Multi-region wind power short-term scene generation method and system

Abstract

The invention discloses a multi-region wind power short-term scene generation method and system, and the method comprises the steps: collecting meteorological data and historical wind power output data of a multi-region wind power plant, and carrying out the cleaning and standardization processing; carrying out feature embedding on the meteorological features by using a pre-trained coding network so as to evaluate a dynamic adjacency matrix between regions and capture spatial correlation; calculating a dynamic adjacency matrix based on the Spearman correlation coefficient, and reflecting the dynamic correlation of wind power output of different regions; mapping a high-dimensional wind power output curve to a low-dimensional hidden space feature vector by using a variational auto-encoder, and reconstructing a high-dimensional output sequence through a decoder; constructing a conditional hidden space diffusion model, gradually adding noise to a hidden space sample to normal distribution, and designing a denoising network to gradually remove noise; and reconstructing the low-dimensional hidden space scene output by the denoising network to generate a high-dimensional wind power output scene. According to the method, the accuracy of scene generation is improved, and effective support is provided for economic dispatching and risk assessment of a power system.

Description

Technical Field

[0001] The present invention relates to the field of renewable energy and smart grid technology, and in particular to a method and system for generating short-term wind power scenarios in multiple regions based on a conditional latent space diffusion model and dynamic graph representation. Background Art

[0002] With the intensification of global climate change and the increase in demand for renewable energy, wind power has become an important part of coping with energy transformation. However, wind power output is affected by the chaos of the atmospheric system and has obvious randomness, volatility and intermittency. With the accelerated grid connection of wind farms, the uncertainty of short-term wind power output has brought huge challenges to the power balance analysis, economic dispatch and safe and stable operation of the power system. On the other hand, affected by atmospheric movement, the uncertainty of short-term wind power output in multiple regions has a certain spatial correlation, and the mismatch between its output prediction results is easy to cause the difficulty of consumption, which restricts the overall improvement of wind power utilization. In order to cope with the uncertainty of wind power output, the random programming method based on scenario generation has been widely used in problems such as power grid economic dispatch. The higher the accuracy of scenario generation, the closer the solution of the corresponding random optimization problem is to the actual optimal value. Therefore, how to accurately construct the short-term output scenario of wind power in multiple regions is of great significance for solving the problem of rolling economic dispatching in the day and day under large-scale wind power access, improving the level of wind power consumption and ensuring the safe and stable operation of the power system.

[0003] At present, most of the research on scenario generation adopts the method of decoupled modeling. Although it has good interpretability, the fitting process strictly relies on the regression of the assumed function parameters. Due to the limitations of the function form, it is difficult to fully characterize the complex temporal, spatial and meteorological correlations of wind power output, which often leads to large deviations in the scenario generation results. In addition, in order to model the spatial correlation of wind power output in multiple regions, existing studies usually model the static spatial correlation of wind power output, which can achieve better results when the research object is stable for a long time. However, the spatial correlation of wind power output is not only related to the distance between sites, but also closely related to real-time environmental conditions such as wind speed, wind direction, and temperature. The short-term fluctuation of wind power makes its spatial correlation in dynamic change. It is difficult to accurately reflect the correlation of short-term changes in wind power output in multiple regions under the influence of weather with constant spatial correlation.

[0004] Definitions of Abbreviations and Key Terms:

[0005] Dynamic Graph Neural Network (DGNN): Dynamic Graph Neural Network is a method to extend traditional Graph Neural Network (GNN) to process graph data that changes over time. DGNN can capture dynamic changes in graph structure, such as the addition and deletion of nodes or changes in edges.

[0006] Representation Learning for Dynamic Graph: Representation Learning for Dynamic Graph refers to generating effective representations for dynamic graphs through learning algorithms. This representation can capture the structure of the graph, node features, and its dynamic characteristics over time.

[0007] Meteorological forecast data (Numerical Weather Prediction, NWP): Meteorological forecast data is generated by numerical weather forecast models and is used to predict future weather conditions. NWP uses mathematical models and computers to simulate atmospheric physical processes and generates weather forecasts for future time periods by analyzing current meteorological observation data.

