Electric power system extreme scene automatic extraction and feature quantification method, equipment and medium

Abstract

The invention relates to a power system extreme scene automatic extraction and feature quantification method and device based on an improved generative adversarial network, and a medium, and the method comprises the following steps: obtaining historical data, and recognizing and extracting an extreme scene from the historical data through an extreme scene recognition network, performing data enhancement on the extreme scene by adopting a data enhancement strategy to form an extreme scene sample set; training by using the extreme scene sample set to obtain an extreme scene generation model, wherein the improved GAN architecture of the extreme scene generation model comprises a condition generator introducing an attention mechanism and a multi-scale discriminator; and extracting and obtaining quantized four-dimensional features based on an extreme scene generated by the extreme scene generation model, constructing a training set based on the extreme scene generated by the extreme scene generation model, and training the extreme scene recognition network. Compared with the prior art, the method has the advantages that the precision is high, the model training problem under extreme data scarcity is solved, and extreme scene feature quantification is realized.

Description

Technical Field

[0001] This invention relates to the field of power system planning technology, and in particular to a method, device and medium for automatic extraction and feature quantification of extreme scenarios in power systems based on an improved generative adversarial network. Background Technology

[0002] As the proportion of intermittent and volatile renewable energy sources such as wind power and photovoltaics in the installed capacity and power generation of the power system continues to rise, the operating characteristics of the power system are rapidly shifting from the traditional "source follows load" to "source-grid-load-storage coordinated interaction." This shift significantly increases the system's dependence on the natural environment, and extreme weather events, such as strong typhoons, persistent cold waves, widespread high-temperature droughts, and torrential rains, are escalating their impact on the power system from localized and short-term effects to systemic risks that may trigger cascading failures. For example, strong typhoons may directly damage coastal wind turbine towers and transmission lines, persistent cold waves can lead to a surge in electricity demand and a sharp drop in photovoltaic module efficiency, and widespread droughts can affect hydropower output and thermal power cooling water supply. These extreme scenarios can disrupt the system's power balance and even cause power outages.

[0003] In the forward-looking work of power system planning, traditional planning models, mostly based on historical average meteorological data and conventional operating scenarios, are no longer sufficient to cover the extreme risks under the integration of a high proportion of renewable energy. Therefore, accurately identifying the core characteristics of extreme scenarios and assessing their potential losses has become a key prerequisite for improving the system's ability to cope with extreme events. This process not only provides data support for grid hardware hardening but also guides the formulation and drills of emergency plans, ultimately helping the power system operate stably under extreme conditions.

[0004] However, existing technologies typically employ data-driven extreme scenario extraction methods. For example, patent application CN116244887A discloses a novel method for generating extreme operating scenarios of a power system, which includes: obtaining empirical distribution functions of new energy sources and new loads; obtaining data-driven power flow empirical formulas; obtaining generalized transfer distribution factors based on the power flow empirical formulas; constructing a novel power system extreme scenario generation model based on the generalized transfer distribution factors, search direction, and active power; solving the novel power system extreme scenario generation model using a random search method to obtain an extreme scenario set; and checking and verifying the extreme scenario set based on the empirical distribution functions of new energy sources and new loads and the AC power flow formulas. However, existing technologies still have the following problems: 1. The problem of scarce extreme data: Extreme events occur infrequently in historical data, resulting in insufficient training data and making it difficult to train an effective recognition model. Traditional methods rely on a large amount of historical data, but data on extreme scenarios is extremely limited, which severely restricts model performance.

[0005] 2. Complex Scene Characteristics: Extreme scenarios are characterized by long duration, wide impact range, and abnormal power output, making it difficult for traditional statistical methods to accurately identify and quantify these characteristics. Existing methods are mainly based on simple threshold judgments, which cannot capture the complex patterns of extreme scenarios.

[0006] 3. Lack of specificity: Existing scene generation methods are mainly designed for general scenarios and lack specific optimization for extreme scenarios. Although traditional GAN ​​models can generate scenes, they cannot be specifically optimized for extreme scenarios.

[0007] 4. Insufficient feature extraction: Existing methods cannot effectively extract key features of extreme scenarios, such as duration, scope of impact, and power output level, which makes it difficult to support subsequent power system planning decisions.

[0008] Therefore, current data-driven extreme scene extraction methods suffer from drawbacks such as data scarcity, insufficient feature extraction, and low recognition accuracy. Summary of the Invention

[0009] The purpose of this invention is to provide a high-precision method for automatic extraction and feature quantization of extreme scenarios in power systems based on improved generative adversarial networks, which solves the model training problem under extreme data scarcity and realizes automatic extraction and feature quantization of extreme scenarios.

