Wind and light scene generation method based on causal convolution generative adversarial network

Abstract

The invention discloses a wind and light output scene generation method based on a causal convolutional generative adversarial network, and belongs to the technical field of new energy power generation. According to the method, a generative adversarial network model taking a time convolutional network as a core is constructed aiming at the defects of an existing wind and light output scene generation technology in the aspects of authenticity, diversity and long-time-sequence modeling efficiency, both a generator and a discriminator adopt a causal convolutional structure, and an RF-Sched receptive field scheduling strategy is introduced, so that the generation efficiency of the wind and light output scene is improved. Long and short term dependence and periodicity characteristics of the wind and light output sequence are effectively captured, and key information omission is avoided. Firstly, daily segmentation and cleaning are performed on historical wind and light data, a training set is constructed, then a generator and a discriminator are alternately trained until a model is converged, and finally an optimal model is loaded to generate a wind and light output scene. Experimental results show that the method is superior to traditional GAN, WGAN and VAE models in indexes such as MAE, RMSE and FID, high-authenticity and diversified wind and light output scenes can be generated, and reliable support is provided for planning and dispatching of a power system.

Description

Technical Field

[0001] This invention relates to the field of new energy power generation, specifically to a method for generating wind and solar scenes based on causal convolutional generative adversarial networks. Background Technology

[0002] In recent years, new energy sources, represented by wind and solar power, have been integrated into the power system on a large scale. However, the randomness, volatility, and intermittency of these new energy sources pose challenges to the stable operation of the power system and increase the difficulty of power system planning and dispatch. Therefore, how to characterize the output distribution patterns of wind and solar power and generate high-quality wind and solar power output scenarios has become a critical issue that urgently needs to be addressed.

[0003] To date, scholars both domestically and internationally have conducted extensive research on the generation technology of wind and solar power output scenarios in new energy power generation. Existing methods can be mainly divided into two categories.

[0004] One type is based on statistical modeling methods. Common methods include Markov chains, Copula functions, and probability distribution fitting. These methods generate typical power scenarios by statistically modeling historical data, which can reflect the uncertainty of photovoltaic output to a certain extent. However, when historical data is scarce or missing, it is difficult to simultaneously capture the temporal characteristics and extreme fluctuations of wind and solar power, resulting in limited realism and diversity of the generated scenarios.

[0005] Another category is machine learning-based methods. In recent years, researchers have introduced neural networks and generative models, such as recurrent neural networks, long short-term memory networks, variational autoencoders, and generative adversarial networks. These methods can automatically extract temporal features through machine learning, especially generative adversarial networks, which have shown advantages in improving the diversity of generated sequences and characterizing nonlinear features.

[0006] However, existing methods also have significant drawbacks: recurrent networks such as Long Short-Term Memory (LSTM) have limited efficiency in long-term sequence modeling and struggle to balance intraday cycles with short-term fluctuations. Traditional generative adversarial networks are prone to pattern collapse during training, resulting in a lack of diversity in the generated scenarios.

[0007] In summary, existing methods for generating photovoltaic power scenarios based on statistical modeling and machine learning still have shortcomings in terms of realism, stability, and physical consistency. In power systems where the proportion of renewable energy is continuously increasing, these methods struggle to meet the demands for high-quality, typical scenarios in dispatch and planning. Summary of the Invention

[0008] The purpose of this invention is to propose a causal convolutional generative adversarial network (GAN) and apply it to solve the problem of extracting wind and solar power output scenarios in new power systems. This invention constructs a scenario generation method based on the causal convolutional GAN ​​to generate representative and diverse photovoltaic and wind power output scenarios, providing a scientific basis for power planning and dispatching operation of new power systems.

[0009] This invention provides a method for generating wind and solar power output scenes based on causal convolutional generative adversarial networks, the method comprising the following steps: S1. Collect historical landscape output data and perform data preprocessing. S1.1 Data segmentation of historical landscape output Wind and solar power output exhibits a diurnal periodicity. The original historical wind and solar power output data are divided into N daily wind power output data points and N daily solar power output data points, corresponding to 24 hours. Each data point contains 24 data points. Combining these data points yields a wind power output data set D of size N×24. W Photovoltaic power output data D S .

