real scene, we can verify whether the model can correctly learn the distribution characteristics of the real scene. The probability density function uses kernel density estimation, and the calculation method is as follows: (11) in, is the density estimate at point x; K(.) is the kernel function; hw is the bandwidth parameter; n is the number of samples.

[0046] 3.1.1 Daily Generation Daily generation refers to generating power output scenarios under given weather data for a specific day. Using the integrated wind and photovoltaic datasets provided by the National Renewable Energy Laboratory (NREL), wind power data from March 19, 2011, and photovoltaic data from April 12, 2011, were selected as validation data for a specific site. Based on the meteorological information for that day, 90 scenarios were generated using the CStyleGAN and C-WGAN-GP models respectively, and the coverage, interval width, and RMSE data of the scenarios generated by each model were calculated using equations (12) and (13). The results... Figure 6 As shown in Tables 1 and 2.

[0047]

[0048]

[0049] from Figure 6 It can be seen that the predicted power based on numerical weather prediction is relatively smooth, while the actual power exhibits small fluctuations. Based on... Figure 6It can be seen that because C-WGAN-GP directly concatenates the conditional vector with the real data before feeding it into the network, its weather labels and predicted power are highly coupled. The generated scenarios are closer to the predicted power values, but the data is concentrated on inaccurate day-ahead predicted power, failing to capture the mapping relationship between predicted and actual values. The generated scenarios have a larger RMSE value and a wider interval width at the same coverage and confidence level, resulting in a coarser characterization of renewable energy distribution. In contrast, CStyleGAN generates power output data with a smaller interval width and lower error, exhibiting the best accuracy. By introducing a mapping network to decouple the meteorological labels, it reduces the impact of erroneous predicted power output data on the generated scenarios to some extent, and its generated scenarios reflect the small fluctuations in actual power output. As shown in Table 1, the RMSE error value of CStyleGAN is reduced by approximately 40% compared to the traditional C-WGAN-GP, demonstrating its ability to capture the complex mapping relationship between weather, predicted power output data, and real data.

[0050] 3.1.2 Seasonal Generation Seasonal generation refers to generating power output scenarios for a given season based on its time information. Two models were used, with four seasonal labels: spring (March, April, May), summer (June, July, August), autumn (September, October, November), and winter (December, January, February). 1200 wind power output scenarios were generated for each season. The frequency distribution for each season was statistically analyzed, and corresponding frequency distribution histogram curves were plotted. The results are shown in Figures 7(a) to 7(c). Figure 7(a) shows the historical wind power frequency distribution, Figure 7(b) shows the C-WGAN-GP wind power frequency distribution, and Figure 7(c) shows the CStyleGAN wind power frequency distribution.

[0051] As can be seen from the frequency distribution plots, the power distribution trends of the three scenarios are quite similar to the actual power, reflecting the temperate maritime climate characteristics of the northwestern coastal region of the United States, characterized by less wind and more sunlight in summer and more wind and less sunlight in winter. According to Figure 7(b), the C-WGAN-GP model has larger errors in the less frequently distributed power range of 0.4–0.7, with data concentrated in the high-output and low-output ranges where the samples appear more frequently, indicating a certain degree of mode collapse. According to Figure 7(c), the frequency distribution of the data generated by CStyleGAN is basically close to the actual data in each range, better capturing the mapping relationship between season and power output. From the frequency distribution histogram, it can be seen that CStyleGAN effectively reduces the mode collapse problem through the progressive synthesis network Gsyn and the discriminant network D, and its generated scenario diversity is significantly improved compared to traditional models.

[0052] 3.2 Spatial Characteristic Indicators Because meteorological conditions are relatively consistent within the same region, the output of different renewable energy power plants in neighboring areas exhibits strong spatial correlation. To verify whether the model can learn the spatial characteristics of renewable energy distribution, the Pearson correlation coefficient is used to construct a Pearson correlation matrix to characterize this feature. The more similar the correlation matrices are, the stronger their ability to capture the spatial distribution of renewable energy.

[0053] (12) Where Xi and Yi represent the i-th power output values ​​of power stations X and Y, respectively. X(—) and Y(—) represent the average power output values ​​of power stations X and Y, respectively. n represents the total number of renewable energy power stations.

[0054] Taking 15 geographically close photovoltaic (PV) farms in Washington State from the NREL dataset as an example, two models were used to generate PV output data for the same day, and their Pearson correlation matrices were calculated separately. The results are as follows: Figure 8 As shown.

[0055] Depend on Figure 8 As can be seen, the distribution trends of the scenarios generated by C-WGAN-GP differ significantly from historical power output data, and the correlation matrix shows low similarity to historical distributions. In contrast, the scenarios generated by CStyleGAN show greater similarity to historical scenarios in terms of intraday variation trends and distribution correlation, accurately reflecting the spatial location characteristics of each site. Taking sites 6-9 as an example, their geographical locations are relatively close, thus this high correlation can be observed in both historical power output data and the CStyleGAN correlation matrix, a feature not reflected in the C-WGAN-GP model. This indicates that the CStyleGAN model has a certain advantage in capturing the spatial distribution characteristics of renewable energy power plants and can more effectively reflect these characteristics.