[0010] The objective of this invention can be achieved through the following technical solutions: An automatic extraction and feature quantization method for extreme scenarios in power systems based on improved generative adversarial networks includes the following steps: Historical data is acquired, extreme scenes are identified and extracted from the historical data using an extreme scene recognition network, and data augmentation strategies are applied to the extreme scenes to form an extreme scene sample set. An extreme scene generation model is obtained by training the extreme scene sample set. The extreme scene generation model is built on an improved GAN architecture, which includes a conditional generator with an attention mechanism and a multi-scale discriminator. The extreme scenes generated by the extreme scene generation model are used to extract quantified four-dimensional features, and a training set is constructed based on the extreme scenes generated by the extreme scene generation model to train the extreme scene recognition network.

[0011] Furthermore, the historical data includes meteorological data, renewable energy output data, and load data.

[0012] Furthermore, the extreme scene recognition network adopts a hybrid architecture of convolutional neural networks and long short-term memory networks.

[0013] Furthermore, the data augmentation strategy includes at least one or more of the following: time window sliding, noise injection, data interpolation, time series transformation, and transfer learning.

[0014] Furthermore, the four-dimensional features include the duration of the extreme scenario, the scope of impact, the level of output, and the severity.

[0015] Furthermore, the multi-scale discriminator includes temporal scale, spatial scale, and feature scale.

[0016] Furthermore, the method also includes: Based on the quantified four-dimensional features, a multidimensional joint probability distribution model is used to obtain the probability of occurrence of each extreme scenario and the correlation between different extreme scenarios. The multidimensional joint probability distribution includes the probability distribution of meteorological conditions, new energy output, and load demand.

[0017] Furthermore, the multidimensional joint probability distribution model is constructed based on a Gaussian mixture model or a variational autoencoder.

[0018] The present invention also provides a computer device, including a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method described above.

[0019] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described above.

[0020] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention addresses the technical challenges of scarce extreme data and complex scenario features in power system planning under extreme weather conditions. By constructing an improved GAN architecture with a condition generator and a multi-scale discriminator, combined with data augmentation and multi-task learning strategies, it achieves automatic extraction of extreme scenarios and quantification of four-dimensional features. This provides important technical support for the coordinated planning of power system sources, grids, loads, and storage, and effectively enhances the ability of new power systems to cope with extreme events.

[0021] 2. This invention effectively solves the model training problem under extreme data scarcity by employing multiple data augmentation strategies, including time window sliding, noise injection, data interpolation, time series transformation, and transfer learning. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the principle architecture of the present invention; Figure 2 This is a flowchart of the data preprocessing and enhancement strategy of the present invention; Figure 3 This is a schematic diagram of the improved GAN architecture of the present invention; Figure 4 This is a flowchart of the four-dimensional feature quantization process of the present invention; Figure 5 This is a flowchart of the multidimensional joint probability distribution modeling process of the present invention; Figure 6 This is a schematic diagram of the extraction results for a continuous rainy and low-temperature scene in an embodiment of the present invention; Figure 7 This is a schematic diagram of the extraction results for a summer high-temperature and scorching heat scene in an embodiment of the present invention; Figure 8 This is a schematic diagram of the extreme scene feature quantization results in an embodiment of the present invention; Figure 9 This is a schematic diagram of the device structure provided in the embodiments of the present invention; Figure 10 This is a schematic diagram of the computer device structure provided in an embodiment of the present invention. Detailed Implementation

[0023] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0024] Example 1 refer to Figure 1 As shown, this embodiment provides a method for automatic extraction and feature quantization of extreme scenarios in power systems based on improved generative adversarial networks, including the following steps: S1. Obtain historical data.

[0025] To improve the quality of extreme scenario identification and generation, it is necessary to collect multi-source heterogeneous data, including meteorological data, renewable energy output data, and load data. Meteorological data includes temperature, humidity, wind speed, and rainfall; renewable energy output data includes actual output data of renewable energy sources such as wind power and photovoltaics; and load data includes load demand data of the power system.

[0026] S2. Using an extreme scene recognition network, identify and extract extreme scenes from the historical data, and apply data augmentation strategies to augment the extreme scenes to form an extreme scene sample set. The data augmentation strategies include at least one or more of the following: time window sliding, noise injection, data interpolation, time series transformation, and transfer learning.