[0010] S1.2 Perform data cleaning on wind and solar power output data. The wind and solar power output dataset obtained above contains missing data and data anomalies. First, the missing data points are marked, and Newton's interpolation method is used to complete the data. Then, data points with values ​​less than 0 are checked and their values ​​are set to 0. This yields the wind power output data D after preprocessing. W Photovoltaic power output data D S .

[0011] S1.3 Constructing the Noise Dataset D n After training, the causal convolutional generative adversarial network generates wind and solar power output scenarios by receiving a noisy dataset. Therefore, an N×24 noise dataset D of the same size as the wind and solar power output datasets is constructed. n Each data point in this dataset follows a standard normal distribution, as shown below:

[0012] S2. Construct and design the generator network architecture The generator's main function is to generate sequential data that resembles real-world landscape power output scenarios. It accepts a noisy dataset D sampled from a standard normal distribution. n As input, and through a series of neural network layers, a set of scenes that conforms to the power distribution of real landscapes is generated. The generator's network architecture consists of several gated residual blocks (GLU-RES), linear layers, and activation functions.

[0013] S2.1 Configure the RF-Sched receptive field scheduling strategy While fixed exponential expansion can rapidly increase the receptive field, it can also lead to sampling voids: for short slope changes and spikes in the wind and solar power output sequence, some layers may not sample key locations, easily missing crucial changes over a short period. This invention introduces an RF-Sched receptive field scheduling strategy to reduce the risk of void sampling.

[0014] During a fixed exponential expansion, d l =1, 2, 4, ..., 32, exhibiting an exponential variation. With increasing layer number, causal convolutional layers can identify more distant historical features in the wind and solar power output sequence. Under the RF-Sched receptive field scheduling strategy, d l =1, 2, 4, 8|1, 2, 4|8, 16, 32, that is, the initial expansion is used to quickly identify the overall characteristics of the solar power generation sequence. Taking photovoltaic as an example, it can identify the basic shape (arch) of the photovoltaic power generation sequence. Then the expansion rate is pulled back for detailed scanning to avoid ignoring small features (small gaps in the photovoltaic power generation sequence when it is covered by clouds). Finally, the large expansion rate in the later stage is restored to ensure the identification of more distant historical features.

[0015] S2.2 Design of gated residual block structure Each gated residual block contains two layers of causal convolutions and gated linear units (GLUs), Layer Normalization, and Dropout. Let the block input be h. in The output is h out .

[0016] First layer of causal convolution and gated linear unit:

[0017] Second-layer causal convolution and gated linear unit:

[0018] Residual synthesis:

[0019] in For causal convolutional layers, is the sigmoid activation function, ⊙ is element-wise multiplication, LayerNorm is layer normalization, and Dropout is the dropout layer.

[0020] The expression for a single-layer causal convolutional layer is as follows:

[0021] in, The output of the causal convolutional layer. Here, k represents the kernel weights, k is the kernel length, and d represents the kernel weights. l The expansion rate of the lth layer, The input sequence contains historical data at the current time step t and the kernel offset k.

[0022] The causal convolutional layer incorporates the RF-Sched receptive field scheduling strategy from step 2.1, which efficiently captures the long-short-term dependencies of time series. Furthermore, compared to traditional recurrent neural networks or long short-term memory networks, it exhibits higher parallelism. Each operation of the causal convolutional layer is based on the current time step and previous time steps, avoiding information leakage. Moreover, by introducing the RF-Sched receptive field scheduling strategy, it can more accurately capture the periodicity and implicit features of wind power and photovoltaic output sequences.

[0023] S2.3 Constructing Linear Layers and Activation Functions The generator network structure ultimately consists of linear layers and activation functions. Based on the results of the gated residual block structure, the linear layers are constructed, and their computation process can be expressed as follows:

[0024] Where Y is the output matrix of the linear layer, X is the input matrix, W is the weight matrix, and b is the bias vector. f The output matrix is ​​the activation function, and LeakyReLU is the selected activation function, which is expressed as max(αY, Y), where α is the negative slope, usually taken as 0.01. This operation maps all negative numbers to a multiple of α, while positive numbers remain unchanged.

[0025] S2.4 Generator Loss Function Generator loss function:

[0026] Where D(G(z)) is the score of the discriminator on the generated data, and the goal of the generator is to make the discriminator score the generated samples as high as possible.

