Enhanced GAN extreme scene generation method and system considering weather factor

Abstract

The invention discloses an enhanced GAN extreme scene generation method and system considering weather factors, and relates to the technical field of scene generation in new energy power system planning and operation, and the method comprises the steps: extracting key factors from multi-time-delay meteorological data, forming an extreme label, converting the extreme label into a generation condition, and achieving the output through two-stage confrontation generation, wherein a scarce sample is stably amplified by WGAN-GP, a joint curve is generated according to conditions, and related structures, climbing and power boundary constraints are synchronously embedded to ensure statistical and physical consistency. According to the method, the authenticity of the distribution tail, the cross-variable correlation and the time structure can be considered when the sample is scarce. Extreme situation deduction of a novel power system can be supported while extreme situation coverage and credibility are improved, and an existing new energy scene analysis and planning process is integrated.

Description

Technical Field

[0001] This invention relates to the field of scene generation technology in the planning and operation of new energy power systems, and more specifically, to an enhanced GAN extreme scene generation method and system that takes weather factors into account. Background Technology

[0002] With the rapid increase in the penetration rate of renewable energy sources such as wind and solar power in the power system, the uncertainty of system operation has increased significantly. Wind and solar power output is random, fluctuating, and intermittent, exhibiting large and sudden changes under extreme weather conditions. For example, drastic changes in wind speed during typhoons may cause wind turbines to shut down, blizzards may block solar panels, causing a sharp drop in solar power output, while cold waves and high temperatures can cause a sharp increase in system load. These extreme scenarios often exhibit the characteristics of "low probability, high impact." Without sufficient historical samples for modeling, power system dispatching, planning, and risk assessment are prone to distortion, potentially threatening the safety and stability of the power grid in severe cases.

[0003] Currently, scene generation methods are mainly divided into three categories: First, scene generation methods based on statistical modeling, typical methods include probability distribution fitting, time series analysis, Markov chains, and Copula functions. These methods generate new samples by statistically modeling historical data, which can better characterize the distribution characteristics and partial correlations of wind and solar power output. Second, scene generation methods based on sampling and optimization, representative methods include Monte Carlo sampling, Latin hypercube sampling, scene reduction, and clustering. These methods rely on assumptions about the distribution of historical data, obtain scene sets through large-scale sampling, and reduce redundant scenes through optimization. Third, generative model methods based on deep learning, such as Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), and their improved models, are gradually becoming research hotspots in new energy scene generation. GANs, through adversarial training between generators and discriminators, can realistically learn complex data distributions and are suitable for dealing with nonlinear and high-dimensional scene generation problems.

[0004] However, existing technologies still face several shortcomings. First, extreme samples are scarce: the number of samples corresponding to extreme weather in historical databases is limited, making it difficult for traditional data-driven models to effectively capture these distribution characteristics. Second, there is a lack of causal relationship characterization: existing methods are mostly based on correlation modeling, with less consideration for the causal relationship between meteorological factors and power system operation, resulting in insufficient interpretability and physical consistency of the generated scenarios. Third, there is a lack of multidimensional coupling: most studies focus on a single energy source (such as wind power or photovoltaics), with less consideration for the multidimensional coupling characteristics of wind power, photovoltaics, and loads, making it difficult to reflect the overall operating status of the system under extreme weather conditions. Summary of the Invention

[0005] To address the aforementioned issues, the present invention aims to provide an enhanced GAN extreme scene generation technology that considers weather factors. This technology effectively overcomes the shortcomings of existing methods in terms of sample scarcity, physical consistency, and multidimensional coupling, thereby providing reliable data support for power system planning and scheduling.

[0006] To achieve the above technical objectives, this application provides an enhanced GAN extreme scene generation method that considers weather factors, comprising the following steps: Based on meteorological factors that have a significant impact on extreme weather, and based on the impact of extreme weather on the operation and planning of the power system, the causal relationship between meteorological time series and power time series is verified. Based on causal relationships, time series clustering and scene classification are performed on multidimensional meteorological factors, extreme sample sets are constructed, and sample augmentation processing is carried out. The enhanced extreme sample set is used as input to a conditional generative adversarial network to finally generate extreme scenarios by coupling wind, light, and load.

