Conditional diffusion model-based Saggoer quarry wind-solar-solar-solar-term scene generation method and system

Abstract

The invention, which belongs to the technical field of the electric power system, discloses a method and a system for generating a Saggohm wind and light solar term scene based on a conditional diffusion model, and the method comprises the steps: firstly, taking 24 solar terms as a division interval of a historical time sequence, and endowing each solar term sample with a label; secondly, on the basis of a feature extraction network of the solar term factors, the solar term factors are mapped into a feature vector set of a hidden space, and the solar term law characteristics of wind and light output are effectively reserved; and finally, introducing a generative network based on a small sample learning technology, constructing a solar-term wind-light scene generation network based on a conditional diffusion probability model C-DDPM, converting prior distribution mapped by a single solar-term feature vector into joint probability distribution of 24 solar terms, capturing uncertainty and time sequence of output by adopting a denoising network architecture, and obtaining a wind-light scene based on the wind-light scene. And after training is completed, noise and corresponding solar term labels are input to generate a solar term scene. The method is especially suitable for regions with scarce data samples, and the model can generate diversified wind and light scenes through solar term conditions and noise input.

Description

Technical Field

[0001] The present invention belongs to the field of electric power systems, and in particular relates to a method and system for generating a Shagohuang wind and solar term scene based on a conditional diffusion model. Background Art

[0002] Building new energy bases in deserts, Gobi and wasteland areas is an important measure for my country to implement the "dual carbon" strategy. The Shagohuang base will become the main power provider. However, due to the volatility and intermittency of wind and solar power output, even the Shagohuang base with rich wind and solar resources will have low output for many consecutive days. Short-term scenarios at the daily level are difficult to describe meteorological fluctuations. Therefore, how to accurately simulate the continuous output of renewable energy for many days and construct medium- and long-term time series scenarios will directly affect the stable operation of the power system and planning decisions.

[0003] Among the existing scenario generation methods, the generative model is a relatively mature method for characterizing the uncertainty of renewable energy output. However, the existing division of historical wind and solar data samples usually does not fully consider the problems of unequal time interval lengths and insufficient distribution regularity. The time interval division in current research is mainly concentrated on weeks, months and quarters, and there are few methods for dividing historical samples based on the solar term scale. At the same time, the data-driven method requires a large amount of historical power data for training and supervision. In the case of missing data or scarce samples, the learning ability of the model is limited. Newly built large-scale wind and solar bases are often located in the deserted area, where historical power data is scarce and training conditions are limited. Existing scenario generation research rarely considers the problem of insufficient actual power samples, which limits the performance of scenario generation.

[0004] In general, for the medium- and long-term wind and solar power uncertainty problem in the Shagohuang area, there is a lack of solar term scenario generation methods that take into account the differences in data subset distribution and insufficient historical samples. Summary of the invention

[0005] The present invention aims to solve the technical problems existing in the background technology, and its purpose is to provide a method and system for generating Shagohuang wind and solar term scenes based on the conditional diffusion model; according to the output characteristics of historical power data, the twenty-four solar terms are used as the time measurement and division system. This interval division method has a high degree of conversion of meteorological types, which can not only accurately capture the performance of wind and solar output in different seasons, but also reveal the changing law of the system with the change of seasons, and provide data support for the subsequent medium- and long-term wind and solar scene generation. From the perspective of data expansion, a sample expansion strategy based on the conditional diffusion probability model is proposed to improve the performance of wind and solar simulation. The solar term label and the twenty-four solar term power data are input into the generative network together. Through the small sample data generation technology based on the generative network, the wind and solar power data of the twenty-four solar terms are jointly modeled, and the conditional joint probability distribution is fitted, so that the historical samples share parameters and the wind and solar scenes are efficiently generated.

