Wind power daily scenario generation method based on tensor self-organizing mapping neural network

Abstract

The application discloses a wind power daily scene generation method based on a tensor self-organizing mapping neural network, and comprises the following steps: obtaining a daily scene data set of multiple wind farms; clustering the daily scene data set by using a tensor self-organizing mapping neural network; constructing a variational autoencoder for each cluster respectively to extract hidden features from the daily scene data; performing random simulation and sampling on the daily scene data of each cluster by using the hidden features to obtain a daily scene hidden variable data set; decoding and reconstructing the daily scene hidden variable data set to obtain reconstructed daily scene data of each cluster; and aggregating the reconstructed daily scene data of each cluster to obtain a reconstructed daily scene data set. After clustering the daily scene samples, the application respectively reduces the dimension and reconstructs the samples of each cluster, improves the accuracy of the reconstructed daily scene data, significantly reduces the spatial and temporal correlation error of the reconstruction by considering the time and space correlation of the scene data, and strengthens the feature expression ability of the generated scene.

Description

Technical Field

[0001] This invention belongs to the field of renewable energy power generation and integrated consumption, specifically involving a method for generating daily wind power scenarios based on tensor self-organizing mapping neural networks. Background Technology

[0002] Scenario analysis methods, by constructing deterministic scenarios to analyze power system uncertainties, have become an effective approach to solving the planning and optimized operation problems of power systems containing renewable energy. Wind power output scenario generation typically employs statistical methods for modeling, assuming wind power output uncertainty as a statistical model. Historical data is used to establish a probabilistic model that conforms to the distribution patterns of wind power output, and sampling methods are combined to randomly generate output scenario samples. However, the difficulty of traditional statistical models lies in how to establish a suitable probabilistic model for wind power output uncertainty, and they are mostly used to describe the uncertainty of single-feature time-series scenarios, and are suitable for generating total wind power output scenarios after aggregation at a specific location or system. Meanwhile, neural networks can, to some extent, solve the problem of probabilistic distribution modeling difficulties in describing wind power output uncertainty, but traditional supervised learning models struggle to fit the probability distribution and require a large amount of training data. Furthermore, the difficulty in generating scenarios involving multiple wind farms lies in the complex temporal-spatial correlations involved. The rapid development of unsupervised learning generative models promises to solve these problems. Generative models can learn the probability distribution of training data and generate new samples that conform to the data's probability distribution. Currently, deep learning methods applied to power system scenario generation mainly fall into two categories: generative adversarial networks (GANs) and variational autoencoders (VAEs). These deep learning techniques have achieved some success in generating renewable energy day-ahead scenarios. However, existing methods all uniformly transform the spatiotemporal power in the sample scenario into a one-dimensional vector before modeling, resulting in models that cannot accurately reflect the true distribution of the original power in the temporal and spatial dimensions, thus affecting the accuracy of scenario generation.

[0003] Therefore, this study investigates methods for generating daily spatiotemporal power scenes of multiple wind farms. Based on the characteristics of wind power data itself, it aims to achieve random generation of daily spatiotemporal power scenes of multiple wind farms. A method combining tensor self-organizing feature map (SOM) clustering and VAE variational autoencoder dimensionality reduction is proposed to achieve the generation of daily spatiotemporal power scenes of multiple wind farms. Summary of the Invention

[0004] The purpose of this invention is to address the aforementioned problems by considering the diversity of spatiotemporal power correlations and accurately reflecting the spatiotemporal distribution of power in the original scene. It provides a scene generation method combining a self-organizing map neural network based on tensor distance and a variational autoencoder neural network, enabling the random generation of daily spatiotemporal power scenes from multiple wind farms. This method uses the second-order spatiotemporal tensor distance of the daily power scene as a basis to cluster historical spatiotemporal power daily scene samples, ensuring that daily scene samples within the same cluster have similar spatiotemporal correlations. A VAE encoding / decoding network is constructed for each cluster of daily scene sets to achieve bidirectional transformation between the high-dimensional actual spatiotemporal power within the daily scene and the low-dimensional latent features following an independent normal distribution. By randomly sampling the latent features using an independent multidimensional normal distribution, the generated samples of each cluster of daily scenes are decoded and aggregated proportionally to obtain a new set of scenes with similar probability distributions and spatiotemporal correlation patterns to the original data, thus improving the effectiveness of scene generation.