[0007] Preferably, when extracting meteorological factors, features of meteorological shadows that have a significant impact on extreme weather are extracted based on historical wind load and meteorological data and using dCor distance correlation ranking.

[0008] Preferably, when verifying the causal relationship between meteorological time series and power time series, the Granger causality test is used to test the causal relationship between meteorological time series and power time series.

[0009] Preferably, when performing time series clustering, the K-multidimensional time series clustering algorithm KMDTSC is used to perform time series clustering on the above-mentioned multidimensional meteorological factors.

[0010] Preferably, when classifying scenes, the elbow method is used to observe the curve of the sum of squared errors within a cluster (SSE) as a function of the number of clusters (K). When the decrease in SSE is no longer significant when K is further increased, this point is taken as the optimal number of clusters (K). Based on the causal relationship passed by the Granger causality test, extreme scene clusters are determined for scene classification.

[0011] Preferably, when performing sample augmentation, the extreme samples of the cluster are augmented using a generative adversarial network WGAN-GP with gradient penalty.

[0012] Preferably, when generating extreme scenes by coupling wind, light, and load on the samples, the enhanced extreme scene sample set is used as the input of the conditional generative adversarial network CGAN to generate the extreme scenes by coupling wind, light, and load on the samples.

[0013] This invention also discloses an enhanced GAN extreme scene generation system that considers weather factors, used to execute the aforementioned enhanced GAN extreme scene generation method that considers weather factors, comprising: The relationship analysis module is used to verify the causal relationship between meteorological time series and power time series based on meteorological factors that have a significant impact on extreme weather and the impact of extreme weather on the operation and planning of the power system. The sample extraction and enhancement module is used to perform time series clustering and scene classification of multidimensional meteorological factors based on causal relationships, construct extreme sample sets, and perform sample enhancement processing. The scene generation module is used to take the enhanced extreme sample set as input to the conditional generative adversarial network and finally generate extreme scenes by coupling wind, light and load on the samples.

[0014] The present invention discloses the following technical effects: 1. Enhanced characterization and explanation of meteorological factors and power system scenarios: This invention introduces a meteorological factor-driven mechanism in scenario generation, using the dCor distance correlation method to screen key meteorological variables and combining Granger causality tests to establish causal relationships between meteorological factors and wind power, photovoltaics, and load curves. This makes the generated extreme scenarios more interpretable and credible.

[0015] 2. Two-stage scene generation mechanism of WGAN-GP+CGAN: WGAN-GP with gradient penalty is used to augment the small samples of extreme scenes obtained from clustering, achieving data augmentation and solving the problem of limited sample size for extreme weather scenes, providing a sufficient data foundation for subsequent modeling. The second step, CGAN, uses scene categories as conditional labels to achieve conditional generation of different extreme weather scenes, ensuring the diversity and relevance of scene generation.

[0016] 3. Generation of Extreme Wind-Solar-Load Coupling Scenarios: Under extreme weather conditions, there is a strong coupling relationship between wind, solar, and load. For example, during blizzards, increased wind speeds cause wind turbines to shut down, solar power output drops sharply due to snow accumulation, and low temperatures simultaneously increase electricity load demand. Considering only a single variable often underestimates system risks. This invention generates coupled scenarios of the three elements through joint modeling, which can reflect the dynamic matching relationship between renewable energy output and load demand under different extreme weather conditions. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is the Granger causality test flowchart described in this invention; Figure 2 This is a schematic diagram of the method described in this invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0020] like Figures 1-2 As shown, this invention provides an enhanced GAN extreme scene generation technology that considers weather factors. It is an enhanced generative adversarial network method that integrates meteorological factors, focusing on the generation of extreme weather scenarios in power systems to address the problem of small sample sizes in extreme scenarios. Firstly, it extracts key meteorological factors through the Hilbert-Schmidt independence criterion and Granger causality test to ensure the physical interpretability of the generated scenarios. Secondly, it constructs diverse extreme sample sets using clustering algorithms and enhances training stability through WGAN-GP. Finally, it uses a conditional generative adversarial network to achieve the generation of extreme scenarios coupled with wind, solar, and load factors, providing reliable input for power system scheduling and planning. Specifically, it includes the following steps: Step 1: Use the distance correlation ranking method to extract features from meteorological factors that have a significant impact on extreme weather.

