Dispersed heterogeneous new energy output typical scene extraction method, system, device and medium

Abstract

The invention discloses a scattered heterogeneous new energy output typical scene extraction method, system, device and medium, and the method comprises the steps: carrying out the preprocessing of a historical data set in a power system, carrying out the dimension reduction of the preprocessed historical data set through a deep belief network, and obtaining a low-dimensional feature data set; constructing a generative adversarial network model based on the low-dimensional feature data set, wherein the generative adversarial network model generates a large-scale wind-light-water scene set by setting condition tags; and establishing an optimization model by taking the maximum scene information retention degree as a target, solving the optimization model by adopting a heuristic iterative algorithm, performing scene reduction on the large-scale wind-light-water scene set, and outputting a typical operation scene set. The method solves the problems that a large-scale wind-light-water scene set generated by an existing scene extraction method is difficult to cover multiple features, and an optimization model is not established with the maximum scene information retention degree as the target during scene reduction, so that the retained scene lacks key output features.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of new energy output typical scene extraction, and in particular relates to a dispersed heterogeneous new energy output typical scene extraction method, system, device and medium. BACKGROUND

[0002] With the development of dispersed heterogeneous new energy industry, the grid-connected proportion of wind, light and water new energy in the power system gradually increases. The output characteristics of such new energy are obviously volatile and intermittent due to natural conditions constraints. The extraction of typical output scenes is a key link to support power system planning and design, operation and dispatching, which requires generating a scene set that meets the actual output characteristics based on historical operation data, and extracting typical scenes to adapt to the actual needs of power system for new energy output characteristic analysis.

[0003] The existing typical scene extraction method fails to process high-dimensional data through effective dimension reduction when processing the historical data set of new energy, which increases the computational complexity of the subsequent scene generation process. At the same time, when generating a large-scale wind, light and water scene set, it fails to guide the combination of condition labels, which makes it difficult for the generated large-scale wind, light and water scene set to cover multiple characteristics. When reducing scenes, an optimization model is not established to maximize the scene information retention, resulting in the retained scenes missing key output characteristics. These problems are interrelated, so that the typical scenes obtained by the existing method cannot match the actual output characteristics of dispersed heterogeneous new energy, and cannot meet the precision and efficiency requirements of power system for scene extraction. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present application provides a dispersed heterogeneous new energy output typical scene extraction method, system, device and medium to solve the problem that the large-scale wind, light and water scene set generated by the existing scene extraction method is difficult to cover multiple characteristics, and when reducing scenes, an optimization model is not established to maximize the scene information retention, resulting in the retained scenes missing key output characteristics.

[0006] To solve the above technical problems, the present application provides the following technical solutions: In a first aspect, the present invention provides a method for extracting typical scenarios of distributed heterogeneous new energy output, comprising: preprocessing historical datasets in a power system and reducing the dimensionality of the preprocessed historical datasets using a deep belief network to obtain a low-dimensional feature dataset; constructing a generative adversarial network model based on the low-dimensional feature dataset, wherein the generative adversarial network model generates a large-scale set of wind, solar and hydropower scenarios by setting conditional labels; establishing an optimization model with the objective of maximizing the retention of scenario information, and solving the optimization model using a heuristic iterative algorithm to reduce the number of scenarios in the large-scale set of wind, solar and hydropower scenarios, and outputting a set of typical operating scenarios.

[0007] As a preferred embodiment of the method for extracting typical scenarios of distributed heterogeneous new energy power output described in this invention, the step of reducing the dimensionality of the preprocessed historical dataset using a deep belief network includes: constructing a deep belief network composed of multiple layers of stacked restricted Boltzmann machines, and inputting the preprocessed historical dataset into the deep belief network; performing bottom-up layer-by-layer pre-training of the deep belief network using unsupervised learning, using the output of the lower layer restricted Boltzmann machine as the input of the upper layer restricted Boltzmann machine, and initializing the parameters of the deep belief network by passing them layer by layer; and performing global fine-tuning based on the pre-trained deep belief network using supervised learning, mapping the preprocessed historical dataset to a low-dimensional feature space to obtain the low-dimensional feature dataset.

[0008] The beneficial effects of this preferred technical solution are as follows: A deep belief network constructed from multiple layers of stacked Restricted Boltzmann Machines (RBMs) can capture complex features in the preprocessed historical dataset layer by layer through a multi-layered structure. Unsupervised learning is used for bottom-up, layer-by-layer pre-training, with the output of the lower-layer RBM serving as the input to the upper layer. This ensures that the initialization process of each layer's parameters aligns with the hierarchical feature distribution of the data. Through layer-by-layer propagation, the deep belief network parameters are initially adapted to the data characteristics. Based on the pre-trained deep belief network, supervised learning is used for global fine-tuning, further optimizing the network's mapping relationship. This allows the preprocessed historical dataset to be accurately mapped to a low-dimensional feature space. The resulting low-dimensional feature dataset retains the key output features of the original data while reducing the data dimensionality. The entire process, through the synergy of multi-layer stacking, layer-by-layer pre-training, and global fine-tuning, allows the dimensionality reduction operation to directly correspond to the feature structure of the historical dataset, achieving effective compression of data dimensionality and preservation of key information.

[0009] As a preferred embodiment of the method for extracting typical scenarios of distributed heterogeneous new energy power output described in this invention, the step of constructing a generative adversarial network model based on the low-dimensional feature dataset includes: constructing a generator network and a discriminator network respectively; outputting generated samples through the generator network; and inputting the generated samples and real samples from the low-dimensional feature dataset into the discriminator network for discrimination; establishing an adversarial training mechanism based on Wasserstein distance; calculating the loss functions of the generator network and the discriminator network through the adversarial training mechanism; and iteratively training the generator network and the discriminator network based on the loss functions to complete the construction of the generative adversarial network model.

[0010] The beneficial effects of this preferred technical solution are as follows: A generator network and a discriminator network are constructed separately. The generator network outputs generated samples, which are then input into the discriminator network along with real samples from the low-dimensional feature dataset for discrimination. This process allows for direct comparison between the generated samples and real samples, providing a clear basis for the training of the generative adversarial network (GAN) model. Establishing an adversarial training mechanism based on Wasserstein distance enables more stable calculation of the loss functions of the generator and discriminator networks, avoiding training fluctuations that may occur in traditional adversarial training models. Iteratively training the generator and discriminator networks based on the calculated loss functions allows the generated samples output by the generator network to gradually approximate the new energy output characteristics carried by the real samples. Simultaneously, the discriminative ability of the discriminator network is optimized synchronously during the training process. The entire process, through the adversarial and collaborative interaction between the generator and discriminator, completes the construction of the GAN model, enabling it to directly correspond to the feature distribution in the low-dimensional feature dataset.

