A two-stage intelligent generation method for water and wind power scheduling scenarios

Abstract

The application discloses a two-stage intelligent generation method of water, wind and light scheduling scenes, and belongs to the technical field of water, wind and light scene generation. The method comprises the following steps: S1, historical data collection and missing value interpolation; S2, data normalization and standardized dataset construction; S3, improved VAE-GAN model training and decoding; and S4, spatiotemporal correlation enhancement. The two-stage intelligent generation method of water, wind and light scheduling scenes is adopted, so that the stability of model training is improved; only small sample historical data is needed for training; runoff, wind power and photovoltaic three elements and multi-variable combined scene generation are supported, day-level and year-level long-term scales are covered, and the accuracy of restoring the complex spatiotemporal characteristics of water, wind and light data is improved.

Description

Technical Field

[0001] This invention belongs to the field of water, wind and light scene generation technology, specifically involving a two-stage intelligent generation method for water, wind and light scheduling scenes. Background Technology

[0002] The increasing depletion of traditional fossil fuels and the intensification of climate change have driven the rapid development of multi-energy complementary systems dominated by renewable energy sources such as hydropower, wind power, and photovoltaics. However, the runoff and wind and solar power output within a river basin have strong uncertainties, complex spatiotemporal correlations, and high-dimensional nonlinear characteristics, making them difficult to accurately characterize and quantify. This increases the uncertainty and complexity of multi-energy complementary systems, posing significant challenges to their planning and scheduling.

[0003] However, existing technologies have the following shortcomings: mainstream generative models are prone to training crashes, generate blurry samples, and are difficult to adapt to the strong uncertainty, complex spatiotemporal correlations, and high-dimensional nonlinear characteristics of water, wind, and light data; most methods require sufficient historical data support, and the model's generalization ability decreases in small sample scenarios, failing to meet the needs of data-scarce scenarios; they focus on single-element or two-element scene generation, with time scales mostly short-term, making it difficult to achieve high-dimensional (three-element and above) and long-term scene generation; traditional statistical methods have weak mapping capabilities for high-dimensional nonlinear data, making it difficult to accurately characterize the spatiotemporal correlations of water, wind, and light data.

[0004] Therefore, a new method is urgently needed. Summary of the Invention

[0005] The purpose of this invention is to provide a two-stage intelligent generation method for water, wind and solar power scheduling scenarios. This method improves the stability of model training; it only requires a small sample of historical data for training; it supports the generation of scenarios combining runoff, wind power and solar power as three elements and multiple variables, covering daily and grade-level long-term scales, filling gaps; and it improves the accuracy of restoring the complex spatiotemporal characteristics of water, wind and solar power data.

[0006] To achieve the above objectives, this invention provides a two-stage intelligent generation method for water, wind, and solar power scheduling scenarios, comprising the following steps: S1. Collect multi-year daily-scale historical data of runoff, wind power output, and photovoltaic power output in the basin, and use the Lagrange interpolation method to impute the missing data to obtain the imputed data. S2. Normalize the data after interpolation in S1 to obtain a standardized dataset; S3. Generate the initial scene for the first stage. Based on the standardized dataset from S2, construct an improved VAE-GAN model framework to obtain the initial scene; this includes the following steps: S301. Construct a VAE-GAN hybrid model framework; S302. Based on the complex spatiotemporal correlation and high-dimensional features of the data, a convolutional neural network is adopted as the core network structure of the encoder, decoder and discriminator. S303. Set the parameters of the VAE-GAN model; S304. Based on the parameter settings in S303, the runoff-wind power-photovoltaic data in the standardized dataset transmitted in S2 are dimensionally integrated. Based on the network structure designed in S302, the integrated dataset is input into the improved VAE-GAN model, and the model is trained according to the parameter configuration in S303 until the model converges. S305. Extract the decoder module from the improved VAE-GAN model trained in S304 and fix its network structure and parameters; input noise and feed it into the fixed decoder, which then outputs the corresponding data samples; obtain the initial scenario of runoff-wind power-photovoltaic combined. S4. Correct the temporal and spatial distribution characteristics of the initial scene transmitted in S3 by using quantile mapping and Cholesky decomposition methods to enhance spatiotemporal correlation, including the following steps: S401. Enhance the temporal relevance of the initial scene transmitted by S3 using the quantile mapping method; S402. Perform Cholesky decomposition on the scene obtained in S401 to enhance the spatial correlation of the scene and obtain the final runoff-wind power-solar scene set.