[0006] In order to solve the technical problem, the technical solution of the present invention is:

[0007] A method for generating Shagohuang scenery solar term scenes based on a conditional diffusion model, the method comprising:

[0008] The 24 solar terms are used as the division intervals of the historical time series, the solar term samples are divided, and each solar term sample is assigned a label;

[0009] The feature extraction network is used to map the solar term label to the latent space, and the generated solar term feature vector c is used as the conditional vector;

[0010] Based on a generative network with small sample learning, a solar-solar-wind scene generation network based on the conditional denoising diffusion probability model C-DDPM was constructed. The prior distribution of a single solar-solar-term feature vector mapping was transformed into the joint distribution of the twenty-four solar terms. Through the iterative process of forward noising and reverse denoising, the uncertainty and temporal nature of the output were captured. The noise and solar-solar-term feature vector c were input into C-DDPM to generate the wind-solar-power scene under specific solar-solar-term conditions. Finally, through the collaborative modeling of the twenty-four solar-term data, the wind-solar-power scene that meets the characteristics of each solar term was generated.

[0011] Furthermore, the historical wind and solar power data are preprocessed. Considering the different number of days in leap years and common years and the different number of days of each solar term, the historical samples of each year are divided into 24 solar term samples with equal number of days according to the actual date and time of each solar term to construct a training set. Considering the periodicity of solar terms within the year, two-dimensional sine and cosine vectors are used to represent the solar term category, that is, the i-th solar term is encoded as c = [sin(2πi / 24), cos(2πi / 24)], and the solar term features are encoded as values ​​in the interval [-1,1] as the solar term label corresponding to each solar term sample.

[0012] Furthermore, the solar term labels are mapped into latent space through the solar term factor feature extraction network, and the solar term factors are mapped into a set of feature vectors in latent space, which effectively retains the solar term regularity characteristics of wind and solar power output. The generated solar term feature vector is used as the conditional vector, and the solar term feature vector c is used as the input condition of the conditional denoising diffusion probability model.

[0013] Furthermore, the solar term wind and light scene generation network based on the conditional denoising diffusion probability model C-DDPM consists of two parts: a forward process and a reverse process. In the forward process, the wind and light sequence x0 is converted into Gaussian noise z by iteratively adding prior noise so that it converges to the prior distribution. In the reverse process, the x0 time steps are de-noised from the prior distribution to generate a wind and light sequence by embedding the conditional label c.

[0014] The forward process is defined by a Markov chain, which gradually adds Gaussian noise to the original solar term sequence x0 until it becomes pure noise; x at any time in the process n Only with x n-1 Therefore, the C-DDPM forward process is defined as:

[0015]

[0016] Among them, q(x n |x n-1 ) is a Gaussian distribution, which can be expressed as:

[0017]

[0018] {β1,…,β N} is a time series with increasing variance, where β n ∈(0, 1) represents the forward step α n =1-β n Noise level;

[0019] Let α n =1-β n , By reparameterization technique, we can get any noise level n downsampled x n for:

[0020]

[0021] Where: as sampling noise;

[0022] The reverse process embeds the solar term condition c and adds noise to x N Denoising generates x0, which also follows a Markov chain, from a learnable Gaussian distribution To begin with, it is defined as:

[0023]

[0024] The distribution between two adjacent latent variables is expressed as:

[0025]

[0026] Where: θ is a shared parameter, p θ (x n-1 |x n , c) Parameterization:

[0027]

[0028]

[0029] Where: ∈ is the real noise, ∈ θ is the predicted value of ∈, is a trainable denoising function that determines how much noise should be removed in the current denoising step.

[0030] Furthermore, the training goal of the C-DDPM is to minimize the error between the actual noise and the predicted noise under different diffusion step sizes. The objective function of the training process is defined as:

[0031]

[0032] Furthermore, the input of the denoising network architecture consists of the noise-added samples The diffusion step number n and the solar term feature c are composed, and the output is the denoised noise ∈ θ ;

[0033] Use a one-dimensional convolutional layer to increase the number of channels of the input sequence x to C, and use a Sigmoid linear unit activation:

[0034] SiLu(x)=x*Sigmoid(x) (19)

[0035] Similar to the position embedding method proposed in Transformer, a sinusoidal encoding formula is applied to embed the noise level n:

