A power system extreme weather scenario small sample generation and identification method based on a multi-generator-evaluator collaborative mechanism and a storage medium

Abstract

The application relates to a power system extreme weather scene small sample generation and identification method based on a multi-generator-evaluator collaborative mechanism and a storage medium. Through acquisition of power system historical operation data and meteorological observation data, an extreme weather scene real sample set is constructed and preprocessed. A small sample generation model containing a discriminator, an independent evaluator and multiple collaborative generators is constructed. Through a joint training mechanism of alternating iteration of the discriminator, the evaluator and the generator, high-quality expanded sample sets are generated in combination with discriminant constraints and similarity measurement constraints. After effectiveness verification, the expanded samples and the training samples are combined to obtain an enhanced sample set, which is used for training a scene identification model to realize accurate identification of the extreme weather scene. The application effectively solves the problems of small sample data scarcity, poor model training collaboration and low identification precision, improves the model training stability and sample expansion quality, and provides reliable technical support for power system extreme weather prevention and control.

Description

Technical Field

[0001] This application relates to the field of power system security and control, and in particular to a method and storage medium for generating and identifying small samples of extreme weather scenarios in power systems based on a multi-generator-evaluator collaborative mechanism. Background Technology

[0002] As the penetration rate of new energy sources such as wind power and photovoltaics in the power system continues to increase, the volatility, randomness, and intermittency of power supply output are becoming increasingly pronounced. Especially under extreme weather conditions such as high temperatures, torrential rains, cold waves, and strong winds, the coupling relationship between new energy output, load demand, and meteorological conditions becomes more complex, easily leading to power system supply-demand imbalances, localized overloads, difficulties in peak shaving, and increased operational risks. Therefore, accurately identifying extreme weather scenarios is of great significance in power system operation analysis, short-term power balance assessment, risk warning, and dispatch decision-making.

[0003] Most existing extreme weather scene identification methods employ data-driven models, extracting features from multi-dimensional time-series data such as historical load, temperature, wind speed, and irradiance, and then using models such as support vector machines, random forests, long short-term memory networks, gated recurrent unit networks, or convolutional neural networks for classification and identification. These methods can achieve good recognition results when there are sufficient samples. However, extreme weather events typically occur infrequently in actual historical data, resulting in a small number of corresponding samples and an imbalanced class distribution. This makes the identification models prone to overfitting, reducing their generalization ability and making it difficult to meet the needs of practical applications.

[0004] To alleviate the scarcity of extreme weather samples, existing technologies employ generative models such as Generative Adversarial Networks (GANs) to augment minority class samples, thereby improving the training performance of subsequent recognition models. GANs can learn distribution features from original samples and generate new samples with a certain degree of realism, showing strong application potential in data augmentation. However, when GANs are directly used to generate small-sample extreme weather scenarios, the pattern collapse problem easily occurs. This means that different random inputs are mapped to a small number of highly similar output samples, resulting in a lack of diversity in the augmented samples and difficulty in covering the multimodal feature distribution present in real extreme weather scenarios. This problem is particularly pronounced when extreme weather samples are scarce and concentrated in their distribution.

[0005] To address pattern collapse, existing technologies have proposed improvements such as multi-generator structures, regularization constraints, variational autoencoder fusion, and diffusion models. However, these methods still have the following shortcomings: First, while multi-generator structures can expand the feature mapping range, different generators are prone to learning repetitive patterns, making it difficult to form effective complementarity. Second, simply increasing generation diversity may compromise sample authenticity, failing to balance sample quality and distribution coverage. Third, existing discriminators are mainly used to distinguish between real and augmented samples, making it difficult to directly constrain the similarity between multiple generators, thus failing to effectively suppress pattern collapse from a mechanistic perspective.

[0006] Therefore, there is an urgent need for a small sample generation method for identifying extreme weather scenarios in power systems, which can improve the diversity and distribution coverage of the expanded samples while maintaining the authenticity of the expanded samples, and further enhance the accuracy and generalization performance of downstream extreme weather scenario identification. Summary of the Invention

[0007] The purpose of this application is to address the aforementioned problems in the existing technology by providing a method and storage medium for generating and recognizing small samples of extreme weather scenarios in power systems based on a multi-generator-evaluator collaborative mechanism. This method generates expanded samples and merges them with training samples to obtain an enhanced sample set, which is then used to train a scene recognition model. This enables accurate recognition of extreme weather scenarios, improves model training stability and sample expansion quality, and effectively solves problems such as scarce small sample data, poor model training collaboration, and low recognition accuracy. This provides reliable technical support for the prevention and control of extreme weather in power systems.