[0008] Generative Adversarial Networks (GAN): It consists of two main parts: the generator and the discriminator. The generator is responsible for generating data samples from random noise, while the discriminator tries to distinguish these generated samples from real samples. These two models are trained simultaneously through a "zero-sum game" process.

[0009] Variational Auto-Encoder (VAE): VAE consists of an encoder and a decoder. The encoder is responsible for converting the input data into a probability distribution (usually a Gaussian distribution), and the decoder samples from the distribution and tries to reconstruct the original input.

[0010] Diffusion Model (DM): The diffusion model is an emerging generative model that solves the posterior expression problem through an explicit diffusion process and has demonstrated excellent performance in various tasks. Summary of the invention

[0011] In view of the defects of the prior art, the present invention provides a method and system for generating a multi-regional wind power short-term scenario.

[0012] In order to achieve the above invention object, the technical solution adopted by the present invention is as follows:

[0013] A method for generating a short-term wind power scenario in multiple regions comprises the following steps:

[0014] S1: Collect meteorological data and historical wind power output data of wind farms in multiple regions. Clean and standardize the collected data to eliminate noise and outliers.

[0015] S2: Use the pre-trained encoding network to embed the meteorological features NWP of each region. The embedding results are used to evaluate the dynamic adjacency matrix between regions to capture the spatial correlation between regions.

[0016] S3: Based on the Spearman correlation coefficient of wind power output in each region, the dynamic adjacency matrix is ​​calculated in real time to reflect the dynamic correlation between wind power output in different regions.

[0017] S4: Using variational autoencoder VAE, the high-dimensional wind power output curve is mapped to a low-dimensional latent space feature vector. The encoder E is used to extract the trend characteristics of wind power output, and the decoder D is used to reconstruct the latent space features into a high-dimensional output sequence.

[0018] S5: Construct a conditional latent space diffusion model, including diffusion process and denoising process. Set the diffusion process to gradually add noise to the latent space samples to normal distribution. Design a denoising network, use conditional information as a covariate, gradually remove noise, and restore latent space samples.

[0019] S6: The denoising network is trained using a residual connection and spatiotemporal convolutional layer structure to improve the learning ability of spatiotemporal related features. The training goal is to minimize the KL divergence between the posterior distribution and the prior distribution to ensure the effectiveness of the denoising process.

[0020] S7: Reconstruct the low-dimensional latent space scenario output by the denoising network to generate a high-dimensional wind power output scenario. The generated scenario is used for economic dispatch and risk assessment of the power system.

[0021] Furthermore, for each region i, the input meteorological data is expressed as T represents the time step and C represents the feature dimension.

[0022] Use the pre-trained encoding network f e x i Embedding is performed to generate a high-dimensional representation of the features.

[0023] The encoding network is a convolutional neural network structure, including: input layer, intermediate layer and output layer;

[0024] Input layer: adjust the number of channels to accommodate subsequent convolutional layers;

[0025] Middle layer: Use multiple convolutional layers and increase the receptive field through residual connection structure to prevent network degradation;

[0026] Output layer: Use skip connections to combine the results of each convolutional layer to achieve the fusion of multi-time scale features and finally obtain the embedding result

[0027] Furthermore, the encoder network is trained using L1 loss for deterministic predictive regression:

[0028]

[0029] Among them, y i is the actual wind power output in region i, The result of embedding.

[0030] Furthermore, all regions are embedded to obtain a feature embedding set: And construct a dynamic adjacency matrix based on this

[0031] Furthermore, the sub-steps of S4 are as follows:

[0032] S41: High-dimensional wind power output sequence Enter encoder E.

[0033] Encoder E generates the mean vector and the standard deviation vector Thus, the probability distribution p of the latent space is defined θ (z|y).

[0034] S42: Sampling latent space variables from the above probability distribution

[0035] S43: Input the latent space variable z into the decoder D to generate the reconstructed wind power output sequence

[0036] S44: Calculate the VAE loss function, including KL divergence and reconstruction error:

[0037]

[0038] After introducing the weight coefficient λ, the updated loss function is:

[0039]

[0040] Where: ε(·) and are the mapping functions of the encoder and decoder respectively; q φ (z|y) represents the posterior distribution of the encoder output, which represents the distribution of the latent variable z after a given input y. θ (z|y) represents the reconstructed distribution generated by the decoder, which represents the probability of recovering the data from the latent variable z.