[0036]

[0037] Then, a shared SiLu activated fully connected FC layer and a residual layer specific FC layer are used to project the encoded noise level into a C-dimensional vector and add it as a bias term to the convolution x; the solar term label condition c is transformed through the FCN U-Net network to transform the temporal resolution of the scenery, and then the channel is transformed through a one-dimensional convolution and added to the bias term of n and x for projection;

[0038] The main body of the model consists of N consistently structured layers connected in a residual manner so that the input information can be passed directly to the final layer; in each residual layer, a bidirectional dilated convolution Bi-DilConv is used to process the input so that it grows exponentially in the receptive field, and in the i-th residual layer, the spacing between the kernel points is set to 2i modulus gated units GU to activate the sum of the processed noise signal and the conditional signal; then, the convolution vector is divided into two parts and passed as residual output and skip output respectively; finally, all skip connections are summed and two convolutional layers are used to obtain a noise vector output with the same shape as x.

[0039] A Shagohuang scenery solar term scene generation system based on a conditional diffusion model, the system is applied to any of the above methods, and the system includes:

[0040] Solar term sample division module: used to use the 24 solar terms as the division interval of the historical time series, divide the solar term samples, and assign a label to each solar term sample;

[0041] Feature extraction module: used to map the solar term label to the latent space using the feature extraction network, and the generated solar term feature vector c is used as the conditional vector;

[0042] Conditional denoising diffusion probability model module: used for generative networks based on small sample learning, a solar-wind scene generation network based on the conditional denoising diffusion probability model C-DDPM is constructed, the prior distribution of a single solar-wind feature vector mapping is transformed into the joint distribution of the twenty-four solar terms, and the uncertainty and timing of the output are captured through the iterative process of forward denoising and reverse denoising. Noise and solar-wind feature vector c are input into C-DDPM to generate wind-wind power scenes under specific solar-wind conditions;

[0043] Collaborative modeling module: used to generate wind and solar power scenarios that meet the characteristics of each solar term through collaborative modeling of the 24 solar terms data.

[0044] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, any one of the above-mentioned methods for generating Shagohuang scenery and solar terms based on a conditional diffusion model is implemented.

[0045] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the above-mentioned methods for generating Shagohuang scenery and solar terms based on a conditional diffusion model.

[0046] Compared with the prior art, the advantages of the present invention are:

[0047] 1) By dividing historical samples into solar term samples of the same time interval and giving them labels, a feature extraction network is used to map the solar term labels into a set of feature vectors in the latent space. This not only effectively retains the solar term characteristics of the original wind and solar power output, but also enhances the interpretability of the model.

[0048] 2) The solar term label is used as the conditional input of C-DDPM, and the historical output data of the 24 solar terms are jointly trained to fit the joint probability distribution, so that the model can share parameters, thereby effectively solving the training difficulty caused by the scarcity of medium- and long-term simulated wind and solar historical samples in the Shagohuang area.

[0049] 3) The scenario generator generated by the proposed C-DDPM model can better capture the variability and uncertainty of solar and wind output during the solar term. This method can also be applied to the generation of wind and solar extremes, which is of great significance to the power balance problem faced by the power system and provides a new idea for the generation of renewable energy scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 , Solar term scene generation flow chart;

[0051] Figure 2 , C-DDPM basic structure diagram;

[0052] Figure 3 , denoising network structure design diagram; DETAILED DESCRIPTION

[0053] The specific implementation mode of the present invention is described below in conjunction with embodiments:

[0054] It should be noted that the structures, proportions, sizes, etc. shown in this specification are only used to match the contents disclosed in the specification so that people familiar with this technology can understand and read them, and are not used to limit the conditions under which the present invention can be implemented. Any structural modification, change in proportional relationship or adjustment of size should still fall within the scope of the technical content disclosed in the present invention without affecting the effects and purposes that can be achieved by the present invention.