[0056] Baseline method: The C-WGAN-GP model based on traditional conditional generative adversarial networks directly concatenates data labels with historical power output data and feeds them into the generator network. New energy power output scenarios are generated through adversarial training between the generator and discriminator. This is generally used as a common technique for the controllable generation of new energy scenarios. Although its structure is simple and easy to implement, it suffers from significant feature entanglement, leading to training instability, pattern collapse, and limited generalization ability (as shown in Figure 7). Figure 8 (As already shown in the comparison method), and the generated scenario has a large root mean square error (RMSE) and a larger interval width at the same confidence level, which limits its ability to characterize the seasonal characteristics of renewable energy and the spatial correlation of the power station (Table 1, Table 2).

[0057] The main difference between the CStyleGAN model proposed in this invention and its predecessor lies in the decoupling of the mapping network implementation conditions and output features, the adoption of a progressive generation structure in the generator and discriminator, and the targeted optimization of the loss function. The main improvements are: firstly, the addition of a mapping network G... map First, Gaussian noise and conditional vectors are mapped to a disentangled latent space to eliminate feature entanglement and improve the accuracy of controllable generation. Second, a progressive structure is adopted in the generator and discriminator to improve training instability and mode collapse, thereby enhancing the model's generalization ability. Third, the loss function is improved: R1 regularization is introduced in the discriminator to prevent overfitting, and path length regularization is introduced in the generator to make the changes between the noise space and the image space smoother, further improving the quality and diversity of the generated scenes.

[0058] By generating (in a single day) Figure 6 Season generation (Figure 7), spatial coordinate generation ( Figure 8 Comparing the three generation modes, it can be seen that the CStyleGAN model proposed in this invention has performance advantages in both time and space metrics compared to the traditional C-WGAN-GP model, and has better results in generating tagged landscape scenes.

[0059] Based on the description and accompanying drawings of this invention, those skilled in the art can easily create or use the labeling new energy scene generation method based on an improved style generative adversarial network of this invention, and can produce the positive effects described in this invention.

[0060] Furthermore, in practicing the claims of this invention, those skilled in the art can understand and influence variations to the disclosed embodiments through a study of the drawings, the disclosure, and the appended claims. Additionally, in the claims and description, words such as "comprising" and "containing" do not exclude other elements or steps, and non-plural nouns do not exclude their plural forms.

[0061] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of the present invention. All equivalent changes and modifications made in accordance with the present invention are covered by the scope of the claims of the present invention, and will not be listed here.

Claims

1. A method for generating labeled new energy scenarios based on an improved style generative adversarial network, characterized in that, Includes the following steps: Step 1: Obtain the condition vector c and the real data x of the new energy power output scenario. The real data x of the new energy power output scenario includes historical power output data p and predicted power output data f. The condition vector c includes meteorological data, time information and new energy power station location data. Step 2: Construct the CStyleGAN model, which includes a neural network generator G and a discriminator D. The neural network generator G includes a mapping network G. map With the synthetic network G syn ; Step 3: Input the Gaussian noise z and the conditional vector c into the mapping network G. map Through the mapping network G map Generate decoupled style control parameters y i ;; Step 4: Set the style control parameter y i Layer-by-layer injection of the synthetic network G syn Generate sample data of new energy power output scenarios x' ; Step 5: Input the real data x of the new energy power output scenario, the sample data x' of the new energy power output scenario, and the condition vector c into the discriminator D, and output the discriminant values ​​D(x) and D(x') through the discriminator D. Step 6: Calculate the loss function based on the discriminant value and update the parameters of the discriminator D and the parameters of the neural network generator G; Step 7: Repeat steps 3 to 6. As training continues, the distribution difference between the generated samples and the real data will gradually decrease. When Nash equilibrium is reached, the repetition will terminate. Extract the generator G after training is completed, which can then be used to generate the specified new energy power output scenario.

2. The method for generating labeled new energy scenarios based on an improved style generative adversarial network as described in claim 1, characterized in that, To control various indicators in the generated wind and light scene, the historical power output data, including month (m), weather conditions (w), temperature (t), humidity (h), wind speed (v), solar radiation intensity (r), and station number (num), are used to form a condition vector c, which is specifically composed of: （1） Where m is a 12-bit unique hot code representing the month, with each bit representing a month; w is a 4-bit unique hot code representing the weather, with each bit representing one of the four weather conditions: sunny, partly cloudy, cloudy, or rain / snow; t max , t min , t ave These represent the maximum, minimum, and average temperatures of the day, respectively, h ave r represents the average humidity of the day. ave v represents the solar radiation intensity of the day. ave Represents the average wind speed of the day, and num is a series of unique hot codes, representing the current serial number of the new energy power station.