[0007] Preferably, S1 is as follows: Historical data on runoff, wind power output, and photovoltaic power output within the basin were collected over many years, with the data time scale standardized to "day". Lagrange interpolation was used to fill in any missing values ​​in the data.

[0008] Preferably, S2 is as follows: The runoff, wind power output, and photovoltaic power output data transmitted by S1 were normalized respectively; to ensure consistency when the dataset is input into the model, the length of "one year" in the dataset was fixed at 365 days.

[0009] Preferably, the VAE-GAN hybrid model framework is constructed in S301 as follows: Using the "generator-discriminator" adversarial structure of the GAN model as the core framework, the "encoder-decoder" structure of VAE is integrated into the generator structure. The input samples are sequentially encoded, reconstructed and discriminated. The distribution characteristics of historical data are initially captured through the minimax game of GAN. By introducing Wasserstein distance to replace JS divergence, the improved generator loss, discriminator loss, and model training objective function formulas are as follows: ; in, The expected value corresponding to the distribution; The output value of the discriminator; This is real data; To generate data; For generator loss; For discriminator loss; The formula for the VAE training objective is: ; in, The loss is the reconstruction error between the generated sample and the real sample; For the posterior probability distribution With prior probability distribution KL divergence between them; The final loss function of the improved VAE-GAN model is expressed as: ; in, , , These are the coefficients of the loss function.

[0010] Preferably, S302 is as follows: The encoder contains five two-dimensional convolutional layers. The first four layers have a kernel size of 3×3, a stride of 1, and padding of 1. The fifth layer has a kernel size of 3×4, a stride of 1×2, and padding of 1. By connecting two fully connected layers, the processed data is mapped to a latent vector with a dimension of 182. The decoder network structure is completely consistent with the encoder, with a convolution kernel size of 3×3, a sliding stride of 1, and padding of 1, which is used to reconstruct the latent vectors into water, wind and light data samples. The discriminator network structure and parameter settings are consistent with those of the encoder, and it is used to distinguish whether the input sample is real historical data or model-generated data.

[0011] Preferably, S303 is as follows: Set the key training parameters for the VAE-GAN model, including batch size, number of iterations, learning rate, and loss function coefficients.

[0012] Preferably, noise is input into S305, and this noise is input into a fixed decoder, specifically as follows: Generated by random sampling These noises are input into a fixed decoder, which outputs corresponding data samples to obtain... An initial scenario combining runoff, wind power, and solar power.

[0013] Preferably, S401 is as follows: For each variable (runoff, wind power, solar power) and for each day of the year, calculate the cumulative distribution function of the initial scenario simulation data. Cumulative distribution function of historical data ; For each simulated scene point Calculate its quantile in the cumulative distribution function of the simulated data, using the following formula: ; Applying quantile thresholds avoids mapping simulated data to extreme quantiles; the formula is: ); in, It is a positive number; The adjusted quantiles are calculated using interpolation. The formula for mapping the corresponding value to the cumulative distribution function of historical data is: ; Among them, a smoothing process is applied to the minimum value, if ,but: ; in, This is the minimum value in historical data.

[0014] Preferably, S402 is as follows: Obtain the year-by-year runoff output sequence from the S401 output scenario set. Wind power output sequence Photovoltaic power output sequence The three sequences were standardized using Z-score, and the standardized data were used to construct a sample matrix. ; Based on the sample matrix Calculate the original Kendall rank correlation matrix of the initial scene. Calculate the target Kendall rank correlation matrix based on historical data. ; respectively and The Cholesky decomposition is performed using the following formula: ; ; in, , It is a lower triangular matrix; Obtain the corresponding lower triangular matrix and ,according to and Calculate the transformation matrix By applying changes to the sample matrix, samples with enhanced spatial correlation are obtained. The calculation formula is: ; ; The enhanced sample matrix Perform Z-score denormalization, and the resulting data is the scene with enhanced spatial correlation after Cholesky decomposition.

[0015] This invention also provides a two-stage intelligent generation system for water, wind, and solar energy scheduling scenarios, comprising: The data acquisition module is used to collect multi-year daily-scale historical data of runoff, wind power output, and photovoltaic power output within the watershed. The Lagrange interpolation method is used to impute missing data to obtain the imputed data. A data preprocessing module, connected to the data acquisition module, is used to normalize the interpolated data to obtain a standardized dataset. The initial scene generation module is connected to the data preprocessing module and is used to generate the initial scene in the first stage. Based on the standardized dataset, an improved VAE-GAN model framework is constructed to obtain the initial scene. The scene post-processing module, connected to the initial scene generation module, is used to correct the temporal and spatial distribution characteristics of the initial scene through quantile mapping and Cholesky decomposition methods, thereby enhancing the spatiotemporal correlation.