[0027] The extreme scene recognition network can automatically identify extreme scenes from historical data, specifically: 1. Scene recognition network structure: A hybrid architecture combining convolutional neural networks (CNN) and long short-term memory networks (LSTM) is employed. CNN is used to extract spatial features, while LSTM is used to capture time series features; An attention mechanism is introduced to focus on key time points and input data dimensions (such as temperature, wind speed, irradiance, load, etc.) of extreme events.

[0028] 2. Definition of recognition rules: Continuous overcast and low-temperature scenario: If there are ≥3 consecutive days of overcast and rainy weather, the temperature is below the threshold, and the photovoltaic output decreases by more than 50%; Summer high temperature and heat scenario: ≥5 consecutive days of high temperature, temperature exceeding 35℃, and load increase exceeding 20%; Extreme wind speed scenario: When the wind speed exceeds the threshold, the wind power output fluctuates abnormally.

[0029] 3. Recognition Algorithm: The sliding window method is used to analyze historical data segment by segment; By combining meteorological conditions and new energy output data, a comprehensive judgment is made as to whether it is an extreme scenario; It outputs basic recognition information such as scene type, start time, and end time, providing input for subsequent feature quantization.

[0030] The extreme scene recognition network can automatically identify various types of extreme scenes from historical data, providing a foundation for subsequent feature quantification.

[0031] Because extreme events occur infrequently in historical data, training data is scarce, making it difficult to train effective recognition models. Therefore, such as Figure 2 As shown, the data preprocessing and augmentation steps in this embodiment include: first, acquiring raw data, including meteorological data, new energy output data and load data from historical data, and collecting and cleaning the raw data; then, using data augmentation strategies (including time window sliding, noise injection, data interpolation, and time series transformation) to expand the training samples; and finally, further improving the model's learning ability for extreme scenarios through transfer learning, ultimately forming an augmented extreme scenario training sample set.

[0032] Specifically, this embodiment employs various data augmentation strategies to expand the training samples for extreme scenarios: 1. Sliding time window: By sliding the time window, multiple training samples are extracted from continuous time series, increasing the number of samples.

[0033] 2. Noise injection: Add appropriate Gaussian noise to the original data to increase data diversity and improve the model's generalization ability.

[0034] 3. Data interpolation: For missing extreme scenario data, time series interpolation methods are used to supplement it.

[0035] 4. Time series transformation: New training samples are generated by transforming time series data, such as time scaling and time shifting.

[0036] 5. Transfer learning: Pre-train the model using general scenario data, and then fine-tune it on extreme scenario data to make full use of limited data resources.

[0037] The data augmentation strategies described above can effectively expand the training samples for extreme scenarios, solve the problem of data scarcity, and provide sufficient data support for subsequent model training.

[0038] In this embodiment, the specific implementation method of the data augmentation strategy is as follows: (1) Enhanced sliding time window: Let the original time series data be X = {x1, x2, ..., x...} n If the window size is w and the stride is s, then the generated samples are: X_i = {x_{i s+1}, x_{i s+2}, ..., x_{i*s+w}} Among them, i = 0, 1, 2, ..., ⌊(nw) / s⌋; (2) Noise injection enhancement: Add Gaussian noise to the original data: X_noise = X + ε Where, ε ~ N(0, σ 2 ), where σ is the noise standard deviation; (13) Data interpolation enhancement: For missing extreme scenario data, linear interpolation or spline interpolation is used: x_interp = f(x_known, t_known, t_missing) Where f is the interpolation function; (4) Time series transformation enhancement: Generate new samples by time scaling and translation: X_scale = {x_{αt1}, x_{αt2}, ..., x_{αt n}} X_shift = {x_{t1+β}, x_{t2+β}, ..., x_{t n +β}} Where α is the scaling factor and β is the translation amount.

[0039] S3. An extreme scene generation model is obtained by training an extreme scene sample set after data augmentation. The extreme scene generation model is built on an improved GAN architecture, which includes a conditional generator with an attention mechanism and a multi-scale discriminator.

[0040] While traditional GAN ​​models can generate scenes, they cannot be optimized specifically for extreme scenarios. Therefore, this embodiment improves the traditional GAN ​​architecture by constructing a conditional generator and a multi-scale discriminator specifically for extreme scenarios, such as... Figure 3 As shown.