[0005] The technical solution of this invention is a method for generating daily wind power scenes based on tensor self-organizing map neural networks. First, tensor self-organizing map neural networks are used to cluster daily spatiotemporal power scenes of multiple wind farms. The dimensionality of each cluster is reduced, and the implicit features of the dimensionality reduction are independently sampled. Then, the sampled features are decoded to obtain a new set of daily spatiotemporal power scenes for multiple wind farms. The method for generating daily spatiotemporal power scenes of multiple wind farms includes the following steps:

[0006] Step 1: Collect daily scene data of wind power output from multiple wind farms to obtain a daily scene dataset;

[0007] Step 2: Use a tensor self-organizing map neural network to cluster the daily scene dataset to obtain daily scene clusters;

[0008] Step 2.1: Input historical spatiotemporal power sample data from multiple wind farms Where x i This represents the scene data for the i-th day, where N represents the number of historical sample days, m is the number of wind farms, and h is the number of time points within the day. Represents an m×h dimensional real space;

[0009] Step 2.2: Assign weights W to each node in the output layer of the Tensor Self-Organizing Map Neural Network. j Initial values ​​are randomly assigned to j = 1, 2, ... J;

[0010] Step 2.3: Analyze the daily spatiotemporal power data of multiple wind farms x i Calculate with x i The connection weight matrix with the shortest distance yields the winning unit j of the tensor self-organizing map neural network. * The calculation formula is as follows:

[0011]

[0012]

[0013] In the formula j * Represents the winning unit; J represents the number of neurons in the input layer of the tensor self-organizing map neural network; ||·|| represents the distance function; g lm Here, G represents the element position metric coefficient, and G is the metric matrix related to the distance between element positions.

[0014] Step 2.4: Define the neighborhood of the winning unit For cells within the neighboring region, adjust their weights to direct them towards x. i The iterative calculation formula for the convergence and weight adjustment is as follows:

[0015]

[0016] In the formula w ij (t), w ij (t+1) represent the weights of neurons i to j at times t and t+1, respectively; α(t,N) represents the weights of the i-th neuron and the winning neuron j in the neighborhood. * A function of the topological distance D between them;

[0017] Step 2.5: Repeat steps 2.3 and 2.4 until the training termination condition α(t) ≤ α is met. min Training stops when α(t) is reached, where α(t) represents the learning rate. min This represents the minimum learning rate and outputs the cluster. Where X k This represents the k-th family of wind power datasets. Indicates the nth k Daily scene data, n k denoted as the number of samples of wind power data in the k-th cluster, where K is the number of clusters.

[0018] Step 3: Construct variational autoencoders for each cluster and use the encoders of the variational autoencoders to extract latent features from the daily scene data;

[0019] Step 4: Use the latent features extracted in Step 3 to perform random simulation and sampling of the daily scene data for each cluster to obtain the daily scene latent variable dataset for each cluster.

[0020] Step 5: Use the decoder of the variational autoencoder to decode and reconstruct the daily scene latent variable dataset to obtain the reconstructed daily scene data for each cluster; aggregate the reconstructed daily scene data for each cluster to obtain the reconstructed daily scene dataset.

[0021] Furthermore, the calculation process of tensor distance in a tensor distance self-organizing map neural network specifically includes:

[0022] Let tensor N>1, x is its vectorized form, for element 1≤i j ≤I j ,1≤j≤N,x l ,x m These are the l-th and m-th elements of its vectorized form, respectively, where Let tensor If it is a tensor in the same space as χ, then and The tensor distance between them can be calculated by equation (4):

[0023]

[0024] In the formula d TD Represents tensor distance, g lm Let G be the element position metric coefficient, and G represent the metric matrix related to the distance between element positions, reflecting the intrinsic relationship between different coordinates of multi-level data. Element position metric coefficient g lm The calculation formula is as follows:

[0025]

[0026] In the formula, σ is the regularization parameter, ||p l -p m ||2 represents x l and x m In the original tensor space, the coordinates of the position (i1, i2, ..., i) are... N )and Distance between:

[0027]

[0028] When G is the identity matrix I, the calculated tensor distance is equal to the Euclidean distance.