[0016] Therefore, the present invention employs the aforementioned two-stage intelligent generation method for water, wind, and solar energy scheduling scenarios. Compared with existing technologies, the technical solution of the present invention has the following beneficial effects: (1) This invention introduces Wasserstein distance to replace JS divergence, and combines it with a three-parameter loss function to improve the stability of model training, avoid mode collapse, and generate clearer scenes; (2) This invention improves the VAE-GAN model structure and two-stage strategy, which can complete the training with only a small sample of historical data, thus breaking the dependence on large-scale data; (3) This invention supports the generation of scenarios combining runoff, wind power, and photovoltaic elements with multiple variables, and covers long-term scenarios at the daily and grade (365 days) time scales, filling the gap in existing technologies; (4) In the second stage of this invention, the temporal distribution is corrected by quantile mapping and spatial correlation is enhanced by Cholesky decomposition, which accurately restores the complex spatiotemporal characteristics of water, wind and light data and has better feature capture capabilities than traditional methods.

[0017] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0018] Figure 1 This is a flowchart of an embodiment of a two-stage intelligent generation method for water, wind and solar scheduling scenarios according to the present invention; Figure 2 This is a schematic diagram of the VAE-GAN model structure of an embodiment of a two-stage intelligent generation method for water, wind and light scheduling scenarios according to the present invention. Figure 2 (a) in the diagram represents the network structure of the discriminator; Figure 2 (b) in the diagram represents the network structure of the encoder in the generator; Figure 2 (c) in the generator represents the network structure of the decoder. Figure 3 This is a confidence interval diagram of a runoff-wind power-photovoltaic scenario generated by an embodiment of a two-stage intelligent generation method for water, wind and solar power scheduling scenarios according to the present invention. Figure 4 This is a bar chart showing the absolute error of the average Kendall correlation coefficient between an embodiment of the two-stage intelligent generation method for water, wind and solar scheduling scenarios of the present invention and scenarios generated by VAE-GAN, GAN and VAE models. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Unless otherwise defined, the technical or scientific terms used in the present invention should have the ordinary meaning understood by those skilled in the art.