[0055] At the same time, the terms such as "upper", "lower", "left", "right", "middle" and "one" cited in this specification are only for the convenience of description and are not used to limit the scope of implementation of the present invention. Changes or adjustments to their relative relationships should be regarded as the scope of implementation of the present invention without substantially changing the technical content.

[0056] Conditional Diffusion Model: The conditional diffusion model is a generative model that adds conditional information to the diffusion model so that the generated samples meet specific conditions or characteristics. It uses conditional probability to guide the data generation process, gradually transforms noisy data into target data, and gradually removes noise during the generation process. The core idea is to achieve accurate generation and control of data by simulating the diffusion process from noise to an ordered state. This model can generate high-quality data that meets specific conditions in many fields, such as image generation and text generation.

[0057] Solar term scenario generation: Solar term scenario generation is to perform probabilistic modeling on the uncertainty in a solar term period in the power system, and discretize these probabilistic models into finite solar term time series scenarios. Compared with directly using probabilistic models to describe uncertainty, solar term scenario generation is an extended study of the application of probabilistic models, which can more effectively apply theoretical innovations of probabilistic models to uncertain optimization problems in actual power systems. According to the time scale of the optimization decision-making problem, solar term scenario generation is mainly used for medium- and long-term scenario generation in power systems.

[0058] Embodiment 1:

[0059] The present invention includes important steps such as data preprocessing, feature extraction, model building, collaborative modeling and generation process. The following is an overview of the logic:

[0060] 24 solar terms division and sample labeling: The historical time series data is divided according to the 24 solar terms, and each solar term sample corresponds to a specific label. This step segments the data according to the characteristics of the solar terms (such as date and time).

[0061] Feature extraction network generates conditional vectors: The feature extraction network is used to map the solar term labels to the latent space, and the latent space feature vector c is input as the conditional vector into the subsequent generation model. This process retains the characteristic information of the solar term.

[0062] Conditional probability network structure: Through the small sample learning method, the C-DDPM model is constructed to fit the conditional probability distribution. The C-DDPM model introduces the solar term feature vector c, and through the iterative process of forward denoising and reverse denoising, it realizes the generation of wind and solar power scenarios under specific solar term conditions.

[0063] Furthermore, the conditional denoising diffusion probability model (C-DDPM): This model consists of two processes: forward denoising and reverse denoising. In the forward process, the original wind-solar power sequence is gradually converted into Gaussian noise by iteratively adding noise; in the reverse process, the noise is denoised by introducing the solar term label condition, and the wind-solar output curve is gradually restored. The forward process is a Markov chain that gradually adds noise to the sequence, while the reverse process restores the power sequence by learning the denoising steps.

[0064] Denoising network captures uncertainty and temporality: The denoising network plays an important role in the reverse denoising process. It can capture the volatility (uncertainty) of wind and solar power output and its changing trend over time (temporality), so that the generated power sequence is more in line with the actual situation.

[0065] Furthermore, the denoising network uses residual connections and bidirectional dilated convolution structures to process the input signal and gradually remove noise. The solar terms transform the data through the U-Net network and then embed it into the model for joint training. After convolution processing, the output is divided into two parts for residual transmission, and finally summed through jump connections to generate a noise vector output with the same shape as the input.

[0066] Collaborative modeling to generate solar term scenarios: Through collaborative modeling of all 24 solar term data, C-DDPM can learn the correlation and change patterns between different solar terms, and ultimately generate wind and solar power scenarios that conform to the characteristics of each solar term.

[0067] Furthermore, by conducting joint training on wind and solar power data in the Shagohuang area, this method can capture the spatiotemporal correlation and fluctuation characteristics of wind and solar multimodal variables. After the training is completed, random noise and solar term labels are input to generate wind and solar power output scenarios related to specific solar terms.

[0068] This paper uses the C-DDPM model to fit the joint probability distribution law of wind and solar output under the conditions of the 24 solar terms through collaborative modeling and denoising, and can effectively generate the 24 solar terms scenes, which is particularly suitable for wind and solar power prediction and expansion in areas with a small number of samples. This overall logic accurately describes the solar term wind and solar scene generation process based on C-DDPM, highlighting the model's capture of solar term characteristics and the processing of data uncertainty and temporality. It shows the entire process from data partitioning, feature extraction joint modeling and model generation.