[0041] S45: Map the wind power sequence of each region through the encoder to obtain the latent space set Z0.

[0042] S46: Use latent space features for downstream tasks of the diffusion model to reconstruct the low-dimensional scene output by the denoising network into a high-dimensional wind power output scene to ensure accurate characterization of detailed fluctuation characteristics.

[0043] Further, the diffusion process is as follows:

[0044] The diffusion process gradually adds noise to the latent space sample Z0 to Z N , to achieve a gradual transition to a standard normal distribution, as follows:

[0045]

[0046] Where: β∈(0,1) represents the scaling factor of the nth step; ∈ is the noise sample, β n represents the scaling factor of the nth step, controlling the degree of noise addition, and ∈ represents the random noise added during the diffusion process.

[0047] The Markov chain process is expressed as:

[0048]

[0049] As n increases, the latent space samples tend to be standard normal distribution

[0050] Furthermore, the denoising process is expressed as:

[0051]

[0052] The denoising goal is to minimize the KL divergence between the posterior distribution and the prior distribution, which is expressed as:

[0053]

[0054] In the formula, l ELBO represents the lower bound of evidence in variational inference, represents the expectation of all possible Z0 and noise ∈ sampled from a standard normal distribution, ||∈-∈ θ (Z n ,n,C)||2 represents the reconstruction error.

[0055] Furthermore, the structural design of the denoising network aims to enhance dynamic spatial correlation, combining residual connections, skip connections and graph convolutions.

[0056] The graph convolutional layer is represented as:

[0057]

[0058] In the formula, H l-1 and H l are the input and output of the lth layer, W dThe parameters of the adaptive matrix, for degree matrix; W and b are the network learnable parameters of GCN; ⊙ represents the Hadamard product; σ is the activation function.

[0059] The present invention also discloses a multi-regional wind power short-term scenario generation system, which can be used to implement the above-mentioned multi-regional wind power short-term scenario generation method, specifically including:

[0060] Data collection module: collects meteorological data and historical wind power output data of wind farms in multiple regions, and performs data cleaning and standardization to eliminate noise and outliers;

[0061] Feature embedding module: Use a pre-trained encoding network to embed the meteorological characteristics (NWP) of each region, and use the embedding results to evaluate the dynamic adjacency matrix between regions to capture the spatial correlation between regions;

[0062] Dynamic adjacency matrix calculation module: Based on the Spearman correlation coefficient of wind power output in each region, the dynamic adjacency matrix is ​​calculated in real time to reflect the dynamic correlation between wind power output in different regions;

[0063] Latent space mapping module: Use variational autoencoder (VAE) to map the high-dimensional wind power output curve to a low-dimensional latent space feature vector, extract the trend characteristics of wind power output through the encoder, and use the decoder to reconstruct the high-dimensional output sequence;

[0064] Conditional latent space diffusion model module: Construct a conditional latent space diffusion model including diffusion process and denoising process, set the diffusion process, gradually add noise to the latent space samples to normal distribution, and design a denoising network with conditional information as covariate to gradually remove noise and restore latent space samples;

[0065] Denoising network training module: The denoising network is trained using the structure of residual connection and spatiotemporal convolution layer to improve the learning ability of spatiotemporal related features. The training goal is to minimize the KL divergence between the posterior distribution and the prior distribution.

[0066] Scenario generation module: Reconstruct the low-dimensional latent space scenario output by the denoising network to generate a high-dimensional wind power output scenario, and finally use the generated scenario for the economic dispatch and risk assessment of the power system.

[0067] The present invention also discloses a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned method for generating a short-term multi-regional wind power scenario when executing the program.

[0068] The present invention also discloses a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the method for generating a short-term multi-regional wind power scenario is implemented.