[0020] Example 1 like Figures 1-4 As shown, the present invention provides a two-stage intelligent generation method for water, wind, and solar power scheduling scenarios, comprising the following steps: S1. Collect 22 years of historical data on runoff, wind power output, and photovoltaic power output within the basin, with a time scale of days. For data with missing values, use the Lagrange interpolation method to imput the missing values. S2. Normalize the runoff, wind power output, and photovoltaic power output data respectively to construct a standardized dataset; to ensure consistency when inputting the dataset into the model, set the length of a year in the dataset to 365 days. S3. Generate the initial scene for the first stage. Based on the standardized dataset from S2, construct an improved VAE-GAN model framework to obtain the initial scene; this includes the following steps: The S301 and VAE-GAN models use the "generator-discriminator" adversarial structure of GAN models as their core framework, integrating the "encoder-decoder" structure of VAEs into the generator structure. Input samples are sequentially encoded, reconstructed, and discriminated. Through the minimax game within the GAN model, the VAE-GAN model can initially capture the distribution characteristics of historical data. In the early stages of training, if the real data... and generating data Since the distributions do not overlap, the Jensen-Shannon (JS) divergence, which describes the difference between the two distributions, will be a constant value, leading to training difficulties and pattern collapse. Therefore, when the two probability distributions do not overlap, the Wasserstein distance is introduced as an alternative. Finally, the generator loss is improved using Wasserstein. Discriminator loss The formula for the model training objective function is as follows: ; in, The expected value corresponding to the distribution; The output value of the discriminator; The VAE model uses an encoder to process real data. Latent variables are generated through encoding and learned through a decoder. With data distribution The mapping relationship between them is determined, and samples are reconstructed. Latent variables and real data can form a joint probability density distribution. The training objective of VAE, after being simplified, is formulated as follows: ; in, The loss is the reconstruction error between the generated sample and the real sample; For the posterior probability distribution With prior probability distribution KL divergence between them; The VAE-GAN model, by sequentially encoding, reconstructing, and discriminating input samples, can more effectively utilize learned latent features to measure sample similarity. To ensure the stability of model training and the quality of generated scenes, Wasserstein distance is introduced to calculate the distance between the probability distributions of generated and real samples. , , The three parameters are used to weight the various loss coefficients to balance the quality and diversity of the samples generated by the model. The final improved loss function formula is as follows: ; in, , , These are the coefficients of the loss function; S302. A Convolutional Neural Network (CNN) is used as the core network structure. Utilizing the local feature extraction capability of two-dimensional convolution (Conv2d), the CNN kernels slide across the spatiotemporal dimension to fully learn the correlation between data from adjacent time points. The encoder part of the generator contains five two-dimensional convolutional layers (the first four layers have a kernel size of 3×3, a stride of 1, and padding of 1; the fifth layer has a kernel size of 3×4, a stride of 1×2, and padding of 1). By connecting two fully connected layers, the data is mapped to a latent vector of dimension 182. The decoder part has the same structure as the encoder (kernel size of 3×3, stride of 1, and padding of 1). The discriminator's network structure and parameter settings are consistent with the encoder. S303. Set the VAE-GAN model parameters as follows: batch size = 22, epochs = 5000, learning rate (lr) for both discriminator and generator = 0.0001, and loss function coefficients... , , The values ​​are 0.8, 0.001, and 0.5 respectively. S304. Based on the batch size, perform dimensional integration on the runoff-wind power-photovoltaic data in the dataset, with a shape size of [batch_size, 1, 3, 365]. Then, based on the network structure designed in step S302 and the parameters set in step S303, use the integrated data as model input and train the model. S305. Extract the decoder trained in step S304, fix its network structure and parameters, sample N=1000 noises as input, and obtain N=1000 initial scenarios of runoff-wind power-photovoltaic combined. S4. Correct the temporal and spatial distribution characteristics of the initial scene transmitted in S3 by using quantile mapping and Cholesky decomposition methods to enhance spatiotemporal correlation, including the following steps: S401. Enhance the temporal correlation of the initial scene set obtained in S305 using the quantile mapping method: For each variable (runoff, wind power, solar power) and for each day of the year, calculate the cumulative distribution function of the initial scenario simulation data. Cumulative distribution function of historical data ; For each simulated scene point Calculate its quantile in the cumulative distribution function of the simulated data, using the following formula: ; Applying quantile thresholds avoids mapping simulated data to extreme quantiles; the formula is: ); in, It is a positive number; The adjusted quantiles are calculated using interpolation. The formula for mapping the corresponding value to the cumulative distribution function of historical data is: ; Among them, a smoothing process is applied to the minimum value, if ,but: ; in, This is the minimum value in historical data; S402. Obtain the year-by-year runoff output sequence from the output scenario set of S401. Wind power output sequence Photovoltaic power output sequence The three sequences were standardized using Z-score, and the standardized data were used to construct a sample matrix. ; Based on the sample matrix Calculate the original Kendall rank correlation matrix of the initial scene. Calculate the target Kendall rank correlation matrix based on historical data. ; respectively and The Cholesky decomposition is performed using the following formula: ; ; in, , It is a lower triangular matrix; Obtain the corresponding lower triangular matrix and ,according to and Calculate the transformation matrix By applying changes to the sample matrix, samples with enhanced spatial correlation are obtained. The calculation formula is: ; ; The enhanced sample matrix Perform Z-score denormalization, and the resulting data is the scene with enhanced spatial correlation after Cholesky decomposition.

[0021] like Figures 3-4 As shown, a comprehensive evaluation system is constructed from two aspects: spatial relevance and reliability, to verify the superiority of the technical solution: Spatial correlation evaluation uses the average Kendall correlation coefficient absolute error index. By calculating the correlation between the number of scenes, the Kendall correlation coefficient between each group of scenes and the Kendall correlation coefficient of historical data, the smaller the value, the more accurate the generated scene's representation of the spatial correlation between water, wind and light variables.

[0022] The reliability evaluation selects six indicators: coverage percentage (PICP), average power range width (PINAW), average coverage error (ACE), comprehensive evaluation index CWC, WS, and PL, to measure the reliability of the generated scenario from multiple dimensions.

[0023] Using single VAE-GAN, GAN, and VAE models as comparison objects, we also verify the effect of coupling the two-stage strategy with GAN and VAE models.

[0024] Evaluation results show that the scenes generated by this scheme perform well in terms of spatiotemporal correlation representation and reliability. Most indicators are better than those of comparative models such as VAE-GAN, GAN, and VAE. Moreover, the two-stage strategy can effectively improve the scene generation quality of a single model.