3. The method for generating labeled new energy scenarios based on an improved style generative adversarial network as described in claim 2, characterized in that, To eliminate sparse and discontinuous one-hot codes in the condition vector c, an embedding layer is added after the condition vector c to map it into a low-dimensional and continuous embedded condition vector c'. Then, the embedded condition vector c' is concatenated with Gaussian noise z, normalized, and fed into a multilayer perceptron module. Decoupled feature vector w is generated through multiple fully connected layers. The process is described as follows: （2） Furthermore, the eigenvector w undergoes an affine transformation matrix A i Decomposed into multiple style control parameters y i : （3） Among them, y s,i y is the scaling parameter. b,i For bias parameters, Ultimately, the mapping network G map Output the decoupled style control parameters y i y i It contains new energy output style information generated based on various label information such as meteorology, geography and date. By inputting yi into the synthesis network, different styles of new energy output scenarios can be generated.

4. The method for generating labeled new energy scenarios based on an improved style generative adversarial network as described in claim 1, characterized in that, Synthetic Network G syn A progressive growth model is adopted, using multiple upsampling modules to gradually expand the constant tensor x0' into sample data x' of new energy power output scenarios with corresponding styles, and the style control parameter y i The process involves injecting upsampling modules at different levels to correct parameters and gradually control scene generation. This process can be described as follows: （4） Wherein, the noise vector n i It provides detail diversity and randomness to the feature maps of the generated scene.

5. The method for generating labeled new energy scenarios based on an improved style generative adversarial network as described in claim 4, characterized in that, Each upsampling module in the Gsyn synthesis network consists of an adaptive instance normalization layer, a convolutional layer, and an upsampling layer. The convolutional layers use ordinary 2D convolutions, and the upsampling part uses bilinear interpolation to double the image resolution. AdaIN, as the core layer of the upsampling module, is used to apply the style control parameter y. i In the injected feature map, the main structure of this layer can be described as follows: （5） Where, μ(x) i ), σ(x i ) are the mean and standard deviation of the feature map, respectively; y s,i y b,i It is the style vector y i The scaling and bias components.

6. The method for generating labeled new energy scenarios based on an improved style generative adversarial network as described in claim 1, characterized in that, The discriminator D consists of multiple downsampling modules. Each downsampling module consists of a downsampling layer and a convolutional layer. The downsampling layer uses average pooling to halve the length of the feature map. The convolutional module uses a convolutional layer with residual structure to enhance the feature extraction capability, enabling the convolutional layer to learn more complex features of the sample and avoid the gradient vanishing problem. At the end of the downsampling module, the feature map passes through two fully connected layers, and finally outputs a discriminant value D(x) and D(x') that characterizes the authenticity of the sample.

7. The method for generating labeled new energy scenarios based on an improved style generative adversarial network as described in claim 6, characterized in that, R1 regularization is introduced into the loss function of the discriminator D to prevent overfitting by penalizing the gradient of the discriminator D with respect to the real image. The loss function of the discriminator can be expressed as follows: （6） Where ∇xD(x) is the gradient of the discriminator with respect to the real image x, and λ is the regularization coefficient, which controls the weight of the regularization term; For the neural network generator G, path length regularization is introduced. By constraining the generator's mapping network, the changes in the noise space z and the changes in the image space are made smoother, thereby improving the quality and diversity of the generated images. The loss function is as follows: （7） Where γ is the path length regularization coefficient, controlling the weight of the regularization term; J z y is the Jacobian matrix of the generator with respect to the noise vector z; y is the style vector; a is the target value, which is usually set to 1.

8. The method for generating labeled new energy scenarios based on an improved style generative adversarial network as described in claim 1, characterized in that, The quality of the specified new energy output scenario generated by generator G is verified by evaluation indicators, which include time characteristic indicators and spatial characteristic indicators.

9. The method for generating labeled new energy scenarios based on an improved style generative adversarial network as described in claim 8, characterized in that, To verify that the specified renewable energy output scenario generated by generator G can reflect the temporal characteristics of the real-world scenario, a comparative analysis was conducted using two methods: daily generation and seasonal generation. Evaluation metrics such as root mean square error, coverage rate, interval width, and probability distribution were selected for assessment. The calculation methods for each metric are as follows: Root mean square error: （8） In the formula, N is the total number of sampling points in the generated scene; For the value of the i-th generated scene, Values ​​for real-world scenarios; Coverage and Range Width: （9） （10） Where N is the total number of generated scene sampling points; N(1-α) is the total number of sampling points falling within the confidence interval at a confidence level of 1-α; , Let represent the maximum and minimum values ​​of the data at the i-th sampling point of the generated scene with a confidence level of 1-α, respectively; Probability density function: （11） in, is the density estimate at point x; K(.) is the kernel function; hw is the bandwidth parameter; n is the number of samples.

10. The method for generating labeled new energy scenes based on an improved style generative adversarial network as described in claim 8, characterized in that, Using the Pearson correlation coefficient, a Pearson correlation matrix is ​​constructed to characterize spatial properties. （12） Among them, X i and Y i Let X and Y represent the i-th output values ​​of the stations respectively. and Let X and Y represent the average power output of power stations X and Y, respectively, and n represent the total number of new energy power stations.

Images