[0008] The above-mentioned technical objectives of this application are mainly achieved through the following technical solutions: A method for generating and identifying small samples of extreme weather scenarios in power systems based on a multi-generator-evaluator collaborative mechanism includes: Historical power system operation data and meteorological observation data are acquired to construct a real sample set of extreme weather scenarios, and the real sample set is preprocessed. A few-shot generation model is constructed by setting up a discriminator, an evaluator independent of the discriminator, and at least two generators that work together through a pre-defined connection structure. The few-shot generation model is trained through a joint training mechanism. During training, a discriminator is used for discrimination constraints, and an evaluator is used for similarity measurement constraints to suppress mode collapse. The trained small sample generation model is used to generate an expanded sample set (the expanded sample set can also be called the generated sample set), and the expanded sample set is fused with the real sample set to obtain an enhanced sample set; The scene recognition model is trained based on the enhanced sample set, and the trained scene recognition model is used to identify extreme weather scenes of the samples to be identified, and the recognition results are output.

[0009] This technical solution addresses the challenges of scarce small-sample data, overfitting during model training, and low scene recognition accuracy in extreme weather scenarios in power systems by establishing a complete chain of "real sample construction, model construction, joint training, sample expansion, and scene recognition." Through the dual constraints of the discriminator and evaluator, it effectively suppresses mode collapse in multi-generator models, improves the realism and diversity of small-sample generation, and ultimately achieves accurate identification of extreme weather scenarios, providing reliable technical support for extreme weather prevention and control in power systems.

[0010] As a further improvement and supplement to the above technical solution, this application adopts the following technical measures: As a preferred option, the steps of using a real sample set and preprocessing include: The acquired historical power system operation data and meteorological observation data were filtered to construct an extreme weather scenario dataset; The extreme weather scenario dataset is sliced ​​to obtain the real sample set; Preprocessing is performed on the real sample set by at least one of missing value processing, outlier processing, and normalization processing. The preprocessed real sample set is divided into a training sample set and a test sample set. The training sample set is used to train the few-shot generation model, and the test sample set is used to validate the few-shot generation model.

[0011] The preprocessing workflow for real sample sets is refined. By screening and labeling, slice normalization, data augmentation, and outlier removal, the purity and standardization of real sample sets are improved, avoiding interference from outlier and invalid data in subsequent model training. By dividing the real sample set into training and testing sample sets, the training and validation of small sample generation models are separated, ensuring the stability of small sample generation model training. At the same time, it provides a standard reference for the validity verification of subsequent expanded sample sets, further enhancing the rigor of the technical solution.

[0012] Preferably, the step of generating the expanded sample set includes: Random latent variables are generated through the latent space input module in the small sample generation model; Each generator generates augmented samples based on the random latent variable mapping, ensuring that the dimensions of the augmented samples are consistent with those of the real samples; The discriminator receives samples from the real sample set and generated expanded samples, determines the source of the samples, and applies discriminative constraints to the expanded samples to make them approximate the real samples and ensure that the two are distributed in a consistent manner. The evaluator imposes similarity measurement constraints on the expansion samples and on the expansion samples and the real samples; After completing the discriminant constraints and similarity measurement constraints, an expanded sample set is formed.

[0013] The steps for generating the augmented sample set are clearly defined. Random latent variables are generated through the latent space input module to ensure the diversity of augmented samples generated by each generator. By combining the discriminant constraint of the discriminator and the similarity metric constraint of the evaluator, the augmented samples are guaranteed to approximate the distribution of real samples and improve the authenticity of the samples, while avoiding redundancy among the augmented samples. This solves the problems of "insufficient authenticity and lack of diversity" in small sample augmentation and lays a high-quality data foundation for the construction of subsequent augmented sample sets.

[0014] Preferably, the discriminator performs the discrimination constraint as follows: After receiving samples from the real sample set and the generated expanded samples, the discriminator outputs the probability value that the sample belongs to the real distribution. Based on the probability values, a discriminator loss function and a generator adversarial loss function are constructed; By minimizing the discriminator loss function, the discriminator's ability to distinguish between the real sample set and the augmented sample set is improved. By minimizing the generator adversarial loss function, the generator is optimized to make the augmented sample distribution approximate the true sample distribution.

[0015] The discriminant constraint steps are refined, and the true distribution probability values ​​of the output samples provide accurate data support for the construction of the loss function. By minimizing the two types of loss functions, the discriminant's distinguishing ability and the generator's generation accuracy are improved respectively, strengthening the effectiveness of the discriminant constraint, ensuring that the distribution of the expanded samples is consistent with that of the real samples, further making up for the shortcoming of scarce small sample data, and providing core technical support for the joint training of small sample generation models.

[0016] Preferably, the step of the evaluator using cosine similarity to measure sample similarity includes: Calculate the mean sample of the augmented sample set and the mean sample of the true sample set respectively; Based on two mean samples, calculate the mean cosine similarity between each augmented sample and the mean sample of the augmented sample set, and the mean cosine similarity between each real sample and the mean sample of the real sample set. Based on the deviation between the two sets of mean cosine similarity, a similarity score is determined to characterize the diversity between the augmented sample set and the real sample set. The generator is subjected to diversity constraints based on the similarity score, so that the diversity of the expanded sample set is close to that of the real sample set.