[0069] Compared with the prior art, the advantages of the present invention are:

[0070] 1. Efficient modeling: By using the conditional latent space diffusion model, the present invention significantly reduces the computational complexity of high-dimensional wind power output scenarios. The model solves the posterior expression problem through an explicit diffusion process, so that the generated scenario not only conforms to the trend of the wind power output mode, but also retains the real detailed fluctuation characteristics.

[0071] 2. Dynamic graph representation: A dynamic graph neural network is introduced to capture the real-time spatial correlation between wind power output in multiple regions using a dynamic adjacency matrix. This method improves the denoising network’s accurate representation of the dynamic correlation between wind power in multiple regions, making the generated scenarios more consistent with actual wind power output.

[0072] 3. Optimize feature extraction: Through the supervised method of feature embedding, the model is guided to effectively extract key features, avoiding feature filtering of the denoising network in each step of the denoising process, thereby focusing on the mapping modeling task and significantly improving the performance of the denoising model.

[0073] 4. Detail reconstruction capability: The variational autoencoder (VAE) is used to map high-dimensional modeling tasks to low-dimensional latent space, which enhances the diffusion model’s ability to depict the detailed fluctuation characteristics of wind power output scenarios and further improves the model’s learning ability and generation effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 It is an architecture diagram of a method for generating short-term wind power scenarios in multiple regions according to an embodiment of the present invention;

[0075] Figure 2 is a flow chart of feature embedding and dynamic graph representation according to an embodiment of the present invention;

[0076] Figure 3 This is a flowchart of VAE latent space embedding according to an embodiment of the present invention;

[0077] Figure 4 is a diffusion and denoising flow chart of the conditional latent space diffusion model of an embodiment of the present invention;

[0078] Figure 5 4 is a diagram of a denoising network structure according to an embodiment of the present invention. DETAILED DESCRIPTION

[0079] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and examples.

[0080] The present invention provides a method for generating short-term wind power scenarios in multiple regions, which is realized by conditional latent space diffusion model technology and dynamic graph representation method, such as Figure 1 It mainly includes three parts: feature embedding and dynamic graph representation, latent space mapping, and conditional latent space diffusion model.

[0081] 1) Feature embedding and dynamic graph representation.

[0082] Using the pre-trained encoding network f e Embed the NWP input features of each region and use the deterministic prediction results Evaluate the real-time dynamic adjacency matrix Will As a conditional embedding of the diffusion model to exploit the denoising network f D The dynamic spatial relationship of wind power in multiple regions is mined, and the diffusion model does not need to filter high-dimensional NWP features at each step of denoising, which reduces the amount of calculation of the diffusion model. Multiple independent embedding networks are used to perform deterministic prediction regression on the high-dimensional NWP features of wind power in each region, and the Spearman correlation coefficient between the deterministic prediction results of wind power in multiple regions is used as the corresponding element of the dynamic adjacency matrix, thereby achieving real-time measurement of the dynamic correlation between different regions. The schematic diagram of feature embedding and dynamic graph representation is shown in the figure. Figure 2 shown.

[0083] When the NWP input data for the i-th region When embedding, the input layer convolutional neural network is used to adjust x i The number of channels is increased to obtain a high-dimensional representation of the features. The intermediate layer uses the residual connection structure of multiple convolutional layers to increase the receptive field of the input data to prevent network degradation. The output layer skips the results of each convolutional layer to achieve multi-time scale feature fusion. After the above steps, we can finally get x i The feature embedding results This embodiment uses L1 loss to perform deterministic prediction regression training on the embedding network, as shown in the following formula:

[0084]

[0085] Where: y i is the actual wind power output in region i, Embedded result for deterministic prediction of wind power in region i.

[0086] Repeat the above steps to get the embedding set of all regional features. Further according to By calculating the Spearman correlation coefficient between regions, we can obtain the dynamic adjacency matrix that represents the real-time spatial correlation of wind power. This method enables the diffusion model to explicitly guide the model to extract NWP key features in a supervised manner, decomposing the tasks of model feature extraction and scene mapping, avoiding feature filtering in each denoising process of the denoising network, and allowing the denoising network to focus on mapping and modeling the uncertainty of the prediction results.