[0029] Furthermore, step 3 includes the following sub-steps:

[0030] Step 3.1: Based on the scenario sample set X of each cluster k Unsupervised training of variational autoencoders (VAEs) corresponding to each cluster. k k = 1, 2, ..., K;

[0031] Step 3.2: Utilize a Variational Autoencoder (VAE) k The encoder extracts the sample set X k The implicit features are used to obtain the sample set X. kThe implicit features μ and σ, where μ represents the sample set X k The mean of the sample set X, σ represents the mean of the sample set X. k The variance.

[0032] Furthermore, step 4 includes the following sub-steps:

[0033] Step 4.1: Using sample set X k The implicit features μ and σ are randomly simulated to obtain the sample set X. k ={x1,x2,...,x nk The hidden variable z of} k =μ(x k )+εσ(x k ), μ(x k ), σ(x) k ) represent the sample set X generated by the encoder. k ={x1,x2,...,x nk The mean and variance of}, where ε is a random number following a standard normal distribution N(0,1).

[0034] Step 4.2: For the latent variable z i Perform independent sampling to obtain n′ k =Mn k / N latent variable samples that follow a standard normal distribution, where M represents the total number of scenarios, N represents the number of historical sample days, and n k This represents the number of samples in the k-th cluster;

[0035] Step 4.3: Combine the extracted latent variable samples into an r-dimensional latent feature vector sample set. z′ j Let represent the j-th hidden feature vector.

[0036] Furthermore, step 5 includes the following sub-steps:

[0037] Step 5.1: Sample the latent feature vectors Z′ k Input the decoder of the corresponding cluster variational autoencoder, and use the decoder to convert Z′ k Reconstruction yields the daily scene sample set generated by the reconstruction of the k-th cluster. 1≤j≤n′ k , This represents the daily scene data generated by the j-th reconstruction;

[0038] Step 5.2: Aggregate the daily scene sample set generated by the K-cluster to obtain M new random simulation scenarios of spatiotemporal power of multiple wind farms.

[0039] Compared with the prior art, the beneficial effects of the present invention include:

[0040] 1) The scene generation method of the present invention performs tensor distance self-organizing mapping neural network clustering on multi-wind farm data, which realizes the accurate reflection of the real distribution of wind power in time or space dimension, and can obtain clustering results with smaller inter-class similarity and more diverse information, thereby improving the diversity and accuracy of clustering results;

[0041] 2) By constructing a generative network model for each cluster, this invention can effectively improve the training effect of the generative network and reduce the reconstruction error, thereby improving the accuracy of the generative network and the diversity of the generated scenarios.

[0042] 3) The scene generation method of the present invention significantly reduces the spatial and temporal correlation error of wind power, effectively improves the accuracy of probability distribution features, and enhances the feature expression capability of the scene generation method. Attached Figure Description

[0043] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0044] Figure 1 This is the SOM neural network model in an embodiment of the present invention.

[0045] Figure 2 A comparison of the characteristics of SOM based on tensor distance and Euclidean distance.

[0046] Figure 3 This is a structural diagram of the VAE model.

[0047] Figure 4 This is a flowchart illustrating the scene generation process of the multi-wind farm scene generation method according to an embodiment of the present invention.

[0048] Figure 5 This is a comparison chart of the sample proportions of each cluster using the Euclidean distance SOM and tensor distance SOM clustering methods according to embodiments of the present invention.

[0049] Figure 6a This is a cluster center diagram of the tensor distance SOM clustering results in an embodiment of the present invention.

[0050] Figure 6b This is a cluster center diagram of the Euclidean distance SOM clustering results in an embodiment of the present invention.

[0051] Figure 7 This is a comparison chart of the changes in loss values ​​during the training process of clustered and non-clustered VAEs according to an embodiment of the present invention.

[0052] Figure 8 This is a comparison chart of MAPE errors between clustered and non-clustered generated scenarios according to an embodiment of the present invention.