[0027] S3. Design the discriminator network architecture The main function of the discriminator is to evaluate the authenticity of the input samples. It receives real landscape scene data or landscape scene data generated by the generator, and outputs a scalar score through its internal network structure to measure the Wasserstein distance between the generated data and the real data, i.e., the authenticity of the input data. In this invention, the discriminator adopts a similar architecture to the generator, but consists of only a few causal convolutional layers.

[0028] The loss function of the discriminator is as follows:

[0029] Where D(x) and D(G(z)) are the outputs of the discriminator on the real data x and the generated data G(z), respectively. These are the weights of the gradient penalty term, ensuring that the gradient of the discriminator does not become excessively large during training. The interpolated sample is generated by interpolating between the real sample x and the generated sample G(z), and the gradient penalty is calculated.

[0030] The gradient penalty term is the third term in the discriminator loss function. It is used to ensure the Lipschitz continuity of the discriminator and avoid instability during training. For the interpolation sample, the expression is:

[0031] in, The gradient penalty term ensures that the gradient changes of the discriminator are smooth, thereby improving the stability of training.

[0032] S4. Training the generator and discriminator After designing the generator and discriminator, train them according to the set procedure.

[0033] S4.1 Discriminator Training When training the discriminator, a batch of real samples are taken from the dataset, and a generator is used to map noise into a batch of fake samples. The discriminator obtains a score for real and fake samples, and randomly interpolates between real and fake samples to obtain a batch of transitional samples. This batch of transitional samples is used to penalize the input gradient of the discriminator to approximate the 1-Lipschitz constraint. Then, only the parameters of the discriminator are updated.

[0034] S4.2 Generator Training The noise is resampled to generate fake samples using a generator, and the generator is trained so that the fake samples get higher scores in the discriminator. At this point, the generator parameters are updated.

[0035] S4.3 Loop Training In each iteration, the discriminator is trained multiple times according to the training ratio, and then the generator is trained once. This continues until the iteration condition is met, at which point the parameters of the generator and discriminator are recorded.

[0036] S5. Generate typical landscape power output scenes Load the optimal generator parameters and set the number of scenes to be generated, N. G The generated power output sequence is obtained by sampling noise. The generated result is then inversely normalized and subjected to necessary physical post-processing to obtain a typical wind and solar power output scenario. Evaluation indicators are then selected to assess it.

[0037] Compared with the prior art, the beneficial effects of the present invention include: (1) This invention proposes a method for generating wind and solar power output scenarios based on causal convolutional generative adversarial networks. When there are few historical wind and solar power output scenarios, the historical scenario set is trained by using causal convolutional generative adversarial networks and RF-Sched receptive field scheduling strategy to fully extract the historical wind and solar power output features and generate typical wind and solar power output scenarios. This has important guiding significance for the planning and operation of the new power system with high wind and solar power integration.

[0038] (2) Compared with traditional scene generation methods, the proposed scene generation method significantly improves the quality of the wind and solar power output scene generated by the causal convolutional network, which can well reflect the authenticity and diversity of the wind and solar power output scene, and is superior to the comparison model in all indicators.

[0039] (3) Wind power and photovoltaic power output scenarios are often quite different. The RF-Sched receptive field scheduling strategy proposed in this invention can quickly identify the overall characteristics of wind and solar power output sequences and avoid ignoring the small features in the output sequence. It can better grasp the output characteristics of the two energy sources at the same time. The generated high-quality wind and solar power output scenarios can be quickly applied to new power systems and have strong practicality and versatility. Attached Figure Description

[0040] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0041] Figure 1 This is a schematic diagram of the overall process of the present invention.

[0042] Figure 2 This is a schematic diagram of the generator network structure of the present invention.

[0043] Figure 3 This is a schematic diagram of the discriminator network structure of the present invention.

[0044] Figure 4 This is a schematic diagram of the causal convolutional generative adversarial network in this invention.

[0045] Figure 5 This is a schematic diagram of a real photovoltaic power output scenario.

[0046] Figure 6 This is a schematic diagram of the photovoltaic power output scenario generated by the method of the present invention.

[0047] Figure 7 A schematic diagram of a photovoltaic power output scenario generated by the WGAN algorithm in existing technologies.

[0048] Figure 8 This is a schematic diagram of a photovoltaic power output scenario generated by the GAN algorithm in existing technologies.

[0049] Figure 9 This is a schematic diagram of a photovoltaic power output scenario generated by the VAE algorithm in existing technologies.

[0050] Figure 10 This is a schematic diagram of a real wind power output scenario.