[0021] In step 1, based on historical wind load and meteorological data, features of meteorological shadows that have a significant impact on extreme weather are extracted using dCor distance correlation ranking. This includes the following steps: Step 1.1: Calculate each meteorological measurement exist The values ​​for each hour are expanded into candidate columns, with the j-th column denoted as... The target is a three-tiered system of wind, solar, and lotus. To eliminate the influence of dimensions and extreme values, median-MAD standardization is performed column-by-column for both candidate and target samples, as follows: in, The median, The median absolute deviation is usually... , They represent wind, light, and lotus respectively.

[0022] Step 1.2: To support non-linear dependency measurement, any standardized column 'a' is converted into a sample pair distance matrix and then doubly centered to make different variables comparable in the same relative distance coordinate system, as follows: Where a is a column vector. To The matrix after doubly centralization Let be the pairwise distance matrix of 'a'. The matrix is ​​centered, where N represents the number of samples and 1 represents a vector of length N consisting entirely of 1s.

[0023] Step 1.3: For each candidate column With three objectives respectively Calculate the distance correlation (dCor), and then take an equal-weighted average to obtain the comprehensive score. dCor is between (0,1), and the larger it is, the stronger the dependency. It can capture both linear and nonlinear relationships, as follows: in, Represents two standardized vectors to be compared, corresponding to ; Represents covariance; This represents the distance variance; Let be the comprehensive score of the j-th candidate feature. and This represents the matrix after doubly centering u and v respectively.

[0024] Step 1.4, Press By screening from high to low lags and merging adjacent selected time lags for the same meteorological quantity into an effective lag interval, the key meteorological factors and their time-dependent ranges can be obtained. It is important to note that feature importance only indicates the contribution of meteorological factors to extreme weather events and does not reflect causal relationships. Step 2: Use Granger causality test to examine the causal relationship between meteorological time series and power time series.

[0025] In step 2, combinations of meteorological factors of varying intensities reflect different types of extreme weather, which have varying impacts on the operation and planning of the power system. Considering meteorological factors such as temperature, precipitation, sunshine, wind speed, and snowfall, correlation analysis is used to quantify their impact on wind power, solar power, and load. A Granger causality test is established, including the following steps: Step 2.1: Clarify the objects of the time series X and Y in the Granger test. Select meteorological factors such as temperature, wind speed, irradiance, precipitation, and rainfall as the time series X, and select power scenario curves such as wind power, photovoltaic power, and load curves as the time series Y. That is, discuss whether X can improve the predictive ability of Y. If so, X is the Granger cause of Y.

[0026] Step 2.2: Perform stationarity (mean, variance, and covariance do not change over time) and cointegration tests on series X and Y. For each time series, first perform the Augmented Dickey-Fuller (ADF) test. If stationary, proceed directly to the Granger test; if non-stationary, differencing is required. If X and Y become stationary after differencing, further testing is needed to determine if a long-term equilibrium relationship exists, i.e., a cointegration test; otherwise, spurious regression may occur.

[0027] Step 2.3: Assuming the lag length is k, establish an unconstrained least squares regression model U containing the sequence x and a constrained model R without x, and perform estimation, defined as follows: In the formula, The coefficient of the lagged term, Indicates a constant coefficient. This represents the time series of wind, solar, and load.