[0025] The scenarios generated in this embodiment outperform the comparison models in most metrics, with all ACE values ​​being positive and reaching the given confidence level. The spatiotemporal correlation error is significantly lower than that of the comparison methods, accurately depicting the complex relationship between the three elements of water, wind, and light. The two-stage strategy can effectively improve the scenario quality of a single model, with a maximum improvement rate of 70.27%. The generated scenarios can provide reliable data support for power system planning.

[0026] Table 1 lists the reliability index results of the runoff-wind power-photovoltaic scenarios generated by the two-stage scenario generation method proposed in this invention and other models in a quantitative manner.

[0027] Table 1 Reliability Index Data for Runoff-Wind Power-Photovoltaic Combined Scenarios

[0028] Table 2 shows the results of coupling GAN and VAE models with the two-stage strategy after replacing the first-stage scene generation model, and the improvement rates of three comprehensive indicators (CWC, WS, PL) compared with the corresponding single-stage models.

[0029] Table 2 Comparison of Comprehensive Indicators and Improvement Rates

[0030] Therefore, this invention adopts the above-mentioned two-stage intelligent generation method for water, wind and solar scheduling scenarios. This method improves the stability of model training; it only requires a small sample of historical data for training; it supports the generation of scenarios combining runoff, wind power and photovoltaic elements and multiple variables, covering daily and grade-level long-term scales, filling gaps; and it improves the accuracy of restoring the complex spatiotemporal characteristics of water, wind and solar data.

[0031] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied 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.

[0032] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A two-stage intelligent generation method for water, wind, and light scheduling scenarios, characterized in that, Includes the following steps: S1. Collect multi-year daily-scale historical data of runoff, wind power output, and photovoltaic power output in the basin, and use the Lagrange interpolation method to impute the missing data to obtain the imputed data. S2. Normalize the data after interpolation in S1 to obtain a standardized dataset; S3. Generate the initial scene for the first stage. Based on the standardized dataset from S2, construct an improved VAE-GAN model framework to obtain the initial scene; this includes the following steps: S301. Construct a VAE-GAN hybrid model framework; S302. Based on the complex spatiotemporal correlation and high-dimensional features of the data, a convolutional neural network is adopted as the core network structure of the encoder, decoder and discriminator. S303. Set the parameters of the VAE-GAN model; S304. Based on the parameter settings in S303, the runoff-wind power-photovoltaic data in the standardized dataset transmitted in S2 are dimensionally integrated. Based on the network structure designed in S302, the integrated dataset is input into the improved VAE-GAN model, and the model is trained according to the parameter configuration in S303 until the model converges. S305. Extract the decoder module from the improved VAE-GAN model trained in S304 and fix its network structure and parameters; input noise and feed it into the fixed decoder, which then outputs the corresponding data samples; obtain the initial scenario of runoff-wind power-photovoltaic combined. S4. Correct the temporal and spatial distribution characteristics of the initial scene transmitted in S3 by using quantile mapping and Cholesky decomposition methods to enhance spatiotemporal correlation, including the following steps: S401. Enhance the temporal relevance of the initial scene transmitted by S3 using the quantile mapping method; S402. Perform Cholesky decomposition on the scene obtained in S401 to enhance the spatial correlation of the scene and obtain the final runoff-wind power-solar scene set. 2.The two-stage intelligent generation method of water, wind and light scenes according to claim 1, characterized in that, S1 specifically refers to: Historical data on runoff, wind power output, and photovoltaic power output within the basin were collected over many years, with the data time scale standardized to "day". Lagrange interpolation was used to fill in any missing values ​​in the data.

3. The two-stage intelligent generation method for a water, wind, and solar power scheduling scenario according to claim 1, characterized in that, S2 specifically refers to: The runoff, wind power output, and photovoltaic power output data transmitted by S1 were normalized respectively; to ensure consistency when the dataset is input into the model, the length of "one year" in the dataset was fixed at 365 days.

4. The two-stage intelligent generation method for a water, wind, and solar power scheduling scenario according to claim 1, characterized in that, The VAE-GAN hybrid model framework is constructed in S301 as follows: Using the "generator-discriminator" adversarial structure of the GAN model as the core framework, the "encoder-decoder" structure of VAE is integrated into the generator structure. The input samples are sequentially encoded, reconstructed and discriminated. The distribution characteristics of historical data are initially captured through the minimax game of GAN. By introducing Wasserstein distance to replace JS divergence, the improved generator loss, discriminator loss, and model training objective function formulas are as follows: ； wherein, is a distribution corresponding to a desire; is an output value of a discriminator; is real data; is generated data; is a generator loss; is a discriminator loss; The formula for the VAE training objective is: ； wherein, is a reconstruction error loss between the sample and the real sample; is a KL divergence between the posterior probability distribution and the prior probability distribution ​ The final loss function of the improved VAE-GAN model is expressed as: ； in, , , These are the coefficients of the loss function.