[0011] As a preferred embodiment of the method for extracting typical scenarios of distributed heterogeneous new energy output described in this invention, the step of generating a large-scale set of wind, solar and water scenes by setting conditional labels in the generative adversarial network model includes: classifying real samples in the low-dimensional feature dataset according to preset feature criteria to obtain conditional labels; and inputting each of the conditional labels into the generative adversarial network model in sequence to obtain a large-scale set of wind, solar and water scenes.

[0012] The beneficial effects of this preferred technical solution are as follows: Based on preset feature standards, real samples in a low-dimensional feature dataset are classified to obtain conditional labels. This classification process allows different new energy output features carried by real samples to correspond to specific label types, enabling conditional labels to directly associate with feature categories in the low-dimensional feature dataset. Each conditional label is sequentially input into a generative adversarial network (GAN) model, which generates corresponding scene samples based on the feature category corresponding to each conditional label. These scene samples fit the output features associated with the corresponding labels. The sequential input of multiple conditional labels allows the generated scenes to cover new energy output under different feature standards. The resulting large-scale wind, solar, and hydropower scene set encompasses output scenes with multiple feature categories. The entire process, through the combination of conditional labels and the GAN model, allows the generated scenes to directly correspond to different output types under preset feature standards, enabling the scene set to present diverse output characteristics of dispersed and heterogeneous new energy sources, thus achieving matching between scene generation and the feature categories of real samples.

[0013] As a preferred embodiment of the method for extracting typical scenarios of distributed heterogeneous new energy power output described in this invention, the step of establishing an optimization model with the objective of maximizing scenario information retention includes: defining an original scenario set and a retained scenario set respectively, and constructing an optimization model with the objective of maximizing the similarity between the original scenario set and the retained scenario set, wherein the optimization model is expressed as: ; ; In the formula, R is the set of scenes to be retained, and S is the original set of scenes. To preserve the scene set The scene is the first one in the original scene set. Scene information retention rate of each scene and These represent the number of scenes in the original scene set and the number of scenes in the retained scene set, respectively. and Scenes and scene The probability, For scene probability similarity, and These are peak power retention and valley power retention, respectively. To ensure the retention rate of monthly electricity consumption information, For sufficiently small positive numbers, These are constraints.

[0014] The beneficial effects of this preferred technical solution are as follows: Defining the original scene set and the retained scene set separately allows the scene reduction process to clearly anchor the characteristic basis of the original scene. An optimization model is constructed with the goal of maximizing the similarity between the two. The scene information retention in the optimization model is achieved through a combination of scene probability similarity, peak power retention, valley power retention, and monthly electricity information retention. This simultaneously covers multi-dimensional features such as the probability distribution of the scene, power extremes, and electricity information. This enables the optimization model's similarity measurement to correspond to the core characteristics of dispersed and heterogeneous new energy output scenarios. The probabilities in the optimization model are respectively associated with the proportions of the original and retained scenes. Combined with constraints, the probability distribution of the retained scene set conforms to the basic logic of scene extraction. The entire optimization model construction process, through the integration of multi-dimensional retention and the combination of probability constraints, allows the optimization model to directly correspond to the feature distribution of the original scene set, achieving a scene selection criterion with the goal of maximizing scene information retention. This ensures that the optimization direction of the optimization model aligns with the information integrity requirements of scene extraction.

[0015] As a preferred embodiment of the method for extracting typical scenarios of distributed heterogeneous new energy power output described in this invention, the values ​​of the scenario probability similarity, peak power retention, valley power retention, and monthly electricity information retention are all within the range of [0,1]; wherein, the calculation method of the scenario probability similarity is expressed as follows: ; In the formula, and Scenes and scene The One element, For the scene dimension; wherein, the calculation method for the peak power retention is expressed as: ; In the formula, and Scenes and scene The Middle The venue is at the first Heavenly Power at any moment The number of power stations; wherein, the calculation method for the valley power retention rate is expressed as: ; In the formula, and All are sufficiently small positive numbers; wherein, the calculation method for the monthly electricity information retention rate is as follows: ; In the formula, and Scenes and scene The Middle The first station Power at any moment This refers to the scene's temporal resolution.

[0016] The beneficial effects of this preferred technical solution are as follows: the values ​​of scene probability similarity, peak power retention, valley power retention, and monthly electricity information retention are all within the range of [0,1], ensuring a uniform measurement scale for these indicators. Scene probability similarity is achieved through scene... and scene The calculation of differences among the elements in the data can correspond to the distribution matching between two scenarios. Peak power retention is calculated based on the maximum power of each station in the scenario at different times on different dates, corresponding to the characteristics of peak renewable energy output. Valley power retention is calculated based on the minimum power, while introducing sufficiently small positive numbers to avoid calculation anomalies, corresponding to the characteristics of valley renewable energy output. Monthly electricity information retention is calculated by integrating the power of each station at different times in the scenario with time resolution, corresponding to the electricity dimension characteristics of renewable energy output. These calculation methods are associated with different characteristics of the scenarios, allowing each indicator to correspond to a dimension of dispersed and heterogeneous renewable energy output, so that the measurement of scenario information retention can cover these multiple characteristics. The whole process combines targeted calculation of each indicator with a unified value range, so that the information matching between scenarios can directly correspond to the output characteristics of different dimensions, realizing the quantification of information retention.

[0017] As a preferred embodiment of the method for extracting typical scenarios of distributed heterogeneous new energy power output described in this invention, the step of solving the optimization model using a heuristic iterative algorithm includes: taking the large-scale wind, solar and water scene set as the original scene set, and taking all scenes in the large-scale wind, solar and water scene set as the initial value of the retained scene set; based on the optimization model, calculating the sum of information retention of each scene in the retained scene set relative to all scenes in the original scene set, removing the scene with the lowest sum of information retention from the retained scene set, and updating the probability distribution of the remaining scenes in the retained scene set; until the number of scenes in the retained scene set is reduced to a preset target number, and outputting the retained scene set as the typical operating scene set.

[0018] The beneficial effects of this preferred technical solution are as follows: A large-scale collection of landscape and water scenes is used as the original scene set, and all scenes are used as the initial value of the retained scene set. This ensures that the retained scene set covers the complete features of the original scenes from the outset. Based on the optimization model, the sum of information retention of each retained scene relative to the original scene set is calculated. This accurately identifies the scenes that contribute the least to the overall information and removes them from the retained scene set. This reduces the number of scenes while maximizing the retention of the core output features of the original scenes. After removal, the probability distribution of the remaining scenes in the retained scene set is updated so that the probability proportion of the remaining scenes corresponds to the distribution characteristics of the original scene set, avoiding distortion of probability features due to scene removal. The entire iterative process is repeated according to this logic until the number of retained scene sets is reduced to the preset target number. The output retained scene set serves as a typical operating scene set. This not only achieves effective simplification of scene scale but also, through the measurement of the sum of information retention and probability updates, allows the typical scene set to directly correspond to the multi-dimensional output features of the original scene set, making the scene reduction process highly consistent with the goals of the optimization model.