[0069] Embodiment 2:

[0070] This embodiment is applied to Embodiment 1. This embodiment divides the solar term samples, maps the solar term labels to the latent space using the feature extraction network, and uses the generated solar term feature vector as the conditional vector. A conditional diffusion model suitable for sample expansion is designed, and the model is trained through the forward denoising and reverse denoising process to capture the correlation changes and fluctuation characteristics of wind and solar power output. Through the collaborative modeling of the 24 solar term data, the conditional joint probability distribution law is fitted, so as to efficiently generate the 24 solar term scenes.

[0071] The specific solar term scene generation process is as follows: Figure 1 shown.

[0072] First, the historical wind and solar power data are preprocessed. Considering the different number of days in leap years and normal years and the different number of days in each solar term, the historical samples of each year are divided into 24 solar term samples with equal number of days according to the actual date and time of each solar term to construct a training set. Since the solar terms are periodic within the year, a two-dimensional sine and cosine vector is used to represent the solar term category, that is, the i-th solar term is encoded as c = [sin (2πi / 24), cos (2πi / 24)], and the solar term features are encoded as values ​​in the interval [-1, 1] as the solar term label corresponding to each solar term sample.

[0073] The solar term labels are mapped into latent space through the solar term factor feature extraction network, and the solar term factors are mapped into feature vector sets in latent space, which effectively retains the solar term regularity characteristics of wind and solar power output. The obtained solar term feature vector is used as the conditional vector, and the proposed conditional probability model uses the solar term feature c as the condition.

[0074] As a generative model, the conditional denoising diffusion probabilistic model (C-DDPM) does not randomly generate a set of power sequences, but generates a solar term output curve that follows a certain distribution based on the solar term characteristics. C-DDPM is used to establish the mapping relationship between the conditional input and the latent space vector. Figure 2 .

[0075] C-DDPM consists of two parts: the forward process and the reverse process. In the forward process, the scenery sequence x0 is converted into Gaussian noise z by iteratively adding prior noise, so that it converges to the prior distribution. The reverse process removes the noise from the prior distribution of x0 time steps by embedding the conditional label c to generate the scenery sequence.

[0076] The forward process is defined by a Markov chain, which gradually adds Gaussian noise to the original solar term sequence x0 until it becomes pure noise. n Only with x n-1It is related to , but not to the solar term label c. Therefore, the forward process of C-DDPM is defined as:

[0077]

[0078] Among them, q(xx|x n-1 ) is a Gaussian distribution, which can be expressed as:

[0079]

[0080] {β1,…,β N} is a time series with increasing variance, where β n ∈(0, 1) represents the forward step α n =1-β n noise level.

[0081] Let α n =1-β n , By reparameterization technique, we can get any noise level n downsampled x n for:

[0082]

[0083] Where: as sampling noise.

[0084] The reverse process embeds the solar term condition c and adds noise to x N Denoising generates x0, which also follows a Markov chain, from a learnable Gaussian distribution To begin with, it is defined as:

[0085]

[0086] The distribution between two adjacent latent variables is expressed as:

[0087]

[0088] Where: θ is a shared parameter, p θ (x n-1 |x n , c) Parameterization:

[0089]

[0090]

[0091] Where: ∈ is the real noise, ∈ θ is the predicted value of ∈, is a trainable denoising function that determines how much noise should be removed in the current denoising step.

[0092] The training goal of C-DDPM is to minimize the error between the actual noise and the predicted noise under different diffusion step sizes. Therefore, the objective function of the training process is defined as:

[0093]

[0094] The input of the denoising network consists of the noisy samples The diffusion step number n and the solar term feature c are composed, and the output is the denoised noise ∈ θ , its structure diagram is as follows Figure 3 shown.