[0041] 1. Condition generator design: Input a scenario condition vector, including condition information such as season, weather type, and severity. Generate corresponding extreme scenario sequences, ensuring that the generated scenarios meet the specified conditions. Introducing an attention mechanism to focus on key time points and feature dimensions of extreme events. 2. Multi-scale discriminator design: Time-scale discriminator: Determines the realism of the generated scene from a time dimension. Spatial Scale Discriminator: Determines the realism of the generated scene from a spatial dimension Feature Scale Discriminator: Determines the realism of the generated scene based on feature dimensions. Multi-scale discriminators can more comprehensively evaluate the quality of generated scenes. 3. Adversarial training strategy: An improved adversarial training strategy is adopted to address the problem of extreme data scarcity. Using Wasserstein distance as the loss function improves training stability. Introducing a gradient penalty mechanism to prevent pattern collapse The improved GAN architecture can generate high-quality extreme scenes, providing data support for subsequent scene recognition and feature extraction.

[0042] In this embodiment, the specific design of the improved GAN architecture is as follows: (1) Condition generator structure: The input of the condition generator G includes the noise vector z and the condition vector c, and the output is the generated extreme scene sequence: G(z, c) = x_generated, where z ~ N(0, I) and c is the scene condition vector; (2) Multi-scale discriminator structure: The multi-scale discriminator D contains three sub-discriminators: Time-scale discriminator: D_temporal(x) Spatial scale discriminator: D_spatial(x) Feature scale discriminator: D_feature(x) The final judgment result is: D(x) = w1D_temporal(x) + w2D_spatial(x) + w3D_feature(x) Where w1, w2, and w3 are weighting coefficients; (3) Loss function design: Wasserstein distance is used as the loss function: L_GAN = E[D(x_real)] - E[D(G(z, c))] + λGP Where λ is the gradient penalty coefficient and GP is the gradient penalty term; (4) Attention Mechanism: Introduce an attention mechanism into the generator and discriminator: Attention(Q, K, V) = softmax(QK^T / √d_k)V Where Q, K, and V are the query, key, and value matrices, respectively.

[0043] S4. Based on the extreme scene generation model, extract quantified four-dimensional features from the generated extreme scene, such as... Figure 4 As shown.

[0044] In this step, a four-dimensional feature quantization model is used to extract features from extreme scenarios. In this embodiment, the four-dimensional feature quantization model is a rule-based calculation model based on mathematical formulas.

[0045] In this embodiment, the four-dimensional features include the duration of the extreme scenario, the scope of impact, the output level, and the severity. Specific quantitative data include: 1. Duration characteristic: Calculate the duration of the scene Analyze the start and end times of the scenario. Quantitative metrics for continuous scenarios 2. Scope of Influence: Geographical range affected by the calculation scenario System Scale for Analyzing the Impact of Scenarios Quantitative scenario coverage metrics 3. Output Level Characteristics: Calculating the degree of anomaly in new energy output Analyze the amplitude and frequency of power output fluctuations. Quantitative output anomaly indicators 4. Severity characteristics: The impact of computing scenarios on power systems Analyze the load changes caused by the scenario Severity indicators for quantitative scenarios Each dimension's features are quantified using mathematical formulas to ensure comparability and interpretability.

[0046] In this embodiment, the specific calculation method for each feature is as follows: (1) Duration characteristics: D = t_end - t_start Where t_start is the scene start time and t_end is the scene end time; (2) Scope of influence: S = (A_affected / A_total) × (P_affected / P_total) Where A_affected is the area of ​​the affected region, A_total is the total area of ​​the region, P_affected is the affected load, and P_total is the total load; (3) Output Level Characteristics: O = |P_actual - P_expected| / P_expected Where P_actual is the actual output and P_expected is the expected output; (4) Severity characteristics: SEV = (ΔP_max / P_base) × (D / D_max) Where ΔP_max is the maximum load change, P_base is the baseline load, D is the duration, and D_max is the maximum possible duration.

[0047] In a preferred embodiment, a training set can be constructed based on the extreme scenes generated by the extreme scene generation model described above, and the extreme scene recognition network can be trained. The generation model can generate more similar extreme scene samples for subsequent training, planning analysis, and testing.

[0048] This embodiment addresses the technical challenges of scarce extreme data and complex scenario features in power system planning under extreme weather conditions. By constructing an improved GAN architecture with a condition generator and a multi-scale discriminator, combined with data augmentation, transfer learning, and multi-task learning strategies, it achieves automatic extraction and four-dimensional feature quantization of extreme scenarios. This invention solves the model training challenge under extreme data scarcity, realizes automatic extraction and feature quantization of extreme scenarios, provides important technical support for the coordinated planning of power system generation, grid, load, and storage, and effectively improves the ability of new power systems to respond to extreme events.