[0019] Secondly, the present invention provides a typical scenario extraction system for distributed heterogeneous new energy power output, comprising: Data preprocessing and dimensionality reduction module: used to preprocess historical datasets and reduce the dimensionality of the preprocessed data through a deep belief network to obtain a low-dimensional feature dataset; Conditional Scene Generation Module: Used to generate a large-scale collection of landscape and water scenes by constructing a generative adversarial network model based on the low-dimensional feature dataset and conditional labels. Scene Reduction and Optimization Module: Used to build an optimization model with the goal of maximizing the retention of scene information; Heuristic Iterative Reduction Module: Used to execute reduction strategies, iteratively calculating and eliminating scenarios that contribute the least to the overall information retention until a preset number of typical running scenarios are obtained; Typical Scene Output Module: Used to output the final set of typical running scenes after reduction when the number of retained scenes reaches the preset target.

[0020] Thirdly, the present invention provides an electronic device, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the method for extracting typical scenarios of distributed heterogeneous new energy power output are implemented.

[0021] Fourthly, the present invention provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the steps of the method for extracting typical scenarios of distributed heterogeneous new energy power output.

[0022] Compared with existing technologies, the beneficial effects of this invention are as follows: After preprocessing the historical dataset of the power system, a deep belief network with multiple layers of restricted Boltzmann machines is used to achieve dimensionality reduction through layer-by-layer pre-training and global fine-tuning. While compressing the data dimension, the key output features in the original data are completely preserved, avoiding redundant calculations in the scene generation stage for high-dimensional data. The generative adversarial network model built based on the low-dimensional feature dataset stabilizes the performance of the generative adversarial network model through the adversarial training mechanism of Wasserstein distance. Combined with the conditional labels obtained by classification according to preset feature standards, the generated large-scale wind, solar and hydropower scene set can cover the output of new energy sources with different feature categories, presenting the diverse output characteristics of dispersed and heterogeneous new energy sources. In the scene reduction stage, an optimization model is established with the goal of maximizing the retention of scene information. Multi-dimensional information such as scene probability, peak and valley power and monthly electricity consumption are integrated to measure the feature matching degree. Then, a heuristic iterative algorithm is used to gradually eliminate scenes with low information contribution and update the probability distribution of the remaining scenes simultaneously. The final typical operating scene set is streamlined in scale while completely preserving the key output features in the original scene. The entire process ensures that the extracted typical scenarios not only conform to the actual output patterns of distributed heterogeneous new energy sources, but also meet the actual needs of power system planning, design, operation, and scheduling for scenario accuracy and efficiency, thus achieving coordinated optimization of data processing, scenario generation, and extraction. Attached Figure Description

[0023] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. 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.

[0024] Figure 1 This is a schematic diagram of the overall process of the method for extracting typical scenarios of distributed heterogeneous new energy power output according to an embodiment of the present invention. Detailed Implementation

[0025] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0026] Example 1, referring to Figure 1 As an embodiment of the present invention, a method for extracting typical scenarios of distributed heterogeneous new energy power output is provided, including steps S100 to S300: S100. Preprocess the historical dataset in the power system, and then reduce the dimensionality of the preprocessed historical dataset using a deep belief network to obtain a low-dimensional feature dataset.

[0027] S200: A generative adversarial network model is constructed based on a low-dimensional feature dataset. The generative adversarial network model generates a large-scale collection of landscape and water scenes by setting conditional labels.

[0028] S300: An optimization model is established with the goal of maximizing the retention of scene information. A heuristic iterative algorithm is used to solve the optimization model. Scene reduction is performed on a large-scale set of landscape and water scenes, and a typical set of operating scenes is output.

[0029] It should be noted that renewable energy generation is characterized by intermittency, volatility, and uncertainty. Wind power output fluctuates dramatically due to wind speed, photovoltaic power generation is significantly constrained by sunlight and weather conditions, and hydropower output is affected by water inflow and dispatch strategies. In power system planning and operation analysis, a large number of possible renewable energy output scenarios need to be considered to comprehensively assess the system's safety and economy. However, traditional scenario generation methods rely on prior probability assumptions, making it difficult to accurately characterize the output characteristics and spatiotemporal coupling relationships of heterogeneous renewable energy sources such as wind, solar, and hydropower. Furthermore, an excessive number of scenarios can drastically increase the computational scale of power system optimization and dispatch analyses, reducing solution efficiency or even making solutions impossible. In addition, traditional scenario reduction methods often focus on mathematical distance metrics, failing to fully integrate the probability distribution and physical characteristics of scenarios. The reduced scenario set lacks representativeness, affecting the accuracy of the analysis results. Therefore, the extraction and reduction of typical renewable energy output scenarios are also crucial.

[0030] Therefore, to address the aforementioned issues of insufficient accuracy in scene generation and lack of representativeness in scene reduction, steps S100-S300 are used to preprocess historical datasets and reduce dimensionality using a deep belief network, effectively compressing data dimensions and extracting key features. A conditional generative adversarial network model based on Wasserstein distance is constructed, and combined with conditional labels, a large-scale set of wind, solar, and hydropower scenes conforming to physical laws is generated, accurately characterizing the power output characteristics under different conditions. An optimization model aimed at maximizing information retention is established, and scene reduction is achieved through a heuristic iterative algorithm. While retaining key features such as scene probability distribution, peak and valley power, and electricity information, the large-scale scene set is simplified into a set of typical operating scenes, achieving an accurate characterization of the uncertainty of heterogeneous new energy output from wind, solar, and hydropower.

[0031] Example 2, refer to Figure 1 This is one embodiment of the present invention. Based on the above embodiment, a method for extracting typical scenarios of distributed heterogeneous new energy power output is provided.

[0032] In this embodiment of the application, the step of reducing the dimensionality of the preprocessed historical dataset using a deep belief network in step S100 includes A1~A3: A1. Construct a deep belief network consisting of multiple layers of restricted Boltzmann machines stacked together, and input the preprocessed historical dataset into the deep belief network; Because historical datasets have high dimensionality, direct use leads to low computational efficiency, feature redundancy, and increased risk of overfitting. Therefore, dimensionality reduction is necessary for the preprocessed historical datasets. In this embodiment, a deep belief network (DPN) is used for dimensionality reduction. The DPN consists of several layers of restricted Boltzmann machines stacked together, which can reduce data dimensionality by extracting abstract features from the data layer by layer. In the process of constructing the DPN, the number of nodes in the input layer of the DPN is first set to match the number of data features according to the actual feature dimensionality of the preprocessed historical dataset, and several hidden layers are set, with the number of nodes in each layer decreasing sequentially to achieve a gradual mapping from high dimension to low dimension. After the DPN is constructed, the preprocessed historical dataset is input into the DPN to prepare for the training phase.