[0095] like Figure 3 As shown, a one-dimensional convolutional layer is used to increase the number of channels of the input sequence x to C, and a Sigmoid linear unit (SiLU) activation is adopted:

[0096] SiLu(x)=x*Sigmoid(x) (29)

[0097] Similar to the position embedding method proposed in Transformer, a sinusoidal encoding formula is applied to embed the noise level n:

[0098]

[0099] Then, the encoded noise level is projected into a C-dimensional vector using a shared SiLu activated fully connected (FC) layer and a residual layer specific FC layer, which is added as a bias term to the convolution x. The solar term label condition c is transformed through the FCN U-Net network to transform the temporal resolution of the scenery, and then the channel is transformed through a one-dimensional convolution and added to the bias term of n and x for projection.

[0100] The main body of the model consists of N layers of consistent structure connected in a residual manner so that the input information can be passed directly to the final layer. In each residual layer, a bidirectional dilated convolution (Bi-DilCon v ) to process the input so that it grows exponentially in the receptive field, and in the i-th residual layer, the spacing between kernel points is set to 2i modulus gated units (GUs) are used to activate the sum of the processed noise signal and the conditional signal. Then, the convolution vector is split into two parts and passed as residual output and skip output respectively. Finally, all skip connections are summed and two convolutional layers are used to obtain a noise vector output with the same shape as x.

[0101] By building a solar term scene generation network based on C-DDPM and using the iterative process of forward denoising and reverse denoising, the samples with less data in the Shagohuang area are jointly trained and expanded. This process can learn the spatiotemporal correlation and fluctuation characteristics between wind and light multimodal variables. After the model training is completed, random noise and solar term labels are input to generate the corresponding wind and light solar term scenes.

[0102] Embodiment 3:

[0103] The present invention provides a Shagohuang scenery and solar term scene generation system based on a conditional diffusion model, which can be used to implement the above-mentioned Shagohuang scenery and solar term scene generation method based on a conditional diffusion model, and the system includes:

[0104] Solar term sample division module: used to use the 24 solar terms as the division interval of the historical time series, divide the solar term samples, and assign a label to each solar term sample;

[0105] Feature extraction module: used to map the solar term label to the latent space using the feature extraction network, and the generated solar term feature vector c is used as the conditional vector;

[0106] Conditional denoising diffusion probability model module: used for generative networks based on small sample learning, a solar-wind scene generation network based on the conditional denoising diffusion probability model C-DDPM is constructed, the prior distribution of a single solar-wind feature vector mapping is transformed into the joint distribution of the twenty-four solar terms, and the uncertainty and timing of the output are captured through the iterative process of forward denoising and reverse denoising. Noise and solar-wind feature vector c are input into C-DDPM to generate wind-wind power scenes under specific solar-wind conditions;

[0107] Collaborative modeling module: used to generate wind and solar power scenarios that meet the characteristics of each solar term through collaborative modeling of the 24 solar terms data.

[0108] Embodiment 4:

[0109] This embodiment provides a terminal device, which includes a processor and a memory, wherein 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. It is the computing core and control core of the terminal, which is suitable for implementing one or more instructions, and is specifically suitable for loading and executing one or more instructions to implement the corresponding method flow or corresponding function; the processor described in the embodiment of the present invention can be used for the operation of the Shagohuang scenery solar term scene generation method based on the conditional diffusion model, including the following steps:

[0110] The 24 solar terms are used as the division intervals of the historical time series, the solar term samples are divided, and each solar term sample is assigned a label;

[0111] The feature extraction network is used to map the solar term label to the latent space, and the generated solar term feature vector c is used as the conditional vector;

[0112] Based on a generative network with small sample learning, a solar-solar-wind scene generation network based on the conditional denoising diffusion probability model C-DDPM was constructed. The prior distribution of a single solar-solar-term feature vector mapping was transformed into the joint distribution of the twenty-four solar terms. Through the iterative process of forward noising and reverse denoising, the uncertainty and temporal nature of the output were captured. The noise and solar-solar-term feature vector c were input into C-DDPM to generate the wind-solar-power scene under specific solar-solar-term conditions. Finally, through the collaborative modeling of the twenty-four solar-term data, the wind-solar-power scene that meets the characteristics of each solar term was generated.