[0087] 2) Hidden space mapping.

[0088] The pre-trained VAE is used as the upstream module of the diffusion model, where the encoder ε encodes and maps the high-dimensional wind power output curve Y of each region to obtain a low-dimensional latent space feature vector Z0, and the decoder D restores the high-resolution scene result S from Z0 through decoding mapping to improve the learning ability of the diffusion model. The VAE model is used to map and reconstruct the wind power output sequence in the latent space, and the wind power output trend learning task is assigned to the diffusion model, while the detail fluctuation learning task is assigned to the VAE decoder. The VAE latent space embedding diagram is shown in Figure 3 shown.

[0089] The encoder ε and decoder D of the VAE model are composed of two independent networks, where the encoder ε inputs the wind power output sequence Output mean vector and the standard deviation vector Then from the probability distribution p θ The latent space variables can be obtained by sampling (z|y) The decoder decodes and maps z to obtain the reconstructed wind power output sequence The training target of VAE is the posterior distribution p θ (z|y) and the prior distribution q φ The KL divergence between (z|y) is the smallest. After using variational inference to optimize the lower bound of the maximum likelihood function, the VAE loss function can be expressed as:

[0090]

[0091] Where: The first term is the posterior probability distribution p of the latent variable fitted by the encoding network θ The KL (Kullback-Leibler) divergence between (z|y) and the prior standard normal distribution p(z); the second term is the reconstruction error loss, which describes the difference between the generated scene and the historical scene.

[0092] KL divergence focuses on evaluating the difference in the overall trend of wind power output, but lacks the description of specific fluctuation information, which will cause the generated time series to be too smooth and unable to reconstruct the detailed information of wind power fluctuation. To this end, this embodiment introduces a weight coefficient λ to balance the reconstruction error loss and the KL divergence weight. The improved loss function is shown in the following formula:

[0093]

[0094] Where: ε(·) and are the mapping functions of the encoder and decoder respectively; λ is the weight coefficient.

[0095] By mapping the wind power sequence of each region into latent space through the encoder ε, the latent space set Z0 can be obtained, thereby converting the mapping modeling of the diffusion model in high-dimensional space to low-dimensional latent space, which helps to improve the learning ability of the diffusion model. As a downstream model of the diffusion model, the low-dimensional scene output by the denoising network is reconstructed into a high-dimensional time series scene, which can accurately characterize the detailed fluctuation characteristics of the wind power output scene.

[0096] 3) Conditional latent space diffusion model.

[0097] It includes diffusion process and denoising process. The diffusion process gradually adds noise to the latent space feature vector Z0 of each region to Z N In the denoising process, C is used as a covariate, and the denoising network is used to integrate spatiotemporal prior knowledge to achieve Z N The step-by-step denoising of The conditional latent space diffusion model is used to realize the mapping modeling of the joint probability distribution of wind power conditions in multiple regions. The conditional latent space diffusion model consists of two mutually inverse Markov chain processes: diffusion and denoising. Figure 4 shown.

[0098] The diffusion process gradually adds noise to the latent space sample Z0 within N steps, and finally makes the sample tend to be normally distributed. Assume that the time series sample after the nth step diffusion is represented by Z n , then from Z n-1 To Z n The diffusion process can be expressed as:

[0099]

[0100] Where: β∈(0,1) represents the scaling factor of the nth step; ∈ is the noise sample.

[0101] Z n It can be represented by Z0 through an n-order Markov chain process:

[0102]

[0103] make Then Z n The relationship between and Z0 can be expressed as:

[0104]

[0105] From the above formula, we can see that as n gradually increases, tends to 0, and finally Z n It will converge to a prior distribution, usually the standard normal distribution as a prior distribution.