[0033] A2. Unsupervised learning is used to perform bottom-up layer-by-layer pre-training of the deep belief network. The output of the lower layer restricted Boltzmann machine is used as the input of the upper layer restricted Boltzmann machine. The parameters of the deep belief network are initialized by passing them layer by layer. In this embodiment, the training of the deep belief network is divided into two stages: pre-training and fine-tuning. The first stage is pre-training, which aims to initialize the parameters of the deep belief network using unsupervised learning, avoiding training difficulties caused by random initialization. Specifically, the preprocessed historical dataset is input into the bottom-level Restricted Boltzmann Machine (RBM). The bottom-level RBM trains its connection weights and bias parameters using an unsupervised learning algorithm, learning the feature representation of the input data. After training, the output of this layer is used as the input of the next RBM to continue training the parameters of the previous layer. This process is repeated layer by layer from bottom to top, with each layer trained based on the features learned from the previous layer, until all layers have completed pre-training. Through this layer-by-layer pre-training method, the low-level features of the original data are gradually combined into more compact high-level features, while simultaneously initializing the parameters of the entire deep belief network, narrowing the parameter optimization space to a more optimal region.

[0034] A3. Based on the pre-trained deep belief network, a supervised learning approach is used for global fine-tuning, mapping the pre-processed historical dataset to a low-dimensional feature space to obtain a low-dimensional feature dataset.

[0035] After pre-training, the fine-tuning phase begins. Since unsupervised learning can only perform a rough initialization of parameters and cannot optimize the parameters of the entire deep belief network, supervised learning is required for global fine-tuning. Specifically, based on the pre-trained deep belief network, a traditional global learning algorithm is used, combined with the pre-processed historical dataset and its own label information, to further optimize and adjust the parameters of each layer in the deep belief network from top to bottom. During the fine-tuning process, the deep belief network can finely adjust the parameters according to the actual task requirements while retaining the abstract features learned in the pre-training phase, making the feature extraction capability of the deep belief network more suitable for the dimensionality reduction task. After multiple rounds of iterative adjustments, the pre-processed historical dataset is input again into the trained deep belief network. The deep belief network maps the input historical dataset to a low-dimensional feature space, and the output is the low-dimensional feature dataset.

[0036] In an optional implementation, the dimensionality reduction of the preprocessed historical dataset using a deep belief network in step S100 can also be achieved using a staged dimensionality reduction strategy. Specifically, considering the significant differences in feature dimensions between different types of new energy data (wind power, photovoltaic, hydropower) and meteorological parameters, when constructing the deep belief network, the preprocessed historical dataset is first grouped according to data type. For each group of data, a corresponding restricted Boltzmann machine sub-network is constructed, and the number of hidden layer nodes in each sub-network is independently set according to the initial dimensionality and redundancy of the data group. During the pre-training stage, each sub-network performs unsupervised learning in parallel and completes its own parameter initialization. After pre-training, the intermediate layer features extracted by each sub-network are concatenated to form a fused feature vector. An additional restricted Boltzmann machine layer is then constructed to perform secondary dimensionality reduction on the fused features. Finally, supervised learning is used to perform global fine-tuning of the entire network.

[0037] In this embodiment of the application, the step of constructing a generative adversarial network model from the low-dimensional feature dataset in step S200 includes B1~B2: B1. Construct a generator network and a discriminator network respectively. The generator network outputs generated samples, and the generated samples and real samples in the low-dimensional feature dataset are input into the discriminator network for discrimination. In this embodiment, the Generative Adversarial Network (GAN) model consists of a generator network and a discriminator network. Scene generation capability is achieved through adversarial training between the generator and discriminator networks. Specifically, the generator network employs a multi-layer neural network structure, receiving a Gaussian-distributed noise vector as input. After upsampling, it outputs generated samples. The output dimension of the generator network is consistent with the dimension of the real samples in the low-dimensional feature dataset to ensure comparability between generated and real samples in the feature space. The discriminator network also employs a multi-layer neural network structure. Its input includes two types of data: generated samples output by the generator network and real samples extracted from the low-dimensional feature dataset. The discriminator network performs feature extraction and analysis on the input sample data through multi-layer downsampling operations, ultimately outputting a continuous value to measure the probability that the input sample data belongs to the true distribution. In the early stages of training, the quality of the generated samples output by the generator network is relatively low, and the discriminator network can easily distinguish between the generated samples and real samples. As training progresses, the generator network continuously optimizes to improve the realism of the generated samples, and the discriminator network also enhances its discrimination ability, forming a game-theoretic relationship between the two.

[0038] For example, the generator network uses a 4-layer fully connected neural network structure, with the number of nodes in each layer being 100, 256, 512, and the dimension of the low-dimensional feature dataset, respectively. ReLU activation is used between each layer, and the last layer uses Tanh activation to constrain the output within a reasonable range. The discriminator network also uses a 4-layer fully connected structure, with the number of nodes in each layer reversed compared to the generator network: the dimension of the low-dimensional feature dataset, 512, 256, and 1, respectively. The first three layers use LeakyReLU activation, and the last layer outputs continuous values ​​for discrimination. During training, 64 real samples are randomly selected from the low-dimensional feature dataset as a batch, and the generator simultaneously generates 64 generated samples. Both are then input into the discriminator for discrimination.

[0039] B2. Establish an adversarial training mechanism based on Wasserstein distance, calculate the loss functions of the generator network and the discriminator network through the adversarial training mechanism, and iteratively train the generator network and the discriminator network based on the loss functions to complete the construction of the generative adversarial network model. To improve the training stability and generation quality of the generative adversarial network (GAN) model, this embodiment establishes an adversarial training mechanism based on Wasserstein distance. Specifically, in the adversarial training mechanism, a gradient penalty mechanism is introduced to enhance the Lipschitz continuity of the discriminator network. The gradient penalty is achieved by randomly interpolating between real and generated samples, calculating the gradient norm of the discriminator network output at the interpolation point, and imposing a constraint on the gradient norm to make it approach 1, thereby stabilizing the training process. Then, based on the discrimination results of the discriminator network, the loss functions of the generator network and the discriminator network are calculated separately. The loss function of the generator network is used to minimize the discriminator's ability to distinguish generated samples, making the generated samples closer to the distribution of real samples, while the loss function of the discriminator network is used to maximize the ability to distinguish between real and generated samples, and includes a gradient penalty term to ensure training stability. During the training process, the parameters of the generator network and the discriminator network are updated alternately. Through multiple rounds of iterative training, the distribution of generated samples output by the generator network gradually approximates the distribution of real samples. When the training reaches convergence, the construction of the GAN model is completed.