[0028] Step 2.4: Propose the null hypothesis H0: X is not a Granger cause of Y, i.e., all lag coefficients... To test the constraint F, separate regressions were performed on the lag coefficients including X and lag coefficients excluding X, and the sum of squared residuals of RSSR and RSSU were recorded. The formula for the F statistic is as follows: In the formula, m is the number of lagged terms of X, n is the sample size, and k is the number of parameters to be estimated in the unconstrained regression model that includes possible constant terms and other variables. and These are the residual sums of squares for the constrained model and the unconstrained model, respectively. If the calculated F-value is greater than the critical value of the F-distribution at a given significance level α... Then the null hypothesis can be rejected. This leads to the conclusion that meteorological time series is a Granger trait of power time series.

[0029] Step 3: Use the K-multidimensional time series clustering algorithm to perform time series clustering on multidimensional meteorological factors, classify scenarios, and construct an extreme sample set.

[0030] In step 3, meteorological factors have higher dimensions and more significant characteristics compared to power system scenarios. Scenario classification based on meteorological factors can further clarify scenario characteristics and guide the generation of extreme scenarios. The K-Multi-Dimensional Time Series Clustering Algorithm (KMDTSC) is used to perform time series clustering of the above multidimensional meteorological factors, including the following steps: Step 3.1: Key meteorological factors (temperature, precipitation, radiation, wind speed, snowfall, etc.) selected through dCor distance correlation and Granger test are used as input. These meteorological factors are often multi-dimensional, time-series data. The K-MDTSC algorithm is selected, which is an improvement on the traditional K-means algorithm. It is mainly designed for multi-dimensional time-series data and updates the distance metric, allowing clustering to consider both time variation and multi-factor coupling.

[0031] Step 3.2: Update the distance calculation for the multidimensional time series. Using N as the dimension of the clustered samples and T as the time scale of the clustered samples, the distance calculation formula for K-MDTSC is as follows: In the formula, L is the distance metric. and It is a synchronous multidimensional meteorological time series. and These represent meteorological data at different times t.

[0032] Step 3.3: Define the performance of the clustering algorithm. Similar to traditional clustering algorithms, the square of the sum of distances between sample points and their cluster centers, i.e., the sum of squared errors within each cluster (SSE) or cluster inertia, is used to evaluate the algorithm's performance and determine the optimal number of clusters. The formula for calculating SSE is as follows: In the formula, C is a cluster. It is the set of all clusters.

[0033] Step 3.4: Identify extreme scenario clusters. Using the elbow method, observe the curve of SSE changing with the number of clusters K. When the decrease in SSE becomes insignificant as K continues to increase, take that point as the optimal number of clusters K. Using the previous Granger causality test, select those clusters that have a significant causal relationship with power scenarios to enhance the interpretability of the clustering results. Examples include clusters for typhoons, high temperatures, cold waves, blizzards, and heavy rain, which will be used for subsequent targeted generation and expansion of GANs.

[0034] Step 3.5: Construct a labeled sample set. Label the selected extreme scenarios with category c, for example, Class 1: low wind speed and low wind power; Class 2: high temperature and high load, etc. Establish the learning sample set format: ,in For wind power series, For photovoltaic power series, For load power sequence, Labels for extreme scenarios.

[0035] Step 4: Use a generative adversarial network with gradient penalty to augment the clustered extreme samples, generating an augmented extreme scenario sample set.

[0036] In step 4, the extreme samples of the clustering are augmented using a generative adversarial network (WGAN-GP) with gradient penalty to generate an enhanced extreme scene sample set, thereby improving the training set quality of the generative adversarial network. This includes the following steps: Step 4.1: Define the basic model structure of a Generative Adversarial Network (GAN). The model structure contains two deep neural networks: a generator and a discriminator. These two neural networks engage in an adversarial game to gradually generate fake samples similar to real samples. The generator takes random noise as input and aims to generate samples similar to real data, while the discriminator distinguishes between real and generated data, guiding the generator to optimize its output. The optimization objective of GAN is to minimize the distribution difference between generated and real data, equivalent to minimizing the JS divergence, which is defined as follows: In the formula, and The distributions of real data and generated data are respectively; It is a mixed distribution of the two; Let KL divergence be defined as: ; Step 4.2: Define the loss function of the generator. The main task of the generator is to generate samples that approximate the true distribution from random noise, and the optimization objective is to minimize the loss function. The definition is as follows: In the formula, Samples generated by the generator; This is the score given by the evaluator to the generated samples. This loss function guides the generator to produce samples that are closer to the true distribution.