[0053] Figure 9aThis is a spatial correlation matrix diagram of historical scene data in an embodiment of the present invention.

[0054] Figure 9b The absolute error map of the spatial correlation matrix of the scene generated by the day scene generation method of the present invention.

[0055] Figure 9c The absolute error plot of the spatial correlation matrix of the scene is generated for the daily scene generation method that does not apply VAE to clustering.

[0056] Figure 9d An absolute error map of the spatial correlation matrix of a daytime scene generated by the Euclidean distance (SOM) combined with VAE method.

[0057] Figure 10a This is a time correlation matrix diagram of historical scene data in an embodiment of the present invention.

[0058] Figure 10b The absolute error diagram of the time correlation coefficient of the scene generated by the day scene generation method of the present invention is shown.

[0059] Figure 10c Generate the absolute error map of the temporal correlation coefficient of the scene for the daily scene generation method that does not apply VAE to clustering.

[0060] Figure 10d The absolute error map of the temporal correlation coefficient of the daytime scene generation method using Euclidean distance SOM combined with VAE is generated.

[0061] Figure 11 The cumulative probability distribution of scene average power is generated using three methods in the embodiments of the present invention. Detailed Implementation

[0062] like Figure 1 As shown, the wind power daily scene generation method based on Tensor Self-Organizing Map Neural Network (SOM) employs a SOM neural network. It simulates the self-organizing mapping function of the brain's nervous system, using unsupervised competitive learning to extract important features or inherent patterns from a set of data for classification. The SOM network can map arbitrarily high-dimensional inputs to a low-dimensional space, and makes certain similar properties within the input data appear as geometrically adjacent feature maps, preserving the invariance of the data topology.

[0063] like Figure 2 As shown, the tensor distance and Euclidean distance in the wind power daily scene generation method based on tensor self-organizing map neural network preserve the spatiotemporal characteristics between scene data and can reflect the spatiotemporal positional relationship of data within the scene through weights.

[0064] like Figure 3As shown, the variational autoencoder network structure in the wind power daily scene generation method based on tensor self-organizing map neural network consists of two parts: an encoder and a decoder. The encoder generates the latent variable z corresponding to the input sample data, denoted as the recognition model q. φ (z|x), while the decoder generates the observed data from a series of latent variables z. Let p be the generative model θ (x|z). In this embodiment, both the decoder and encoder employ a feedforward neural network containing multiple hidden layers.

[0065] To address the issue that the distribution of the latent variable z is not directly observable, VAE introduces a recognition model q into the inference network. φ (z|x) is used to replace the uncertain true posterior distribution p. θ (x|z), and assume the recognition model q φ (z|x) is a known distribution, usually a standard normal distribution, which helps identify the model q. φ (z|x) can then be used as the inference network part of the VAE, with the conditional distribution p θ (x|z) is used as the generator network part. To make the recognition model q... φ (z|x) and the true posterior distribution p θ Since (x|z) are approximately equal, VAE uses KL divergence to measure the similarity between them. The complete formula for calculating the loss function L(θ,φ,x) of VAE is as follows:

[0066] L(θ,φ,x)=KL((q φ (z|x)),(p θ (z|x)))

[0067] -E qφ(z|x) [log(p θ (x|z)] (7)

[0068] In the formula, the first term KL((q) φ (z|x)),(p θ (z|x))) represents the probability distribution of the latent variable z and the prior distribution q φ The similarity of (z|x) is such that the more similar their probability distributions are, the smaller the KL divergence, and the smaller the latter term E. qφ(z|x) [log(p θ [x|z)] represents the error between the reconstructed sample and the original sample. That is, in the training of VAE network, the training objective is to minimize the reconstruction error between the reconstructed sample and the original sample, and to make the probability distribution of the latent variable z as close as possible to the prior distribution, such as the standard normal N(0,1).