[0040] For example, during training, the loss function of the discriminator network is the difference between the expected value of the real sample discrimination and the expected value of the generated sample discrimination, minus the gradient penalty term. The gradient penalty term is calculated as follows: in each batch, the real samples and generated samples are linearly interpolated with random weights to obtain interpolated samples. The L2 norm of the gradient of the discriminator network at the interpolated sample is calculated, and the square of the difference between the L2 norm of the gradient at the interpolated sample and 1 is multiplied by the penalty coefficient 10. The loss function of the generator network is the negative of the discriminator's expected value of the generated sample discrimination. In each iteration, the discriminator network parameters are updated 5 times, each time using data from a different batch, and then the generator network parameters are updated once. This unbalanced update strategy ensures that the discriminator always maintains a strong discrimination ability. The learning rate is set to 0.0001, and the Adam optimizer is used for parameter updates. After about 1000 iterations of training, the generator loss and discriminator loss tend to stabilize, at which point the construction of the generative adversarial network model is completed.

[0041] In an optional implementation, the construction of the generative adversarial network model from the low-dimensional feature dataset in step S200 can also be achieved using a conditional information-guided construction strategy. First, the real samples in the low-dimensional feature dataset are classified according to preset feature criteria, labeling them with different conditional categories, such as four categories based on season (spring, summer, autumn, winter) or three levels based on average daily power generation (high, medium, and low). At the input of the generator network, in addition to the noise vector, the corresponding conditional label encoding is input simultaneously, enabling the generator network to learn the scene generation rules under specific conditions. In the discriminator network, the conditional labels are concatenated with the input sample data, allowing the discriminator network to not only judge the authenticity of the sample data but also whether the sample data conforms to the feature distribution of the corresponding conditional label. When establishing the adversarial training mechanism, a conditional consistency constraint term is added to the loss function to ensure that the generated samples output by the generator network are not only authentic but also match the input conditional labels. Through this conditional information-guided construction method, the generative adversarial network model can simultaneously learn scene features under different conditions during training, and after model construction is completed, it can directly generate scene data with corresponding features by specifying conditional labels.

[0042] In this embodiment of the application, the step S200, in which the generative adversarial network model generates a large-scale set of landscape and water scenes by setting conditional labels, includes C1~C2: C1. Classify the real samples in the low-dimensional feature dataset according to the preset feature criteria to obtain conditional labels; First, pre-defined feature standards are determined based on actual needs. These standards can be key indicators reflecting the characteristics of new energy output, such as the season, weather type, and daily average power generation range of the real samples. After determining the feature standards, each real sample in the low-dimensional feature dataset is analyzed to extract its corresponding feature attribute values. For example, if the season is used as the feature standard, the season is determined based on the time information of the real sample; if the daily average power generation range is used as the feature standard, the daily average power generation of the real sample is calculated and the range is determined. Based on the extracted feature attribute values, the real samples are divided into several categories, each corresponding to a condition label. The condition label can be represented using numerical encoding or one-hot encoding. This classification method establishes a mapping relationship between the real samples in the low-dimensional feature dataset and their corresponding condition labels.

[0043] For example, by selecting season as the feature criterion, the samples are divided into four categories—spring, summer, autumn, and winter—based on the time information of the original historical data corresponding to each real sample in the low-dimensional feature dataset, corresponding to condition labels 1, 2, 3, and 4, respectively. To more accurately characterize the power output characteristics of different time periods, a second feature criterion can be introduced, such as dividing the average daily power generation into three levels: low, medium, and high, corresponding to condition labels A, B, and C, respectively. Specifically, for each real sample in the low-dimensional feature dataset, the power time series in the corresponding preprocessed historical data is traced back, and the average daily power generation of the power time series is calculated. The distribution of the average daily power generation of all real samples is statistically analyzed. Using the 33rd and 67th percentiles as dividing points, samples with average daily power generation below the 33rd percentile are labeled as low (A), those between the 33rd and 67th percentiles are labeled as medium (B), and those above the 67th percentile are labeled as high (C). Through the combination of the two feature criteria, 12 condition labels can be obtained (such as spring-low, spring-medium, spring-high, etc.).

[0044] C2. Input each condition label into the generative adversarial network model in sequence to obtain a large-scale set of landscape and water scenes; Each condition label and a randomly generated noise vector are jointly input into the generator network of the generative adversarial network model. The generator network outputs wind, solar and water scene data that conforms to the features corresponding to the current condition label based on the input condition label and noise vector. Since the noise vector is different each time, even if the condition label is the same, the generator network can output different scene data with a certain degree of randomness, thus ensuring the diversity of generated scenes. After traversing all condition labels and generating a sufficient number of scene data for each condition label, all generated scene data are aggregated to form a large-scale wind, solar and water scene set. The large-scale wind, solar and water scene set covers a variety of scenes under different conditions and can comprehensively reflect the uncertainty and fluctuation characteristics of wind, solar and water power output.

[0045] For example, if there are 12 types of condition labels (such as the season-power generation combination in step C1), generating 1000 scene data for each condition label will generate a total of 12000 scene data, forming a preliminary large-scale wind, solar and water scene set. During the generation process, for the condition label "spring-low", 1000 noise vectors are randomly sampled from a Gaussian distribution, each noise vector having a dimension of 100. The condition label is encoded as a one-hot vector and concatenated with the noise vector, then input into the generator network to obtain 1000 scene data under the "spring-low" condition label. The same operation is performed on the other 11 condition labels in turn, finally obtaining a large-scale wind, solar and water scene set containing 12000 scenes. If a larger scene set is needed, the number of scenes generated for each condition label can be increased. For example, if 5000 scenes are generated for each condition label, the total scale reaches 60000.

[0046] In one optional implementation, the generative adversarial network model in step S200, which generates a large-scale wind, solar, and hydropower scene set by setting condition labels, can also employ a multi-stage iterative generation strategy. First, all condition labels are prioritized, with those appearing frequently or having a significant impact on power system operation set as high priority, such as condition labels corresponding to extreme weather conditions or high-load periods. In the first stage, for high-priority condition labels, a larger number of scene data are generated for each condition label to ensure sufficient coverage of these key scenarios. In the second stage, a suitable amount of scene data is generated for medium-priority condition labels. In the third stage, fewer scene data are generated for low-priority or infrequently appearing condition labels. This phased and differentiated generation strategy ensures the scale of the scene set while making the generated scene distribution more consistent with actual operating characteristics, focusing on typical operating conditions that have a significant impact on the power system. After completing each stage of generation, all scene data are aggregated to obtain a structurally sound large-scale wind, solar, and hydropower scene set.