[0049] If the above methods are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0050] Example 2 refer to Figure 1 As shown in this embodiment, the automatic extraction and feature quantization method for extreme scenarios in power systems based on improved generative adversarial networks also includes: S5. Based on the quantified four-dimensional features, a multi-dimensional joint probability distribution model is used to obtain the probability of occurrence of each extreme scenario and the correlation between different extreme scenarios. The multi-dimensional joint probability distribution includes probability distributions of meteorological conditions, renewable energy output, and load demand. Figure 5 As shown, specifically: 1. Multidimensional joint distribution modeling: Establish a multi-dimensional joint probability distribution of meteorological conditions, new energy output, and load demand. Modeling can be performed using Gaussian mixture models (GMM) or variational autoencoders (VAE). Consider the correlation between the dimensions 2. Extreme scenario probability estimation: Based on the joint distribution model, the probability of extreme scenarios occurring is estimated. Calculate the probability distribution of scenarios with different severity levels. Provide probability prediction and risk assessment 3. Scene correlation analysis: The correlations between different extreme scenarios were analyzed using Pearson correlation coefficient, Spearman rank correlation, or mutual information methods. Granger causality tests or time series cross-correlation analysis are used to study the temporal and causal relationships of scenarios. Using association rule mining algorithms or conditional probability analysis, we can identify the triggering conditions and propagation patterns of a scene. By using multidimensional joint probability distribution models, we can gain a deeper understanding of the statistical characteristics of extreme scenarios and provide a scientific basis for power system planning.

[0051] In this embodiment, historical data from a specific region in a specific year is selected as an example for analysis, including meteorological data, renewable energy output data, and load data. Using the method proposed in this invention, various types of extreme scenarios are successfully identified, including continuous rainy and low-temperature scenarios, and hot summer scenarios.

[0052] Based on the identification results, a four-dimensional feature quantization model is used to extract features from extreme scenes. In this embodiment, the four-dimensional feature quantization model is a rule-based calculation model based on mathematical formulas. The feature quantization results for the continuous rainy and low-temperature scene are as follows: Figure 6 As shown, the feature quantification results of the summer high temperature and scorching heat scene are as follows: Figure 7 As shown, the four-dimensional feature quantization results for extreme scenarios are as follows: Figure 8 As shown.

[0053] The results show that the continuous rainy and low-temperature scenario has a long duration (average 7 days), a large affected area (covering 80% of the city), a significant decrease in photovoltaic output (average decrease of 60%), and a high severity (load deficit reaching 15%). The summer high-temperature and scorching scenario has a medium duration (average 5 days), a medium affected area (covering 60% of the city), a moderate load increase (average increase of 25%), and a high severity (load deficit reaching 20%).

[0054] Analysis using a multidimensional joint probability distribution model revealed a correlation between continuous rainy and low-temperature scenarios and hot and scorching summer scenarios. The probability of both scenarios occurring simultaneously is 0.15, providing an important reference for power system planning.

[0055] Example 3 Corresponding to the method in the above embodiments, Figure 9This diagram illustrates the structural block diagram of the automatic extraction and feature quantization device for extreme scenarios in power systems based on an improved generative adversarial network, as provided in this embodiment. The device is applied to a computer device, which connects to a target database via a pre-defined application programming interface (API). When the target database is driven to perform corresponding tasks, corresponding task logs are generated, which can be collected via the API. For ease of explanation, only the parts relevant to this embodiment are shown.

[0056] See Figure 9 As shown, the device includes: a data preprocessing and augmentation module for collecting multi-source data and performing data augmentation processing; an improved GAN training module for training a condition generator and a multi-scale discriminator; an extreme scene recognition module for automatically identifying extreme scenes in historical data; a feature quantization module for extracting and quantifying four-dimensional features of extreme scenes; a probability modeling module for establishing a multi-dimensional joint probability distribution model; and an output module for outputting extreme scene features, probability analysis structure, and risk assessment results.

[0057] Optionally, the data preprocessing and enhancement module includes: a data collection unit for collecting historical meteorological data, new energy output data, and load data; a data enhancement unit for performing data enhancement through time window sliding, noise injection, data interpolation, and time series transformation methods; and a transfer learning unit for pre-training models using general scenario data and then fine-tuning them on extreme scenario data.