[0106] The denoising process is the inverse process of the diffusion process. N ) is gradually denoised, and finally the latent space sample distribution q(Z0|C) is restored. The conditional distribution p of the denoising process is θ (Z n-1 |Z n ,C) is fitted, and the denoising process can be expressed as an n-order Markov chain process:

[0107]

[0108] The training objective of the denoising process is to minimize the posterior distribution p θ (Z n-1 |Z n ,C) and the prior distribution q(Z n-1 |Z n ,C) KL divergence. However, this prior distribution is usually difficult to obtain directly, so it can be transformed into a Bayesian formula to transform the denoising process into a fitting noise sample, which can be expressed as:

[0109]

[0110] Where: ∈ θ =f D (Z n ,n,C) is the output result of the denoising network, which determines how much noise needs to be removed in the current denoising step; f D (·) is the mapping function of the denoising network.

[0111] The denoising network inputs the denoising sample, the number of denoising steps n, the feature embedding, and the dynamic adjacency matrix A, and outputs the noise that should be removed in this step. The loss function of the denoising network training can be expressed as:

[0112]

[0113] p θ (Z n-1 |Z n ,C)

[0114] To achieve the posterior distribution p θ (Z n-1 |Z n ,C) Effective fitting, this embodiment designs a denoising network structure that considers the dynamic spatial correlation of multi-regional wind power, such as Figure 5 As shown in Figure 2, the denoising network adopts a residual and skip connection network structure. The feature encoding layer and the Linear layer map the discrete denoising step n to a continuous vector space, and the Conv convolution layer implements Z n and The temporal convolutional network (TCN) and the DGNN layer are combined to form a spatiotemporal convolutional layer to learn the spatiotemporal correlation features of the input data, and finally the nth step for sample Z is obtained. n The denoising result ∈ θ .

[0115] In addition, in order to improve the learning effect of the DGNN network, this embodiment adopts a learnable adaptive matrix W d Dynamic adjacency matrix For adaptive adjustment, the calculation process can be expressed as:

[0116]

[0117] Where: H l-1 and H l are the input and output of the lth graph convolutional layer respectively; W d are the parameters of the adaptive matrix; for degree matrix; W and b are the network learnable parameters of GCN; ⊙ represents the Hadamard product; σ is the activation function.

[0118] In another embodiment of the present invention, a multi-regional wind power short-term scenario generation system is provided. The system can be used to implement the above-mentioned multi-regional wind power short-term scenario generation method, specifically comprising: module

[0119] Data collection module: collects meteorological data and historical wind power output data of wind farms in multiple regions, and performs data cleaning and standardization to eliminate noise and outliers;

[0120] Feature embedding module: Use a pre-trained encoding network to embed the meteorological characteristics (NWP) of each region, and use the embedding results to evaluate the dynamic adjacency matrix between regions to capture the spatial correlation between regions;

[0121] Dynamic adjacency matrix calculation module: Based on the Spearman correlation coefficient of wind power output in each region, the dynamic adjacency matrix is ​​calculated in real time to reflect the dynamic correlation between wind power output in different regions;

[0122] Latent space mapping module: Use variational autoencoder (VAE) to map the high-dimensional wind power output curve to a low-dimensional latent space feature vector, extract the trend characteristics of wind power output through the encoder, and use the decoder to reconstruct the high-dimensional output sequence;

[0123] Conditional latent space diffusion model module: Construct a conditional latent space diffusion model including diffusion process and denoising process, set the diffusion process, gradually add noise to the latent space samples to normal distribution, and design a denoising network with conditional information as covariate to gradually remove noise and restore latent space samples;

[0124] Denoising network training module: The denoising network is trained using the structure of residual connection and spatiotemporal convolution layer to improve the learning ability of spatiotemporal related features. The training goal is to minimize the KL divergence between the posterior distribution and the prior distribution.

[0125] Scenario generation module: Reconstruct the low-dimensional latent space scenario output by the denoising network to generate a high-dimensional wind power output scenario, and finally use the generated scenario for the economic dispatch and risk assessment of the power system.

[0126] In another embodiment of the present invention, a terminal device is provided, the terminal device includes a processor and a memory, the memory is used to store a computer program, the computer program includes program instructions, and the processor is used to execute the program instructions stored in the computer storage medium. The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc., which are the computing core and control core of the terminal, which are suitable for implementing one or more instructions, and are specifically suitable for loading and executing one or more instructions to implement corresponding method processes or corresponding functions; the processor described in the embodiment of the present invention can be used for the operation of a multi-regional wind power short-term scenario generation method.