[0047] In this embodiment of the application, step S300 establishes an optimization model with the goal of maximizing scene information retention. This is achieved by defining an original scene set and a retained scene set, and constructing an optimization model with the goal of maximizing the similarity between the original scene set and the retained scene set. The optimization model is expressed as follows: ; ; In the formula, R is the set of scenes to be retained, and S is the original set of scenes. To preserve the scene set The scene is the first one in the original scene set. Scene information retention rate of each scene and These represent the number of scenes in the original scene set and the number of scenes in the retained scene set, respectively. and Scenes and scene The probability, For scene probability similarity, and These are peak power retention and valley power retention, respectively. To ensure the retention rate of monthly electricity consumption information, For sufficiently small positive numbers, These are constraints; The values ​​of scene probability similarity, peak power retention, valley power retention, and monthly electricity information retention are all within the range of [0,1]. The method for calculating scene probability similarity is expressed as follows: ; In the formula, and Scenes and scene The One element, From the perspective of scene; The calculation method for peak power retention is expressed as follows: ; In the formula, and Scenes and scene The Middle The venue is at the first Heavenly Power at any moment Number of stations; The calculation method for valley power retention is expressed as follows: ; In the formula, and All are sufficiently small positive numbers; The calculation method for monthly electricity consumption information retention rate is as follows: ; In the formula, and Scenes and scene The Middle The first station Power at any moment For scene temporal resolution; In this embodiment, to accurately assess the degree to which the retained scene set preserves scene information from the original scene set, four key indicators need to be calculated: scene probability similarity, peak power retention, valley power retention, and monthly electricity consumption information retention. These indicators all range from [0,1], with smaller values ​​indicating higher retention. Specifically, scene probability similarity measures the closeness between the retained and original scenes in the feature space. The calculation method is as follows: first, calculate the sum of the squares of the differences between the feature vectors of the original and retained scenes to obtain the Euclidean distance; then, divide the product of the two scene probabilities by the sum of the probabilities as a weighting coefficient; multiply the weighting coefficient by the square root of the Euclidean distance; finally, subtract the product from 1 to obtain the scene probability similarity. Peak power retention measures the retention of the maximum power feature of the original scene by the retained scene. The calculation method is as follows: for each power station, the maximum power of the original and retained scenes on each day is statistically analyzed. Peak power retention is calculated as the ratio of the sum of the maximum power of the two scenarios, and the absolute value of this ratio is subtracted from 1 to obtain the peak power retention rate. Valley power retention is calculated similarly to peak power retention, but the minimum power for each day is used, and a sufficiently small positive number is added to avoid division by zero. Monthly electricity information retention measures the retention of the original scenario's total power generation by the retained scenario. It is calculated as follows: for each power station, the power at all times is multiplied by the time resolution and summed to obtain the total electricity generation. The ratio of the total electricity generation of the original scenario to the retained scenario is calculated, and the absolute value of this ratio is subtracted from 1 to obtain the monthly electricity information retention rate. The information retention rate of the retained scenario relative to the original scenario is obtained by summing the above four indicators with sufficiently small positive numbers, taking the reciprocal, and then multiplying by the probability of the original scenario. This multi-dimensional comprehensive evaluation method can fully characterize the degree to which the retained scenario retains the original scenario's information in terms of probability distribution, extreme value characteristics, and electricity characteristics.

[0048] For example, calculate a scene in the retained scene set R. Compared to a certain scene in the original scene set S When considering information retention, assume the probabilities of the two scenarios are respectively... and The scene dimension is 48 (representing 24 hours per day × 2 sites); first, the scene probability similarity is calculated by summing the squared differences of the 48 feature elements to obtain the squared Euclidean distance. Assuming the calculated result is 1.5, the Euclidean distance is 1.225, and the weighting coefficient is... The scene probability similarity is Assuming the peak power retention rate is 0.05, the valley power retention rate is 0.08, and the monthly electricity consumption information retention rate is 0.03, let... Then the information retention rate is Similarly, the total information retention rate of the retained scene is obtained by sequentially calculating all 12,000 scenes in the original scene set and summing them up.

[0049] In this embodiment of the application, step S300, which uses a heuristic iterative algorithm to solve the optimization model, includes steps D1 to D3: D1. Use the large-scale landscape and water scene set as the original scene set, and use all the scenes in the large-scale landscape and water scene set as the initial value of the retained scene set. In this embodiment, after establishing the optimization model, a heuristic iterative algorithm is needed to solve the optimization model and select a typical set of operating scenarios from the large-scale landscape and water scene set. The essence of the heuristic iterative algorithm is a gradual reduction process. By repeatedly identifying and eliminating redundant scenarios with low information retention, the most representative subset of scenarios is ultimately retained. Specifically, an initialization operation is first performed, defining the large-scale landscape and water scene set output by the generative adversarial network model as the initial value for both the original scenario set S and the retained scenario set R. The purpose of this initialization is that the original scenario set S remains unchanged throughout the reduction process, serving as the evaluation benchmark; while the retained scenario set R is continuously reduced during the iteration process, eventually converging into a typical set of operating scenarios. In the initial state, the original scenario set S and the retained scenario set R contain identical scenario data and probability distributions. The initial probability of each scenario is determined based on its condition label at the time of generation. If the number of scenarios corresponding to each condition label is equal, then all scenarios under the same condition label have the same initial probability, and the sum of the probabilities of all scenarios is 1.

[0050] For example, suppose a large-scale landscape and water scene collection contains 12,000 scenes, which come from 12 conditional labels, with 1,000 scenes under each conditional label. During initialization, these 12,000 scenes are defined as the original scene set S, and all 12,000 scenes are copied to the initial value of the reserved scene set R. For each scene in the original scene set S, the initial probability is... The initial probability distribution of the retained scene set R is completely consistent with that of the original scene set S. The preset target number of scenes is 100, that is, the heuristic iterative algorithm needs to reduce 12,000 scenes to 100. At this time, the retained scene set R is in an initial saturated state, containing all possible scenes, providing a complete candidate set for iterative reduction.