[0058] Optionally, the improved GAN training module includes: a condition generator unit, used to construct a condition generator that generates corresponding extreme scene sequences from input scene condition vectors; a multi-scale discriminator unit, used to construct a multi-scale discriminator that makes discriminations from the time scale, spatial scale, and feature scale respectively; and an adversarial training unit, used to train the GAN model using an improved adversarial training strategy.

[0059] Optionally, the extreme scene recognition module includes: a scene recognition network unit for constructing a recognition network that combines CNN and LSTM; a recognition rule unit for defining recognition rules for various extreme scenes; and a recognition algorithm unit for implementing the sliding window recognition algorithm.

[0060] Optionally, the feature quantization module includes: a duration feature unit for calculating the duration of the scenario; an impact range feature unit for calculating the geographical scope and system scale of the scenario's impact; an output level feature unit for calculating the degree of anomaly in the output of new energy sources; and a severity feature unit for calculating the degree of impact of the scenario on the power system.

[0061] Optionally, the probability modeling module includes: a joint distribution modeling unit for establishing a multidimensional joint probability distribution model; a probability estimation unit for estimating the probability of extreme scenarios occurring; and a correlation analysis unit for analyzing the correlation between different extreme scenarios.

[0062] It should be noted that the information interaction and execution process between the above modules are based on the same concept as the method embodiments of the present invention. For details on their specific functions and technical effects, please refer to the method embodiments section, which will not be repeated here.

[0063] Example 4 Figure 10 This is a schematic diagram of the structure of a computer device provided in this embodiment. Figure 10 As shown, the computer device of this embodiment includes: at least one processor ( Figure 10 The diagram shows only one instance of a memory and a computer program stored in the memory that can run on at least one processor. When the processor executes the computer program, it implements the steps in any of the above method embodiments. The processor executes the computer program to implement the above methods, and the memory stores the operating system and application programs that access data from the target database via APIs, supporting various external storage devices.

[0064] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for automatic extraction and feature quantization of extreme scenarios in power systems based on improved generative adversarial networks, characterized in that, Includes the following steps: Historical data is acquired, extreme scenes are identified and extracted from the historical data using an extreme scene recognition network, and data augmentation strategies are applied to the extreme scenes to form an extreme scene sample set. An extreme scene generation model is obtained by training the extreme scene sample set. The extreme scene generation model is built on an improved GAN architecture, which includes a conditional generator with an attention mechanism and a multi-scale discriminator. The extreme scenes generated by the extreme scene generation model are used to extract quantified four-dimensional features, and a training set is constructed based on the extreme scenes generated by the extreme scene generation model to train the extreme scene recognition network.

2. The method for automatic extraction and feature quantization of extreme scenarios in power systems based on improved generative adversarial networks according to claim 1, characterized in that, The historical data includes meteorological data, renewable energy output data, and load data.

3. The method for automatic extraction and feature quantization of extreme scenarios in power systems based on improved generative adversarial networks according to claim 1, characterized in that, The extreme scene recognition network adopts a hybrid architecture of convolutional neural networks and long short-term memory networks.

4. The method for automatic extraction and feature quantization of extreme scenarios in power systems based on improved generative adversarial networks according to claim 1, characterized in that, The data augmentation strategies include at least one or more of the following: time window sliding, noise injection, data interpolation, time series transformation, and transfer learning.

5. The method for automatic extraction and feature quantization of extreme scenarios in power systems based on improved generative adversarial networks according to claim 1, characterized in that, The four-dimensional features include the duration of the extreme scenario, the scope of impact, the level of output, and the severity.

6. The method for automatic extraction and feature quantization of extreme scenarios in power systems based on improved generative adversarial networks according to claim 1, characterized in that, The multi-scale discriminator includes time scale, spatial scale, and feature scale.

7. The method for automatic extraction and feature quantization of extreme scenarios in power systems based on improved generative adversarial networks according to claim 1, characterized in that, The method also includes: Based on the quantified four-dimensional features, a multidimensional joint probability distribution model is used to obtain the probability of occurrence of each extreme scenario and the correlation between different extreme scenarios. The multidimensional joint probability distribution includes the probability distribution of meteorological conditions, new energy output, and load demand.

8. The method for automatic extraction and feature quantization of extreme scenarios in power systems based on improved generative adversarial networks according to claim 7, characterized in that, The multidimensional joint probability distribution model is constructed based on a Gaussian mixture model or a variational autoencoder.

9. A computer device comprising a processor, a memory, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 7.

Images