[0069] like Figure 4 As shown, the method for generating daily wind power scenarios based on tensor self-organizing map neural networks specifically includes the following steps:

[0070] Step 1: Collect daily scene data of wind power output from multiple wind farms to obtain a daily scene dataset;

[0071] Step 2: Use a tensor self-organizing map neural network to cluster the daily scene dataset to obtain daily scene clusters;

[0072] Step 2.1: Input historical spatiotemporal power sample data from multiple wind farms Where, x i This represents the scene data for the i-th day, where N is the number of historical sample days, m is the number of wind farms, and h is the number of time points within the day.

[0073] Step 2.2: Assign weights W to each node in the output layer of the Tensor Self-Organizing Map Neural Network. j (j=1,2,…J) are randomly assigned initial values.

[0074] Step 2.3: Analyze the daily spatiotemporal power data of multiple wind farms x i Calculate x i With W j The connection weight matrix with the shortest distance (j = 1, 2, ..., J) is calculated as follows:

[0075]

[0076]

[0077] In the formula j * The winning unit; J is the number of neurons in the output layer; ||·|| is the distance function; g lm Here, G represents the element position metric coefficient, and G is the metric matrix related to the distance between element positions.

[0078] Step 2.4: Define the neighborhood of the winning unit For cells within the neighboring region, adjust their weights to direct them towards x. i The iterative calculation formula for the weight adjustment is as follows:

[0079]

[0080] In the formula, w ij (t), w ij (t+1) represent the weights of neurons i to j at times t and t+1, respectively; α(t,N) represents the weights of the i-th neuron and the winning neuron j in the neighborhood. * A function of the topological distance D between them;

[0081] Step 2.5: Repeat steps 2.3 and 2.4 until the training termination condition α(t) ≤ α is met. min Training stops when α min This represents the minimum learning rate and outputs the cluster. Where n k Let K be the number of samples of wind power data in the k-th cluster, and K be the number of clusters.

[0082] The calculation process of tensor distance specifically includes:

[0083] Let tensor N>1, x is its vectorized form, for element 1≤i j ≤I j ,1≤j≤N,x l ,x m These are the l-th and m-th elements of its vectorized form, respectively, where Let tensor Is with Tensors in the same space, then and The tensor distance between them can be calculated by equation (4):

[0084]

[0085] In the formula d TD Represents tensor distance, g lm G is the element position metric coefficient, and G is the metric matrix related to the distance between element positions, reflecting the intrinsic relationship between different coordinates between multi-level data. lm The definition is as follows:

[0086]

[0087] In the formula, σ is the regularization parameter, ||p l -p m ||2 is x l and x m In the original tensor space, the coordinates of the position (i1, i2, ..., i) are... N )and Distance between:

[0088]

[0089] When G is the identity matrix I, the calculated tensor distance is equal to the Euclidean distance.

[0090] Step 3: Construct variational autoencoders for each cluster and use the encoders of the variational autoencoders to extract latent features from the daily scene data;

[0091] Step 3.1: Based on the scenario sample set X of each clusterk Unsupervised training of VAEs corresponding to each cluster k Network k = 1, 2, ..., K;

[0092] Step 3.2: Utilize a Variational Autoencoder (VAE) k The encoder extracts the sample set X k The implicit features are used to obtain the sample set X. k The implicit features μ and σ, where μ represents the sample set X k The mean of the sample set X, σ represents the mean of the sample set X. k The variance.

[0093] Step 4: Randomly simulate and sample the daily scene data of each cluster using the extracted latent features to obtain the daily scene latent variable dataset for each cluster.

[0094] Step 4.1: Using sample set X k The implicit features μ and σ are randomly simulated to obtain the sample set. The hidden variable z k =μ(x k )+εσ(x k ), μ(x k ), σ(x) k ) represent the sample sets generated by the encoder. The mean and variance of , where ε is a random number following a standard normal distribution N(0,1).

[0095] Step 4.2: Independently sample the latent features to obtain n′ k =Mn k / N latent variable samples that follow a standard normal distribution, where M is the total number of scenarios, N represents the number of historical sample days, and n k This represents the number of samples in the k-th cluster;

[0096] Step 4.3: Combine the extracted latent variable samples into an r-dimensional latent feature vector sample set. z′ j Let represent the j-th hidden feature vector.