[0051] D2. Based on the optimization model, calculate the sum of information retention of each scene in the retained scene set relative to all scenes in the original scene set, remove the scene with the lowest sum of information retention from the retained scene set, and update the probability distribution of the remaining scenes in the retained scene set. After initialization, the iterative reduction phase begins. This phase is the core of the heuristic iterative algorithm, gradually reducing the size of the retained scene set through repeated calculations, identification, and elimination processes. Specifically, in each iteration, each scene in the retained scene set R is first evaluated, and the sum of information retention of that scene relative to all scenes in the original scene set S is calculated. According to the information retention calculation method in step S300, for the j-th scene in the retained scene set R, it is necessary to traverse all scenes in the original scene set S. For each scene, calculate the scene information retention rate of the retained scene relative to each original scene, and then... The information retention values ​​of each scene are summed to obtain the total information retention of the retained scene. Similarly, the information retention of each scene in the retained scene set R is calculated sequentially, and the total information retention of each scene is compared. The scene with the lowest total information retention is identified. The lower the total information retention, the less information the scene contributes to the original scene set, and the more redundant the scene is. The scene with the lowest identified total information retention is removed from the retained scene set R, reducing the size of the retained scene set by 1. After removing the scene, the sum of the probabilities of the remaining scenes in the retained scene set R is no longer equal to 1. Therefore, the probability of the remaining scenes needs to be normalized, that is, the probability of each remaining scene is divided by the sum of the probabilities of all remaining scenes to ensure that the sum of the normalized probabilities is equal to 1 again. After completing one round of iteration, the number of scenes in the retained scene set R is reduced by 1, the probability distribution is updated, and it is ready for the next round of iteration.

[0052] For example, suppose the current retained scene set R contains 200 scenes, and the original scene set S still contains 12,000 scenes. In this iteration, the first scene in the retained scene set R... This requires calculating the information retention relative to all 12,000 scenes in the original scene set S; according to the formula in S300, for each original scene... ,calculate 12,000 Values ​​are accumulated to obtain the scene. The total information retention rate is assumed to be 15.6. This calculation is then repeated for the remaining 199 retention scenarios, assuming scenario... The total information retention rate is the lowest, at only 8.3, so the scenario will be... Remove from the retained scene set R. Before removal, the sum of probabilities of the 200 scenes is 1; after removal, the sum of probabilities of the remaining 199 scenes becomes... Assuming If the sum of the remaining probabilities is 0.995, then the probability of each remaining scenario is normalized by dividing by 0.995, so that the sum of the probabilities of the 199 scenarios is equal to 1 again.

[0053] D3. Continue until the number of scenes in the retained scene set is reduced to the preset target number, then output the retained scene set as the typical running scene set; In this embodiment, the iterative reduction process is repeatedly executed according to step D2. In each iteration, the scenario with the lowest total information retention is eliminated, and the probability distribution of the remaining scenarios is updated. The termination condition of the iteration process is that the number of scenarios in the retained scenario set R is reduced to a preset target number. Specifically, before the heuristic iterative algorithm begins, a target number of scenarios is set according to actual needs. The target number should be as small as possible while ensuring the representativeness of the scenario set, so as to reduce the computational burden of subsequent power system analysis. After each iteration, it is checked whether the current number of scenarios in the retained scenario set R has reached the target number. If it has not yet reached the target number, the next iteration continues; if the target number has been reached, the iteration process is terminated, and the current retained scenario set R is output as the final typical operating scenario set.

[0054] For example, the target number of preset scenarios is 100. Starting with an initial 12,000 scenarios, after the first iteration, the number of retained scenarios becomes 11,999, after the second iteration it becomes 11,998, and so on. When the iteration reaches the 11,900th round, the number of scenarios in the retained scenario set R decreases to 100, reaching the preset target. At this point, the iteration terminates, and a typical running scenario set containing 100 scenarios is output. These 100 scenarios are the most representative scenarios selected through 11,900 iterations, including scenarios from different condition labels, such as 12 high power generation scenarios in spring, 8 medium power generation scenarios in summer, and 15 low power generation scenarios in autumn. The scenario distribution is determined based on information retention optimization.

[0055] In an optional implementation, the heuristic iterative algorithm used to solve the optimization model in step S300 can also be implemented using a phased adaptive reduction strategy. Specifically, the entire scene reduction process is divided into two stages: a coarse reduction stage and a fine reduction stage. The number of scenes to be reduced in each iteration is dynamically adjusted according to the current size of the retained scene set. In the coarse reduction stage, when the number of scenes in the retained scene set is much larger than the target number, a batch reduction method is adopted. In each iteration, multiple scenes with low total information retention are eliminated simultaneously to quickly reduce the scene size. Specifically, in each iteration, the total information retention of all retained scenes is calculated, sorted from smallest to largest, and the k lowest-ranked scenes are eliminated simultaneously. The value of k is dynamically determined according to the difference between the current number of scenes and the target number. For example, when the difference is greater than 1000, k is set to 10, and when the difference is between 500 and 1000, k is set to 5. When the number of retained scenes approaches twice the target number, the fine reduction phase begins. At this point, the k value is set to 1, and only one scene is removed per round to ensure that the scenes retained have the best information retention. After each round of reduction, the probability of the remaining scenes is normalized, and the process is repeated until the target number is reached.

[0056] In summary, after preprocessing the historical dataset of the power system, a deep belief network with multiple layers of restricted Boltzmann machines is used to achieve dimensionality reduction through layer-by-layer pre-training and global fine-tuning. This reduces the data dimensionality while fully preserving the key output features in the original data, avoiding redundant calculations in the scene generation stage for high-dimensional data. The generative adversarial network model built on the low-dimensional feature dataset stabilizes its performance through an adversarial training mechanism based on Wasserstein distance. Combined with conditional labels obtained by classification according to preset feature standards, the generated large-scale wind, solar, and hydropower scene set can cover the output of new energy sources with different feature categories, presenting the diverse output characteristics of dispersed and heterogeneous new energy sources. In the scene reduction stage, an optimization model is established with the goal of maximizing scene information retention. Multi-dimensional information such as scene probability, peak and valley power, and monthly electricity consumption are integrated to measure feature matching degree. Then, a heuristic iterative algorithm is used to gradually eliminate scenes with low information contribution and update the probability distribution of the remaining scenes simultaneously. The final typical operating scene set is streamlined in scale while fully preserving the key output features in the original scenes. The entire process ensures that the extracted typical scenarios not only conform to the actual output patterns of distributed heterogeneous new energy sources, but also meet the actual needs of power system planning, design, operation, and scheduling for scenario accuracy and efficiency, thus achieving coordinated optimization of data processing, scenario generation, and extraction.

[0057] Example 3 illustrates a schematic scheme for a method of extracting typical scenarios of distributed heterogeneous new energy power output. It should be noted that the technical solution of this system for extracting typical scenarios of distributed heterogeneous new energy power output belongs to the same concept as the technical solution of the method described above. Details not described in detail in this embodiment can be found in the description of the technical solution of the method described above.