[0051] Figure 11 This is a schematic diagram of the wind power output scenario generated by the method of the present invention.

[0052] Figure 12 This is a schematic diagram of wind power output scenarios generated by the WGAN algorithm in existing technologies.

[0053] Figure 13 This is a schematic diagram of wind power output scenarios generated by the GAN algorithm in existing technologies.

[0054] Figure 14 This is a schematic diagram of wind power output scenarios generated by the VAE algorithm in existing technologies. Detailed Implementation

[0055] Example 1 The landscape scene generation method based on causal convolutional generative adversarial networks described in this invention has the following overall flowchart: Figure 1 As shown, the specific implementation is as follows: S1. Collect historical photovoltaic power output data and perform data preprocessing. The implementation example uses photovoltaic power output data from a certain location for training. The photovoltaic power output data used is from January 1, 2020 to November 30, 2024.

[0056] S1.1 Data segmentation of historical photovoltaic output Photovoltaic power output exhibits a diurnal periodicity. The original historical photovoltaic power output data was divided into daily segments, yielding 1796 daily photovoltaic power output data points, corresponding to 24 hours, with each data point containing 24 data points. The combined data resulted in a photovoltaic power output data set D of size 1796 × 24. s .

[0057] S1.2 performs data cleaning on the photovoltaic output data. The photovoltaic output dataset obtained above contains missing data and data anomalies. First, the missing data points are marked, and Newton's interpolation method is used to complete the data. Then, data points with values ​​less than 0 are checked and their values ​​are set to 0. This yields the photovoltaic output data D after preprocessing. s .

[0058] S1.3 Constructing the Noise Dataset D n After training, the causal convolutional generative adversarial network generates photovoltaic power output scenarios by receiving a noisy dataset. Therefore, an N×24 noisy dataset D of the same size as the photovoltaic power output dataset is constructed. nEach data point in this dataset follows a standard normal distribution, as shown below:

[0059] S2. Construct and design the generator network architecture The generator's main function is to generate sequential data that resembles real-world landscape power output scenarios. It accepts a noisy dataset D sampled from a standard normal distribution. n As input, and through a series of neural network layers, a scene set that conforms to the power distribution of real-world landscapes is generated. The generator's network architecture consists of several gated residual blocks (GLU-RES), linear layers, and activation functions, and its architecture design is as follows: Figure 2 As shown.

[0060] S2.1 Configure the RF-Sched receptive field scheduling strategy While fixed exponential expansion can rapidly increase the receptive field, it can also lead to sampling voids: for short slope changes and spikes in the wind and solar power output sequence, some layers may not sample key locations, easily missing crucial changes over a short period. This invention introduces an RF-Sched receptive field scheduling strategy to reduce the risk of void sampling.

[0061] During a fixed exponential expansion, d l =1, 2, 4, ..., 32, exhibiting an exponential variation. With increasing layer number, causal convolutional layers can identify more distant historical features in the wind and solar power output sequence. Under the RF-Sched receptive field scheduling strategy, d l =1, 2, 4, 8|1, 2, 4|8, 16, 32, that is, the initial expansion is used to quickly identify the overall characteristics of the solar power generation sequence. Taking photovoltaic as an example, it can identify the basic shape (arch) of the photovoltaic power generation sequence. Then the expansion rate is pulled back for detailed scanning to avoid ignoring small features (small gaps in the photovoltaic power generation sequence when it is covered by clouds). Finally, the large expansion rate in the later stage is restored to ensure the identification of more distant historical features.

[0062] S2.2 Design of gated residual block structure Each gated residual block contains two layers of causal convolutions and gated linear units (GLUs), Layer Normalization, and Dropout. Let the block input be h. in The output is h out .

[0063] First layer of causal convolution and gated linear unit:

[0064] Second-layer causal convolution and gated linear unit:

[0065] Residual synthesis:

[0066] in For causal convolutional layers, is the sigmoid activation function, ⊙ is element-wise multiplication, LayerNorm is layer normalization, and Dropout is the dropout layer.

[0067] The expression for a single-layer causal convolutional layer is as follows:

[0068] in, The output of the causal convolutional layer. Here, k represents the kernel weights, k is the kernel length, and d represents the kernel weights. l The expansion rate of the lth layer, The input sequence contains historical data at the current time step t and the kernel offset k.