[0097] Step 5: Use the decoder of the variation autoencoder to decode and reconstruct the daily scene latent variable dataset to obtain the reconstructed daily scene data for each cluster; aggregate the reconstructed daily scene data for each cluster to obtain the reconstructed daily scene dataset.

[0098] Step 5.1: Sample the latent feature vectors Z′ k Input the decoder of the corresponding cluster variational autoencoder, and use the decoder to convert Z′ k Reconstruction yields the daily scene sample set generated by the reconstruction of the k-th cluster. This represents the daily scene data generated by the j-th reconstruction;

[0099] Step 5.2: Aggregate the daily scene sample set generated by the K-cluster to obtain M new random simulation scenarios of spatiotemporal power of multiple wind farms.

[0100] Figure 5 The figure shows the proportion of samples in each cluster in the Euclidean distance SOM neural network clustering results and the tensor distance SOM neural network clustering results.

[0101] Figure 6a , 6b The images shown are the daily scene cloud maps of the cluster centers of each cluster sample in the Euclidean distance SOM neural network clustering results and the tensor distance SOM neural network clustering results, respectively.

[0102] Figure 7 The image shows a comparison of the network loss values ​​between the method of this invention and a scene generation method that does not use clustering and employs VAE. Figure 7 This indicates that clustering historical samples of daily scenes with similar spatiotemporal correlations and then using VAE for dimensionality reduction helps to extract the essential features of daily scenes with different spatiotemporal correlation patterns, thereby improving the accuracy of the VAE generation network and the diversity of generated scenes.

[0103] Figure 8 The figure shows the mean absolute percentage error (MAPE) between the reconstructed scene generated by the VAE network and the original scene. The calculation formula is as follows:

[0104]

[0105] In the formula P j Let x be the rated power of the j-th wind farm. j,h,i This represents the actual power of the j-th wind farm at time h in the i-th day scene of the original scene dataset. This represents the power of the j-th wind farm in the i-th day scene of the generated scenario at the h-th time point.

[0106] Depend on Figure 8 It is evident that the MAPE of the generated data from the four clusters in this invention is generally lower than that of the direct VAE result without clustering. The reconstruction error MAPE value, calculated by weighting the average absolute percentage error of the four clusters according to their sample proportions, is 5.32%, which is 17.37% lower than the MAPE value of the direct VAE method without clustering.

[0107] Figure 9aThe figure shows the absolute error of the spatial correlation matrix of historical power day scene data of multiple wind farms in the embodiment. Figure 9b , 9c 9d and 9d are the absolute errors of the spatial correlation coefficient matrix of the generated scenes in the daily scene generation method of Zhang's distance SOM combined with VAE, the daily scene generation method of VAE without clustering, and the daily scene generation method of Euclidean distance SOM combined with VAE, respectively.

[0108] Figure 10a The figure shown is the absolute error of the time correlation coefficient of historical power day scene data of multiple wind farms in the embodiment. Figure 10b , 10c 10d and 10d represent the absolute errors of the time correlation coefficients of the scenes generated by the Zhang's distance SOM combined with VAE daily scene generation method, the non-clustering VAE daily scene generation method, and the Euclidean distance SOM combined with VAE daily scene generation method of the present invention, respectively. Figures 9a-9d as well as Figures 10a-10d As can be seen from the image, the darkest color indicates the smallest error in the correlation coefficient between the spatial and temporal correlation of the scene generated by the method of the present invention and the original scene. This means that the scene generated by the method of the present invention can more accurately reflect the spatiotemporal correlation pattern of the original scene.

[0109] Figure 11 The image shows the empirical cumulative probability distribution of all average power after calculating the average power of 18 wind farms at various times in the original scene and the scenes generated by the three methods. Figure 11 A comparison of the cumulative probability distributions of the three methods shows that all three methods perform well in terms of probability distribution, achieving the effect of probabilistic modeling in traditional scene generation methods in an unsupervised manner. However, the scene generated by the method of this invention is closest to the original scene, indicating that the method of this invention can more accurately capture the probability distribution patterns of historical wind power data.