[0113] Embodiment 5:

[0114] This embodiment 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 is understandable 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, such as at least one disk memory.

[0115] The processor may load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the method for generating the Shagohuang scenery solar term scene based on the conditional diffusion model in the above embodiment; the processor may load and execute the following steps:

[0116] The 24 solar terms are used as the division intervals of the historical time series, the solar term samples are divided, and each solar term sample is assigned a label;

[0117] The feature extraction network is used to map the solar term label to the latent space, and the generated solar term feature vector c is used as the conditional vector;

[0118] Based on a generative network with small sample learning, a solar-solar-wind scene generation network based on the conditional denoising diffusion probability model C-DDPM was constructed. The prior distribution of a single solar-solar-term feature vector mapping was transformed into the joint distribution of the twenty-four solar terms. Through the iterative process of forward noising and reverse denoising, the uncertainty and temporal nature of the output were captured. The noise and solar-solar-term feature vector c were input into C-DDPM to generate the wind-solar-power scene under specific solar-solar-term conditions. Finally, through the collaborative modeling of the twenty-four solar-term data, the wind-solar-power scene that meets the characteristics of each solar term was generated.

[0119] 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.

[0120] 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.

[0121] 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.

[0122] 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.

Claims

1. The method for generating Shagohuang scenery solar term scenes based on conditional diffusion model is characterized by: The method comprises: The 24 solar terms are used as the division intervals of the historical time series, the solar term samples are divided, and each solar term sample is assigned a label; The feature extraction network is used to map the solar term label to the latent space, and the generated solar term feature vector c is used as the conditional vector; Based on a generative network with small sample learning, a solar-solar-wind scene generation network based on the conditional denoising diffusion probability model C-DDPM was constructed. The prior distribution of a single solar-solar-term feature vector mapping was transformed into the joint distribution of the twenty-four solar terms. Through the iterative process of forward noising and reverse denoising, the uncertainty and temporal nature of the output were captured. The noise and solar-solar-term feature vector c were input into C-DDPM to generate the wind-solar-power scene under specific solar-solar-term conditions. Finally, through the collaborative modeling of the twenty-four solar-term data, the wind-solar-power scene that meets the characteristics of each solar term was generated.

2. The method for generating Shagohuang scenery and solar term scenes based on the conditional diffusion model according to claim 1 is characterized in that: The historical wind and solar power data are preprocessed. Considering the different number of days in leap years and common years and the different number of days of each solar term, the historical samples of each year are divided into 24 solar term samples with equal number of days according to the actual date and time of each solar term to construct a training set. Considering the periodicity of solar terms within the year, two-dimensional sine and cosine vectors are used to represent the solar term category, that is, the i-th solar term is encoded as c = [sin(2πi / 24), cos(2πi / 24)], and the solar term features are encoded as values ​​in the interval [-1,1] as the solar term label corresponding to each solar term sample.

3. The method for generating Shagohuang scenery and solar term scenes based on conditional diffusion model according to claim 1 is characterized in that: The solar term labels are mapped into latent space through the solar term factor feature extraction network, and the solar term factors are mapped into a feature vector set in latent space, which effectively retains the solar term regularity characteristics of wind and solar power output. The generated solar term feature vector is used as the conditional vector, and the solar term feature vector c is used as the input condition of the conditional denoising diffusion probability model.