[0037] Step 4.3: Define the loss function of the evaluator. The goal of the evaluator is to measure the distributional difference between the input samples and the true data. WGANA-GP uses the Wasserstein distance as the optimization objective, and its mathematical definition is as follows: In the formula, The evaluation function must satisfy the Lipschitz continuity constraint, i.e. .

[0038] Secondly, WGANA-GP improves model training stability by replacing weight clipping with gradient penalty to constrain Lipschitz continuity in a more stable way. The mathematical expression is as follows: In the formula, The generator produces interpolated samples by mixing real samples with the generated samples in a certain proportion. The gradient of the evaluator over the interpolated samples; This represents the gradient penalty strength coefficient.

[0039] Finally, the loss function of the evaluator is defined, which combines the Wasserstein distance and the gradient penalty term, and is defined as follows: In the formula, Let be the score given by the evaluator to the real samples. The first term maximizes the score of the real samples, ensuring the evaluator has high confidence in the real data; the second term minimizes the score of the generated samples, distinguishing between real and generated data; the third term is the gradient penalty term, ensuring that the gradient update conforms to the Lipschitz constraint.

[0040] Step 4.4: Design the model network structure of WGAN-GP. The model network structure is shown in Table 1: Table 1 Step 4.5: Divide the extreme scene wind-solar-load data into different clusters, generate data using WGAN-GP, and add the generated new data to the sample set until the data in the extreme scene sample library is greater than or equal to the data capacity requirement of the scene generation, thus completing the data augmentation of the scene samples.

[0041] Step 5: Use the enhanced extreme scene sample set as input to the conditional generative adversarial network to finally generate extreme scenes by coupling wind, light and load.

[0042] In step 5, the enhanced extreme scene sample set is used as input to the conditional generative adversarial network (CGAN) to finally generate extreme scenes by coupling wind, light, and load. This includes the following steps: Step 5.1: Define the objective functions for the generator and evaluator of CGAN. CGAN combines supervised and semi-supervised learning, allowing additional information to be used as input to the generator during sample generation, guiding the sample generation process. The evaluator's output for both generated samples and augmented real samples is defined as follows: In the formula, CGAN assumes that the input sample of the generator G is noise z, which follows the distribution p(z), and the real sample is x, which follows the data distribution p(x). c represents the additional sample information of the input, i.e., the extreme scene category label. This represents the generated sample, which serves as the input to discriminator D. The output of discriminator D is a scalar. ; , These represent the discriminator's output values ​​for real samples and generated samples, respectively.

[0043] Objective function of generator in CGAN The objective function of the discriminator It can be represented as: In the formula, , This represents the expected value for each batch of samples.

[0044] Ultimately, the two networks iterate continuously until the network parameters stabilize and reach Nash equilibrium, which is expressed as: ; Step 5.2: Design the CGAN model network structure, as shown in Table 2: Table 2 Step 5.3: Use the data-enhanced extreme scene sample set from Step 4 as the training set for CGAN. The information in the additional samples c differs for different categories of scenes, guiding the generation of extreme scenes of different categories. Finally, when the CGAN network parameters reach equilibrium, it can complete the generation of extreme scenes (such as typhoons, high temperatures, cold waves, blizzards, and heavy rain) with weather factors (such as temperature, precipitation, radiation, wind speed, snowfall, etc.).