[0069] The causal convolutional layer incorporates the RF-Sched receptive field scheduling strategy from step 2.1, which efficiently captures the long-short-term dependencies of time series. Furthermore, compared to traditional recurrent neural networks or long short-term memory networks, it exhibits higher parallelism. Each operation of the causal convolutional layer is based on the current time step and previous time steps, avoiding information leakage. Moreover, by introducing the RF-Sched receptive field scheduling strategy, it can more accurately capture the periodicity and implicit features of wind power and photovoltaic output sequences.

[0070] S2.3 Constructing Linear Layers and Activation Functions The generator network structure ultimately consists of linear layers and activation functions. Based on the results of the gated residual block structure, the linear layers are constructed, and their computation process can be expressed as follows:

[0071] Where Y is the output matrix of the linear layer, X is the input matrix, W is the weight matrix, and b is the bias vector. f The output matrix is ​​the activation function, and LeakyReLU is the selected activation function, expressed as max(αY, Y), where α is the negative slope, typically 0.01. This operation maps all negative numbers to a multiple of α, while positive numbers remain unchanged.

[0072] S2.4 Generator Loss Function Generator loss function:

[0073] Where D(G(z)) is the score of the discriminator on the generated data, and the goal of the generator is to make the discriminator score the generated samples as high as possible.

[0074] S3. Design the discriminator network architecture The discriminator's main function is to evaluate the authenticity of the input samples. It receives real landscape scene data or landscape scene data generated by the generator, and outputs a scalar score through its internal network structure to measure the Wasserstein distance between the generated data and the real data, i.e., the authenticity of the input data. In this invention, the discriminator adopts a similar architecture to the generator, but consists of only a few causal convolutional layers, as shown in the following architecture design. Figure 3 As shown.

[0075] The loss function of the discriminator is as follows:

[0076] Where D(x) and D(G(z)) are the outputs of the discriminator on the real data x and the generated data G(z), respectively. These are the weights of the gradient penalty term, ensuring that the gradient of the discriminator does not become excessively large during training. The interpolated sample is generated by interpolating between the real sample x and the generated sample G(z), and the gradient penalty is calculated.

[0077] The gradient penalty term is the third term in the discriminator loss function. It is used to ensure the Lipschitz continuity of the discriminator and avoid instability during training. For the interpolation sample, the expression is:

[0078] in, The gradient penalty term ensures that the gradient changes of the discriminator are smooth, thereby improving the stability of training.

[0079] S4. Training the generator and discriminator After designing the generator and discriminator, train them according to the set process. The overall architecture diagram of the causal convolutional generative adversarial network is as follows: Figure 4 As shown.

[0080] S4.1 Discriminator Training When training the discriminator, a batch of real samples are taken from the dataset, and a generator is used to map noise into a batch of fake samples. The discriminator obtains a score for real and fake samples, and randomly interpolates between real and fake samples to obtain a batch of transitional samples. This batch of transitional samples is used to penalize the input gradient of the discriminator to approximate the 1-Lipschitz constraint. Then, only the parameters of the discriminator are updated.

[0081] S4.2 Generator Training The noise is resampled to generate fake samples using a generator, and the generator is trained so that the fake samples get higher scores in the discriminator. At this point, the generator parameters are updated.

[0082] S4.3 Loop Training In each iteration, the discriminator is trained multiple times according to the training ratio, and then the generator is trained once. This continues until the iteration condition is met, at which point the parameters of the generator and discriminator are recorded.

[0083] S5. Generate typical photovoltaic power output scenarios Load the optimal generator parameters and set the number of scenes to be generated, N. G The generated power output sequence is obtained by sampling noise. The generated result is then inversely normalized and subjected to necessary physical post-processing to obtain a typical wind and solar power output scenario. Evaluation indicators are then selected to assess it.

[0084] S5.1 Selection of Indicators for Evaluating the Quality of Landscape Output Scene Generation Mean absolute error (MAE), root mean square error (RMSE), and Fréchet inception distance (FID) were selected as indicators to evaluate the quality of landscape scene generation. The smaller the MAE, RMSE, and FID, the higher the quality of the generated scene.

[0085] S5.2 Causal Convolutional Generative Adversarial Network Parameter Settings The algorithm was written in Python and ran on Windows 11. The batch size was set to 64, the epochs to 800, the learning rate to 0.0001, and the training ratio to 5.