5. The two-stage intelligent generation method of water, wind and light scenes according to claim 1, characterized in that, S302 specifically refers to: The encoder contains five two-dimensional convolutional layers. The first four layers have a kernel size of 3×3, a stride of 1, and padding of 1. The fifth layer has a kernel size of 3×4, a stride of 1×2, and padding of 1. By connecting two fully connected layers, the processed data is mapped to a latent vector with a dimension of 182. The decoder network structure is completely consistent with the encoder, with a convolution kernel size of 3×3, a sliding stride of 1, and padding of 1, which is used to reconstruct the latent vectors into water, wind and light data samples. The discriminator network structure and parameter settings are consistent with those of the encoder, and it is used to distinguish whether the input sample is real historical data or model-generated data.

6. The two-stage intelligent generation method of water, wind and light scenes according to claim 1, characterized in that, S303 specifically refers to: Set the key training parameters for the VAE-GAN model, including batch size, number of iterations, learning rate, and loss function coefficients.

7. The two-stage intelligent generation method of water, wind and light scenes according to claim 1, characterized in that, The S305 inputs noise, and this noise is fed into a fixed decoder, specifically as follows: By random sampling to generate The noise is input into the fixed decoder, and the corresponding data sample is output by the decoder to obtain The runoff-wind power-photovoltaic combined initial scene is obtained. 8.The two-stage intelligent generation method of water, wind and light scenes according to claim 1, characterized in that, S401 specifically refers to: Cumulative distribution functions of the initial scenario simulation data and of the historical data are calculated for each variable, i.e. runoff, wind power, photovoltaics, and for each day of the year Cumulative distribution functions of the initial scenario simulation data and of the historical data are calculated for each variable, i.e. runoff, wind power, photovoltaics, and for each day of the year ; For each simulated scenario point its quantile in the simulated data cumulative distribution function is calculated, given by: ； Applying quantile thresholds avoids mapping simulated data to extreme quantiles; the formula is: )； wherein is positive; The adjusted quantiles are calculated using interpolation. The formula for mapping the corresponding value to the cumulative distribution function of historical data is: ； where the minimum value is smoothed, and if then: ； wherein is a minimum value in the historical data.

9. The two-stage intelligent generation method of water, wind and light scenes according to claim 1, characterized in that, S402 specifically refers to: Obtaining S401 the runoff output sequence of each year in the output scene set , the wind power output sequence , the photovoltaic output sequence , respectively performing Z_score standardization processing on the three sequences, and constructing the standardized data into a sample matrix ; According to the sample matrix Compute the original Kendall rank correlation matrix of the initial scene , compute the target Kendall rank correlation matrix according to the historical data ; respectively, Cholesky decomposition is performed on and , and the formula is: ； ； wherein , is a lower triangular matrix; corresponding lower triangular matrix is obtained and , according to and a transformation matrix is calculated , the change is applied to the sample matrix, and a sample after enhancement of spatial correlation is obtained , and the calculation formula is: ； ； The enhanced sample matrix Perform Z-score denormalization, and the resulting data is the scene with enhanced spatial correlation after Cholesky decomposition.

10. A two-stage intelligent generation system for water, wind, and light scheduling scenarios, applied to the two-stage intelligent generation method for water, wind, and light scheduling scenarios as described in any one of claims 1-9, characterized in that, include: The data acquisition module is used to collect multi-year daily-scale historical data of runoff, wind power output, and photovoltaic power output within the watershed. The Lagrange interpolation method is used to impute missing data to obtain the imputed data. A data preprocessing module, connected to the data acquisition module, is used to normalize the interpolated data to obtain a standardized dataset. The initial scene generation module is connected to the data preprocessing module and is used to generate the initial scene in the first stage. Based on the standardized dataset, an improved VAE-GAN model framework is constructed to obtain the initial scene. The scene post-processing module, connected to the initial scene generation module, is used to correct the temporal and spatial distribution characteristics of the initial scene through quantile mapping and Cholesky decomposition methods, thereby enhancing the spatiotemporal correlation.

Images