4. The method for generating Shagohuang scenery and solar term scenes based on conditional diffusion model according to claim 1 is characterized in that: The solar term wind and light scene generation network based on the conditional denoising diffusion probability model C-DDPM consists of two parts: a forward process and a reverse process. In the forward process, the wind and light sequence x0 is converted into Gaussian noise z by iteratively adding prior noise so that it converges to the prior distribution. In the reverse process, the x0 time steps are removed from the prior distribution to generate a wind and light sequence by embedding the conditional label c. The forward process is defined by a Markov chain, which gradually adds Gaussian noise to the original solar term sequence x0 until it becomes pure noise; x at any time in the process n Only with x n-1 Therefore, the C-DDPM forward process is defined as: Among them, q(x n ∣x n-1 ) is a Gaussian distribution, which can be expressed as: {β1,…,β N } is a time series with increasing variance, where β n ∈(0,1) represents the forward step α n =1-β n Noise level; Let α n =1-β n , By reparameterization technique, we can get any noise level n downsampled x n for: Where: as sampling noise; The reverse process embeds the solar term condition c and adds noise to x N Denoising generates x0, which also follows a Markov chain, from a learnable Gaussian distribution To begin with, it is defined as: The distribution between two adjacent latent variables is expressed as: Where: θ is a shared parameter, p θ (x n-1 |x n ,c) Parameterization: Where: ∈ is the real noise, ∈ θ is the predicted value of ∈, is a trainable denoising function that determines how much noise should be removed in the current denoising step.

5. The method for generating Shagohuang scenery and solar term scenes based on conditional diffusion model according to claim 1 is characterized in that: The training goal of the C-DDPM is to minimize the error between the actual noise and the predicted noise under different diffusion step sizes. The objective function of the training process is defined as:

6. The method for generating Shagohuang scenery and solar term scenes based on conditional diffusion model according to claim 1 is characterized in that: The input of the denoising network architecture consists of the noisy samples The diffusion step number n and the solar term feature c are composed, and the output is the denoised noise ∈ θ ; Use a one-dimensional convolutional layer to increase the number of channels of the input sequence x to C, and use a Sigmoid linear unit activation: SiLu(x)=x*Sigmoid(x) (9) Similar to the position embedding method proposed in Transformer, a sinusoidal encoding formula is applied to embed the noise level n: Then, a shared SiLu activated fully connected FC layer and a residual layer specific FC layer are used to project the encoded noise level into a C-dimensional vector and add it as a bias term to the convolution x; the solar term label condition c is transformed through the FCN U-Net network to transform the temporal resolution of the scenery, and then the channel is transformed through a one-dimensional convolution and added to the bias term of n and x for projection; The main body of the model consists of N consistently structured layers connected in a residual manner so that the input information can be passed directly to the final layer; in each residual layer, a bidirectional dilated convolution Bi-DilConv is used to process the input so that it grows exponentially in the receptive field, and in the i-th residual layer, the spacing between the kernel points is set to 2i modulus gated units GU to activate the sum of the processed noise signal and the conditional signal; then, the convolution vector is divided into two parts and passed as residual output and skip output respectively; finally, all skip connections are summed and two convolutional layers are used to obtain a noise vector output with the same shape as x.

7. The Shagohuang scenery solar term scene generation system based on the conditional diffusion model is characterized by: The system is applied to the method described in any one of claims 1 to 6, and the system comprises: Solar term sample division module: used to use the 24 solar terms as the division interval of the historical time series, divide the solar term samples, and assign a label to each solar term sample; Feature extraction module: used to map the solar term label to the latent space using the feature extraction network, and the generated solar term feature vector c is used as the conditional vector; Conditional denoising diffusion probability model module: used for generative networks based on small sample learning, a solar-wind scene generation network based on the conditional denoising diffusion probability model C-DDPM is constructed, the prior distribution of a single solar-wind feature vector mapping is transformed into the joint distribution of the twenty-four solar terms, and the uncertainty and timing of the output are captured through the iterative process of forward denoising and reverse denoising. Noise and solar-wind feature vector c are input into C-DDPM to generate wind-wind power scenes under specific solar-wind conditions; Collaborative modeling module: used to generate wind and solar power scenarios that meet the characteristics of each solar term through collaborative modeling of the 24 solar terms data.

8. A computer device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the program, the method for generating the Shagohuang scenery and solar term scene based on the conditional diffusion model as described in any one of claims 1 to 6 is implemented.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the method for generating the Shagohuang scenery and solar term scene based on the conditional diffusion model as described in any one of claims 1 to 6.

Images