[0086] S5.3 Results Analysis Typical landscape power scenes were generated using a causal convolutional generative adversarial network (CCGAN). Additionally, three classic generative algorithms—WGAN, GAN, and VAE—were trained on the same dataset to generate typical landscape power scenes.

[0087] The evaluation results of the generation performance of the causal convolutional generative adversarial network CCGAN and the comparison algorithm are shown in Table 1.

[0088]

[0089] Table 1 In Table 1, the Causal Convolutional Generative Adversarial Network (CCGAN) proposed in this invention achieves better generation quality and numerical error reduction in photovoltaic scenarios compared with other comparative algorithms: On photovoltaic data, compared with the average level of the three algorithms, MAE and RMSE are reduced by about 48.7% and 47.1% respectively, and FID is reduced by about 81.4%, indicating that significant improvements are achieved in both point error and distribution fidelity.

[0090] Figure 5 For real-world photovoltaic power generation scenarios, from Figure 6 It can be seen that the causal convolutional generative adversarial network proposed in this invention can effectively reproduce the distribution and diversity of real-world scenes. The generation quality of other generative algorithms is significantly inferior to that of the method in this invention. Figure 7 In this case, the scene generated by the WGAN network does not fully reproduce the features. Figure 8 In practice, the scenes generated by GAN networks are relatively simple, concentrated in a narrow range, and cannot well reflect the richness of the original real scenes. Figure 9 In practice, the scenarios generated by VAE networks are not stable enough.

[0091] Example 2 The wind and solar scene generation method based on causal convolutional generative adversarial networks described in this invention, while retaining the core logical framework of Embodiment 1, is optimized and improved specifically for wind power scenarios. The specific implementation is as follows: S1.2 Data cleaning of wind power output data. Missing value repair adaptation: Given the strong correlation between wind speed and power output in wind power, for long-term missing data periods exceeding 12 hours, in addition to Newton interpolation, weighted interpolation is performed using historical complete data segments with the most similar wind speed change trends, combined with concurrent wind speed data. Furthermore, the outlier judgment range is set from 0 to 1.05 times the wind farm's installed capacity, adapting to differences in wind power installed capacity and power output characteristics.

[0092] S2.1 Adjustment strategy for the generator RF-Sched based on wind power output data Expansion rate scheduling sequence adaptation: For short-term drastic fluctuations in wind power output, such as rapid ramp-up within 1-2 hours and sudden drops in gusts, the expansion rate sequence is adjusted to d. l =1, 2, 4, 8|1, 2, 4, 8|8, 16, 32, extending the scanning duration of medium and low expansion rates to enhance the capture of short-term fluctuations.

[0093] S5.3 Results Analysis The evaluation results of the generation performance of the causal convolutional generative adversarial network CCGAN and the comparison algorithm are shown in Table 2.

[0094]

[0095] Table 2 In terms of wind power data, CCGAN only achieved minor optimizations in MAE and RMSE, but it has outstanding advantages in distribution quality: FID is reduced by an average of about 89.1% relative to the three baselines and by 53.8% relative to the strongest baseline, showing that it portrays complex wind power fluctuation distributions more realistically.

[0096] Figure 10 To represent real-world wind power output scenarios, among the typical wind power output scenario images generated by various algorithms, only... Figure 11 The method of this invention can accurately reproduce the distribution of real wind power scenarios, while other algorithms such as WGAN... Figure 12 GANs Figure 13 The generated scene does not fully reproduce the features, such as the VAE network. Figure 14 The generated scenarios are too simplistic and lack typicality.

Claims

1. A method for generating landscape and solar power output scenes based on causal convolutional generative adversarial networks, characterized in that, Includes the following steps: S1. Collect historical solar power output data and perform data preprocessing, the data preprocessing including dividing historical solar power output data by day, cleaning the solar power output data and constructing a noisy dataset; S2. Construct a generator network architecture, which consists of gated residual blocks, linear layers and activation functions. An RF-Sched receptive field scheduling strategy is introduced into the causal convolutional layer of the generator to capture the long-short-term dependence and periodicity of the wind and solar power output sequence. S3. Design a discriminator network architecture. The discriminator is constructed using causal convolutional layers and is used to output scalar scores to measure the authenticity of the input sample's real wind and solar power output data or the wind and solar power output data generated by the generator. S4. Alternately train the generator and discriminator: First, train the discriminator multiple times according to the set training ratio, then train the generator once, and repeat the training until the model converges. Record the optimal parameters of the generator and discriminator. S5. Load the optimal generator parameters, set the number of scenes to be generated, sample noise input to the generator to obtain the output sequence, obtain typical wind and solar power output scenes through inverse normalization and physical post-processing, and evaluate the scene quality through preset evaluation indicators.