[0045] This invention introduces a meteorological factor-driven mechanism into scene generation, using the dCor distance correlation method to screen key meteorological variables and combining Granger causality tests to establish causal relationships between meteorological factors and wind power, photovoltaics, and load curves. This makes the generated extreme scenarios more interpretable and credible.

[0046] This invention utilizes WGAN-GP with gradient penalty to augment the small samples of extreme scenarios obtained from clustering, achieving data augmentation and solving the problem of limited sample quantity for extreme weather scenarios, thus providing a sufficient data foundation for subsequent modeling. The second step, CGAN, uses scenario categories as conditional labels to achieve conditional generation of different extreme weather scenarios, ensuring the diversity and relevance of scenario generation.

[0047] This invention generates a coupled scenario of the three elements through joint modeling, which can reflect the dynamic matching relationship between renewable energy output and load demand under different extreme weather conditions. While improving the coverage and reliability of extreme situations, this invention supports the extreme scenario simulation of new power systems and integrates into the existing renewable energy scenario analysis and planning process.

[0048] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0049] In the description of this invention, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0050] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An enhanced GAN method for generating extreme scenes that considers weather factors, characterized in that, Includes the following steps: Based on meteorological factors that have a significant impact on extreme weather, and based on the impact of the extreme weather on the operation and planning of the power system, the causal relationship between meteorological time series and power time series is verified. Based on causal relationships, time series clustering and scene classification are performed on multidimensional meteorological factors, extreme sample sets are constructed, and sample augmentation processing is carried out. The enhanced extreme sample set is used as input to a conditional generative adversarial network to finally generate extreme scenarios by coupling wind, light, and load.

2. The method for generating extreme scenes using an enhanced GAN that considers weather factors according to claim 1, characterized in that: When extracting meteorological factors, based on historical wind load and meteorological data, the dCor distance correlation ranking is used to extract features of meteorological shadows that have a significant impact on extreme weather.

3. The method for generating extreme scenes using an enhanced GAN that considers weather factors according to claim 2, characterized in that: When verifying the causal relationship between meteorological time series and power time series, the Granger causality test is used to examine the causal relationship between the meteorological time series and power time series.

4. The enhanced GAN extreme scene generation method considering weather factors according to claim 3, characterized in that: When performing time series clustering, the K-multidimensional time series clustering algorithm KMDTSC is used to cluster the above-mentioned multidimensional meteorological factors in time series.

5. The method for generating extreme scenes using an enhanced GAN that considers weather factors according to claim 4, characterized in that: When classifying scenes, the elbow method is used to observe the curve of the sum of squared errors within a cluster (SSE) as a function of the number of clusters (K). When the decrease in SSE becomes insignificant as K is further increased, this point is taken as the optimal number of clusters (K). Based on the causal relationship that passes the Granger causality test, extreme scene clusters are determined for scene classification.

6. The method for generating extreme scenes using an enhanced GAN that considers weather factors according to claim 5, characterized in that: When performing sample augmentation, the Generative Adversarial Network WGAN-GP with gradient penalty is used to augment the extreme samples in the cluster.

7. The method for generating extreme scenes using an enhanced GAN that considers weather factors according to claim 6, characterized in that: When generating extreme scenarios by coupling wind, light, and load on the samples, the enhanced extreme scenario sample set is used as the input of the conditional generative adversarial network CGAN to generate the extreme scenarios by coupling wind, light, and load on the samples.

8. An enhanced GAN extreme scene generation system that considers weather factors, used to execute the enhanced GAN extreme scene generation method that considers weather factors as described in claim 1, characterized in that, include: The relationship analysis module is used to verify the causal relationship between meteorological time series and power time series based on meteorological factors that have a significant impact on extreme weather and the impact of the extreme weather on the operation and planning of the power system. The sample extraction and enhancement module is used to perform time series clustering and scene classification of multidimensional meteorological factors based on causal relationships, construct extreme sample sets, and perform sample enhancement processing. The scene generation module is used to take the enhanced extreme sample set as input to the conditional generative adversarial network and finally generate extreme scenes by coupling wind, light and load on the samples.

Images