[0127] In another embodiment of the present invention, the present invention further provides a storage medium, specifically a computer-readable storage medium (Memory), which is a memory device in a terminal device for storing programs and data. It can be understood that the computer-readable storage medium here can include both the built-in storage medium in the terminal device and the extended storage medium supported by the terminal device. The computer-readable storage medium provides a storage space, which stores the operating system of the terminal. In addition, one or more instructions suitable for being loaded and executed by the processor are also stored in the storage space, and these instructions can be one or more computer programs (including program codes). It should be noted that the computer-readable storage medium here can be a high-speed RAM memory, or a non-volatile memory (non-volatile memory), such as at least one disk storage.

[0128] One or more instructions stored in a computer-readable storage medium can be loaded and executed by a processor to implement the corresponding steps of a multi-regional wind power short-term scenario generation method in the above embodiment; one or more instructions in the computer-readable storage medium are loaded and executed by the processor.

[0129] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0130] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0131] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0132] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0133] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the implementation methods of the present invention, and should be understood that the protection scope of the present invention is not limited to such special statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A method for generating short-term wind power scenarios in multiple regions, characterized by: The following steps are involved: S1: Collect meteorological data and historical wind power output data of wind farms in multiple regions; clean and standardize the collected data to eliminate noise and outliers; S2: Use the pre-trained encoding network to embed the meteorological features NWP of each region; use the embedding results to evaluate the dynamic adjacency matrix between regions to capture the spatial correlation between regions; S3: Based on the Spearman correlation coefficient of wind power output in each region, a dynamic adjacency matrix is ​​calculated in real time to reflect the dynamic correlation between wind power outputs in different regions; S4: Using variational autoencoder VAE, the high-dimensional wind power output curve is mapped to a low-dimensional latent space feature vector; the trend characteristics of wind power output are extracted through encoder E, and the latent space features are reconstructed into a high-dimensional output sequence using decoder D; S5: Construct a conditional latent space diffusion model, including a diffusion process and a denoising process. Set the diffusion process to gradually add noise to the latent space samples to normal distribution. Design a denoising network, use conditional information as a covariate, gradually remove noise, and restore the latent space samples. S6: The denoising network is trained using the structure of residual connection and spatiotemporal convolutional layer to improve the learning ability of spatiotemporal related features; The training goal is to minimize the KL divergence between the posterior distribution and the prior distribution to ensure the effectiveness of the denoising process; S7: Reconstruct the low-dimensional latent space scenario output by the denoising network to generate a high-dimensional wind power output scenario; use the generated scenario for economic dispatch and risk assessment of the power system.

2. The method for generating short-term multi-regional wind power scenarios according to claim 1, characterized in that: For each region i, the input meteorological data is expressed as T represents the time step, and C represents the feature dimension; Use the pre-trained encoding network f e x i Embed and generate high-dimensional representation of features; The encoding network is a convolutional neural network structure, including: input layer, intermediate layer and output layer; Input layer: adjust the number of channels to accommodate subsequent convolutional layers; Middle layer: Use multiple convolutional layers and increase the receptive field through residual connection structure to prevent network degradation; Output layer: Use skip connections to combine the results of each convolutional layer to achieve the fusion of multi-time scale features and finally obtain the embedding result 3. The method for generating short-term multi-regional wind power scenarios according to claim 1, characterized in that: The encoder network is trained using L1 loss for deterministic predictive regression: Among them, y i is the actual wind power output in region i, The result of embedding.