Claims

1. A method for generating daily wind power scenarios based on tensor self-organizing map neural networks, characterized in that, Includes the following steps: Step 1: Collect daily scene data of wind power output from multiple wind farms to obtain a daily scene dataset; Step 2: Use a tensor self-organizing map neural network to cluster the daily scene dataset to obtain daily scene clusters; Step 2.1: Input historical spatiotemporal power sample data from multiple wind farms Where x i This represents the scene data for the i-th day, where N represents the number of historical sample days, m is the number of wind farms, and h is the number of time points within the day. Represents an m×h dimensional real space; Step 2.2: Assign weights W to each node in the output layer of the Tensor Self-Organizing Map Neural Network. j Initial values ​​are randomly assigned to j = 1, 2, ... J; Step 2.3: Analyze the daily spatiotemporal power data of multiple wind farms x i Calculate with x i The connection weight matrix with the shortest distance yields the winning unit j of the tensor self-organizing map neural network. * ; Step 2.4: Define the neighborhood of the winning unit For cells within the neighboring region, adjust their weights to direct them towards x. i Approach; Step 2.5: Repeat steps 2.3 and 2.4 until the training termination condition α(t) ≤ α is met. min Training stops when α(t) is reached, where α(t) represents the learning rate. min This represents the minimum learning rate and outputs the cluster. Where X k This represents the k-th family of wind power datasets. Indicates the nth k Daily scene data, n k denoted as the number of samples of wind power data in the k-th cluster, where K is the number of clusters. Step 3: Construct variational autoencoders for each cluster and use the encoders of the variational autoencoders to extract latent features from the daily scene data; Step 4: Use the latent features extracted in Step 3 to perform random simulation and sampling of the daily scene data for each cluster to obtain the daily scene latent variable dataset for each cluster. Step 5: Use the decoder of the variational autoencoder to decode and reconstruct the daily scene latent variable dataset to obtain the reconstructed daily scene data for each cluster; aggregate the reconstructed daily scene data for each cluster to obtain the reconstructed daily scene dataset.

2. The method for generating daily wind power scenarios according to claim 1, characterized in that, In step 2.4, the iterative calculation formula for weight adjustment is as follows: w ij (t+1)=w ij (t)+α(t,D)[x i (t)-w ij (t)] In the formula w ij (t), w ij (t+1) represent the weights of neurons i to j at times t and t+1, respectively; α(t,D) represents the weights of the i-th neuron and the winning neuron j in the neighborhood. * A function of the topological distance D between them.

3. The method for generating daily wind power scenarios according to claim 2, characterized in that, Step 3 includes the following sub-steps: Step 3.1: Based on the scenario sample set X of each cluster k Unsupervised training of variational autoencoders (VAEs) corresponding to each cluster. k k = 1, 2, ..., K; Step 3.2: Utilize a Variational Autoencoder (VAE) k The encoder extracts the sample set X k The implicit features are used to obtain the sample set X. k The implicit features μ and σ, where μ represents the sample set X k The mean of the sample set X, σ represents the mean of the sample set X. k The variance.

4. The method for generating daily wind power scenarios according to claim 3, characterized in that, Step 4 includes the following sub-steps: Step 4.1: Using sample set X k The implicit features μ and σ are randomly simulated to obtain the sample set X. k The implicit variable z; Step 4.2: Perform independent sampling on the latent variable z to obtain n′ k =Mn k / N latent variable samples that follow a standard normal distribution, where M represents the total number of scenarios, N represents the number of historical sample days, and n k This represents the number of samples in the wind power data of the k-th cluster; Step 4.3: Combine the extracted latent variable samples into an r-dimensional latent feature vector sample set. z′ j Let represent the j-th hidden feature vector.

5. The method for generating daily wind power scenarios according to claim 4, characterized in that, Step 5 includes the following sub-steps: Step 5.1: Sample the latent feature vectors Z′ k Input the decoder of the corresponding cluster variational autoencoder, and use the decoder to convert Z′ k Reconstruction yields the daily scene sample set generated by the reconstruction of the k-th cluster. This represents the daily scene data generated by the j-th reconstruction; Step 5.2: Aggregate the daily scene sample set generated by the K-cluster to obtain M new random simulation scenarios of spatiotemporal power of multiple wind farms.

Images