[0058] This embodiment also provides a typical scenario extraction system for distributed heterogeneous new energy power output, including: Data preprocessing and dimensionality reduction module: used to preprocess historical datasets and reduce the dimensionality of the preprocessed data through a deep belief network to obtain a low-dimensional feature dataset; Conditional Scene Generation Module: Used to generate a large-scale collection of landscape and water scenes by building a generative adversarial network model based on a low-dimensional feature dataset and conditional labels; Scene Reduction and Optimization Module: Used to build an optimization model with the goal of maximizing the retention of scene information; Heuristic Iterative Reduction Module: Used to execute reduction strategies, iteratively calculating and eliminating scenarios that contribute the least to the overall information retention until a preset number of typical running scenarios are obtained; Typical Scene Output Module: Used to output the final set of typical running scenes after reduction when the number of retained scenes reaches the preset target.

[0059] This embodiment also provides an electronic device applicable to the extraction of typical scenarios for distributed heterogeneous new energy power output, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the method for extracting typical scenarios for distributed heterogeneous new energy power output as proposed in the above embodiment.

[0060] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the method for extracting typical scenarios of distributed heterogeneous new energy output proposed in the above embodiments.

[0061] The storage medium proposed in this embodiment and the method for extracting typical scenarios of distributed heterogeneous new energy output proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0062] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0063] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. 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 be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for extracting typical scenes of dispersed heterogeneous new energy output, characterized in that, The method comprises the following steps: Preprocessing a historical data set in a power system, and reducing dimensionality of the preprocessed historical data set by a deep belief network to obtain a low-dimensional feature data set; Constructing a generative adversarial network model based on the low-dimensional feature data set, the generative adversarial network model generating a large-scale wind-solar-hydro scene set by setting conditional labels; Establishing an optimization model with the maximum scene information retention degree as the target, and solving the optimization model by using a heuristic iterative algorithm to reduce scenes in the large-scale wind-solar-hydro scene set, and outputting a typical operation scene set. 2.The method of claim 1, wherein, The step of reducing dimensionality of the preprocessed historical data set by the deep belief network comprises: Constructing a deep belief network stacked by multiple layers of restricted Boltzmann machines, and inputting the preprocessed historical data set into the deep belief network; Performing bottom-up layer-by-layer pre-training on the deep belief network in an unsupervised learning manner, taking the output of a lower layer of restricted Boltzmann machines as the input of an upper layer of restricted Boltzmann machines, and completing initialization of parameters of the deep belief network through layer-by-layer transmission; Based on the pre-trained deep belief network, performing global fine-tuning in a supervised learning manner, mapping the preprocessed historical data set to a low-dimensional feature space, and obtaining the low-dimensional feature data set.

3. The method of claim 2, wherein the method comprises: The step of constructing the generative adversarial network model based on the low-dimensional feature data set comprises: Respectively constructing a generator network and a discriminator network, outputting generated samples through the generator network, inputting the generated samples and real samples in the low-dimensional feature data set into the discriminator network for discrimination, and establishing an adversarial training mechanism based on a Wasserstein distance, calculating loss functions of the generator network and the discriminator network through the adversarial training mechanism, and iteratively training the generator network and the discriminator network based on the loss functions to complete construction of the generative adversarial network model. The step of generating the large-scale wind-solar-hydro scene set by the generative adversarial network model by setting conditional labels comprises:

4. The method of claim 3, wherein the method comprises: Classifying real samples in the low-dimensional feature data set according to a preset feature standard to obtain conditional labels; Inputting the conditional labels into the generative adversarial network model in sequence to obtain the large-scale wind-solar-hydro scene set. The step of establishing the optimization model with the maximum scene information retention degree as the target comprises:

5. The method of claim 4, wherein the method comprises: Defining an original scene set and a retained scene set, and constructing an optimization model with the maximum similarity between the original scene set and the retained scene set as the target, the optimization model being represented as: The scene probability similarity, the peak power retention degree, the valley power retention degree and the monthly electricity quantity information retention degree all have a value range of [0, 1]; ； ； where R is the reserved scene set, S is the original scene set, is the scene information reservation degree of the scene in the reserved scene set relative to the scene in the original scene set, and are the scene numbers in the original scene set and the reserved scene set respectively, and are the probabilities of the scene and the scene respectively, is the scene probability similarity degree, and are the peak power reservation degree and the valley power reservation degree respectively, is the monthly electricity information reservation degree, is a sufficiently small positive number, is the constraint condition.

6. The method of claim 5, wherein the method comprises: The calculation method of the scene probability similarity is represented as: The calculation method of the peak power retention degree is represented as: ； In the formula, and Scenes and scene The One element, From the perspective of scene; The calculation method of the valley power retention degree is represented as: ； In the formula, and Scenes and scene The Middle The venue is at the first Heavenly Power at any moment Number of stations; The calculation method of the monthly electricity quantity information retention degree is: ； wherein and are each sufficiently small positive numbers; The step of solving the optimization model by using the heuristic iterative algorithm comprises: ； wherein and are the power at the time and the power at the time of the first of the first of the first is the scenario time resolution.

7. The method of claim 6, wherein the method comprises: Taking the large-scale wind-solar-hydro scene set as the original scene set, and taking all scenes in the large-scale wind-solar-hydro scene set as the initial value of the retained scene set. ​ Based on the optimization model, the information retention sum of each scene in the reserved scene set is calculated relative to all scenes in the original scene set, the scene with the lowest information retention sum is removed from the reserved scene set, and the probability distribution of the remaining scenes in the reserved scene set is updated; Until the number of scenes in the reserved scene set is reduced to the preset target number, the reserved scene set is output as the typical operation scene set.

8. A system for extracting typical scenarios of dispersed heterogeneous new energy output, applying the method of any one of claims 1-7, characterized in that, Comprise: Data preprocessing and dimensionality reduction module: for preprocessing the historical data set, and reducing the dimension of the preprocessed data through a deep belief network to obtain a low-dimensional feature data set; Conditional scene generation module: for constructing a generative adversarial network model based on the low-dimensional feature data set and based on a conditional label, thereby generating a large-scale wind, light and water scene set; Scene reduction optimization module: for constructing an optimization model with the maximum scene information retention as the target; Heuristic iterative reduction module: for executing a reduction strategy, and through iterative calculation, removing the scene with the smallest contribution to the overall information retention, until a preset number of typical operation scene sets are obtained; Typical scene output module: for outputting the final reduced typical operation scene set when the number of reserved scenes reaches the preset target. 9.An electronic device, comprising: a memory and a processor; The memory is used to store computer executable instructions, and the processor is used to execute the computer executable instructions, which realize the steps of the distributed heterogeneous new energy output typical scene extraction method in any one of claims 1 to 7 when executed by the processor. 10.A computer readable storage medium storing computer executable instructions, which realize the steps of the distributed heterogeneous new energy output typical scene extraction method in any one of claims 1 to 7 when executed by the processor.

Images