4. The method for generating short-term wind power scenarios in multiple regions according to claim 1, characterized in that: Embed all regions to obtain a feature embedding set: And construct a dynamic adjacency matrix based on this 5. The method for generating short-term multi-regional wind power scenarios according to claim 1, characterized in that: The sub-steps of S4 are as follows: S41: High-dimensional wind power output sequence Send to encoder E; Encoder E generates the mean vector and the standard deviation vector Thus, the probability distribution p of the latent space is defined θ (z|y); S42: Sampling latent space variables from the above probability distribution S43: Input the latent space variable z into the decoder D to generate the reconstructed wind power output sequence S44: Calculate the VAE loss function, including KL divergence and reconstruction error: After introducing the weight coefficient λ, the updated loss function is: Where: ε(·) and are the mapping functions of the encoder and decoder respectively; q φ (z|y) represents the posterior distribution of the encoder output, which represents the distribution of the latent variable z after the given input y; p θ (z|y) represents the reconstructed distribution generated by the decoder, which represents the probability of data recovered from the latent variable z; S45: Map the wind power sequence of each region through the encoder to obtain the latent space set Z0; S46: Use latent space features for downstream tasks of the diffusion model to reconstruct the low-dimensional scene output by the denoising network into a high-dimensional wind power output scene to ensure accurate characterization of detailed fluctuation characteristics.

6. The method for generating short-term multi-regional wind power scenarios according to claim 1, characterized in that: The diffusion process is as follows: The diffusion process gradually adds noise to the latent space sample Z0 to Z N , to achieve a gradual transition to a standard normal distribution, as follows: Where: β∈(0,1) represents the scaling factor of the nth step; ∈ is the noise sample, β n represents the scaling factor of the nth step, controlling the degree of noise addition, ∈ represents the random noise added during the diffusion process; The Markov chain process is expressed as: As n increases, the latent space samples tend to be standard normal distribution 7. The method for generating short-term multi-regional wind power scenarios according to claim 6, characterized in that: The denoising process is expressed as: The denoising goal is to minimize the KL divergence between the posterior distribution and the prior distribution, which is expressed as: In the formula, l ELBO represents the lower bound of evidence in variational inference, represents the expectation of all possible Z0 and noise ∈ sampled from a standard normal distribution, ||∈-∈ θ (Z n ,n,C)||2 represents the reconstruction error.

8. The method for generating short-term multi-regional wind power scenarios according to claim 7, characterized in that: The structural design of the denoising network aims to enhance dynamic spatial correlation and combines residual connections, skip connections, and graph convolutions. The graph convolutional layer is represented as: In the formula, H l-1 and H l are the input and output of the lth layer, W d The parameters of the adaptive matrix, for degree matrix; W and b are the network learnable parameters of GCN; ⊙ represents the Hadamard product; σ is the activation function.

9. A multi-regional wind power short-term scenario generation system, characterized by: The system can be used to implement the multi-regional wind power short-term scenario generation method according to any one of claims 1 to 8, specifically comprising: Data collection module: collects meteorological data and historical wind power output data of wind farms in multiple regions, and performs data cleaning and standardization to eliminate noise and outliers; Feature embedding module: Use the pre-trained encoding network to embed the meteorological feature NWP of each region, and use the embedding results to evaluate the dynamic adjacency matrix between regions to capture the spatial correlation between regions; Dynamic adjacency matrix calculation module: Based on the Spearman correlation coefficient of wind power output in each region, the dynamic adjacency matrix is ​​calculated in real time to reflect the dynamic correlation between wind power output in different regions; Latent space mapping module: Use variational autoencoder (VAE) to map the high-dimensional wind power output curve to a low-dimensional latent space feature vector, extract the trend characteristics of wind power output through the encoder, and use the decoder to reconstruct the high-dimensional output sequence; Conditional latent space diffusion model module: Construct a conditional latent space diffusion model including diffusion process and denoising process, set the diffusion process, gradually add noise to the latent space samples to normal distribution, and design a denoising network with conditional information as covariate to gradually remove noise and restore latent space samples; Denoising network training module: The denoising network is trained using the structure of residual connection and spatiotemporal convolution layer to improve the learning ability of spatiotemporal related features. The training goal is to minimize the KL divergence between the posterior distribution and the prior distribution. Scenario generation module: Reconstruct the low-dimensional latent space scenario output by the denoising network to generate a high-dimensional wind power output scenario, and finally use the generated scenario for the economic dispatch and risk assessment of the power system.

10. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the program is executed by a processor, the method for generating a multi-regional wind power short-term scenario as described in one of claims 1 to 8 is implemented.

Images