2. The method for generating wind and solar power output scenes based on causal convolutional generative adversarial networks according to claim 1, characterized in that, The specific operation of dividing the historical wind and solar power output data by day in step S1 is as follows: using the daily periodicity of wind and solar power output, the original historical wind and solar power output data is divided into 24-hour segments to obtain N wind power daily output data and N photovoltaic daily output data. Each data segment contains data points of 24 time steps, which are combined to form a wind power output data matrix or a photovoltaic output data matrix of size N×24.

3. The method for generating wind and solar power output scenes based on causal convolutional generative adversarial networks according to claim 1, characterized in that, The process of cleaning the wind and solar power output data in step S1 includes marking missing points in the dataset, using Newton's interpolation method to fill in the missing data, detecting data points in the dataset with values ​​less than 0, and correcting them to 0.

4. The method for generating wind and solar power output scenes based on causal convolutional generative adversarial networks according to claim 1, characterized in that, The specific method for constructing the noise dataset in step S1 is to construct an N×24 noise dataset with the same size as the wind power output data matrix or the photovoltaic output data matrix. Each data point in the noise dataset follows a standard normal distribution and is used to provide input for the trained generator to generate wind and solar power output scenarios.

5. The method for generating wind and solar power output scenes based on causal convolutional generative adversarial networks according to claim 1, characterized in that, The core design of the RF-Sched receptive field scheduling strategy in step S2 is: to address the problem of sampling hole effect caused by fixed exponential expansion, an expansion rate scheduling logic is adopted that first identifies the overall features, then performs detailed scanning, and finally captures the long-term features. The segmented expansion rate sequence achieves full feature coverage of the wind and solar power output sequence.

6. The method for generating wind and solar power output scenes based on causal convolutional generative adversarial networks according to claim 1, characterized in that, In step S2, each gated residual block consists of two layers of causal convolution, gated linear units, layer normalization, and a temporary de-extension layer.

7. The method for generating wind and solar power output scenes based on causal convolutional generative adversarial networks according to claim 1, characterized in that, The expressions for the single-layer causal convolutional layer in steps S2 and S3 are as follows: ； in, The output of the causal convolutional layer. Here, k represents the kernel weights, k is the kernel length, and d represents the kernel weights. l The expansion rate of layer l is scheduled by the RF-Sched policy. The input sequence contains historical data at the current time step t and the kernel offset k.

8. The method for generating wind and solar power output scenes based on causal convolutional generative adversarial networks according to claim 1, characterized in that, The expression for the construction process of the linear layer and activation function in step S2 is as follows: ； Where Y is the output matrix of the linear layer, X is the input matrix, W is the weight matrix, and b is the bias vector. f The output matrix is ​​the activation function, and LeakyReLU is the selected activation function, which is expressed as max(αY, Y), where α is the negative slope, usually taken as 0.

01. This operation maps all negative numbers to a multiple of α, while positive numbers remain unchanged.

9. A method for generating wind and solar power output scenes based on causal convolutional generative adversarial networks according to claim 1, characterized in that, The loss function expression for the generator in step S2 is as follows: ； Where D(G(z)) is the discriminator's score for the generated data. The training objective of the generator is to minimize this loss function, that is, to make the discriminator's score for the generated data as high as possible, thereby improving the similarity between the generated scene and the real scene.

10. A method for generating wind and solar power output scenes based on causal convolutional generative adversarial networks according to claim 1, characterized in that, The loss function expression for the discriminator in step S3 is as follows: ； Where D(x) and D(G(z)) are the outputs of the discriminator on the real data x and the generated data G(z), respectively. These are the weights of the gradient penalty term, ensuring that the gradient of the discriminator does not become excessively large during training. These are interpolated samples, generated by interpolating between the real sample x and the generated sample G(z). A gradient penalty is calculated, and this penalty term is the third term in the discriminator's loss function. It is used to ensure the discriminator's Lipschitz continuity and prevent instability during training. For the interpolation sample, the expression is: ； in, The gradient penalty term ensures that the gradient changes of the discriminator are smooth.

Images