[0017] Cosine similarity is used to measure the similarity of the evaluator. By calculating the mean sample, mean cosine similarity and bias, the diversity difference between the expanded sample set and the real sample set is accurately characterized. Based on the similarity score, the generator is subjected to diversity constraints, which effectively avoids the problems of redundancy and monotonous patterns in the expanded samples of multiple generators. It works in synergy with the discriminator's discrimination constraints to further suppress model pattern collapse and improve the diversity and rationality of the expanded sample set.

[0018] Preferably, a joint loss function is constructed based on the generator adversarial loss function and the similarity score; By minimizing the joint loss function, the samples generated by each generator can approximate the real samples. The steps for jointly training the few-shot generation model are as follows: By fixing the parameters of each generator and evaluator, the parameters of the discriminator are updated using real and expanded samples to improve the discriminator's ability to distinguish between real and expanded samples. The parameters of the discriminator and each generator are fixed, and the parameters of the evaluator are updated according to the similarity scores between the output samples of each generator to enhance the evaluator's ability to represent the redundant patterns of the expanded samples. By fixing the discriminator and evaluator parameters, and updating multiple generator parameters based on the joint loss function, the generator can improve the authenticity of samples while reducing the redundancy of output. Repeat the parameter update steps in sequence, and continuously iterate and optimize the discriminator, evaluator and generators until the preset number of training rounds or the joint loss function converges.

[0019] By fusing the generator adversarial loss function with similarity scores to construct a joint loss function, the "realism constraint" and "diversity constraint" are unified, ensuring that the expanded samples approximate the real distribution while possessing sufficient diversity. Through alternating iterative optimization of the "discriminator, evaluator, and generator," the synergistic improvement of each module of the small sample generation model is achieved, solving the problems of module optimization disconnect and adversarial training mechanism fragmentation in traditional joint training. This significantly improves the convergence speed and stability of model training, ensuring that the model outputs high-quality expanded samples.

[0020] As a preferred approach, latent variables are input into the trained small sample generation model, and each generator outputs candidate augmented samples. Candidate augmentation samples are selected or fused according to preset rules to obtain an augmentation sample set; Based on the aforementioned test sample set, the effectiveness of the expanded sample set is verified; The augmented samples that pass the validity verification are merged with the training sample set to obtain the enhanced sample set.

[0021] The quality and diversity of the expanded sample set are further improved by selecting / merging candidate augmented samples output by each generator; the validity of the expanded sample set is verified based on the test sample set, and samples that do not meet the requirements are removed to ensure the reliability of the expanded samples; the verified expanded samples are combined with the training sample set to construct an enhanced sample set, which effectively makes up for the lack of small sample data, provides sufficient and high-quality training data for the subsequent training of the scene recognition model, and improves the training effect and recognition accuracy of the scene recognition model.

[0022] As a preferred option, the steps for training the scene recognition model and performing extreme weather scene recognition include: Based on the enhanced sample set, the scene recognition model extracts features from the enhanced sample set; Based on the extracted features, prediction results are obtained for the corresponding samples, and these prediction results are used to distinguish extreme weather scenarios. Based on the enhanced sample set and the prediction results, the scene recognition model is trained using the cross-entropy loss function; Based on the trained scene recognition model, it is applied to the test sample set and the sample set to be recognized to output the extreme weather scene recognition results.

[0023] This method refines the training and recognition steps of the scene recognition model. Through feature extraction and cross-entropy loss function training, it improves the scene recognition model's ability to extract features and classify extreme weather scenes. The trained model is then applied to the augmented sample set and the sample set to be recognized, enabling rapid and accurate recognition of extreme weather scenes. This solves the problems of low recognition accuracy and poor generalization ability of traditional recognition models in small sample scenarios, providing technical support for early warning and timely prevention and control of extreme weather in the power system.

[0024] Preferably, the scene recognition model is a convolutional neural network model, which includes at least two convolutional layers, at least two pooling layers, and at least one classification output layer. The connection structure can be any one of the following: parallel connection, series connection, or hierarchical connection.

[0025] The scene recognition model is clearly defined as a convolutional neural network model. Its multi-layer convolution and pooling structure can accurately extract deep features of extreme weather scene samples, improving recognition accuracy and generalization ability. The generator has multiple connection structures, which can be flexibly selected according to different data characteristics and training requirements of extreme weather scenes in the power system, enhancing the flexibility and applicability of the technical solution and expanding its application scope.

[0026] The technical solution related to the second technical topic in this technical solution is as follows: A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned method for generating and identifying small samples of extreme weather scenarios in power systems based on a multi-generator-evaluator collaborative mechanism.

[0027] This facilitates the storage of the aforementioned small sample generation and recognition method based on the multi-generator-evaluator collaborative mechanism in the form of a computer program, enabling the method to be stored, invoked, and portable. This makes it easier to promote and apply the technical solution in different power system scenarios, reduce the cost of technology implementation, and provide support for the engineering implementation of the method, ensuring that the technical solution can be quickly transformed into practical applications and play its role in extreme weather prevention and control.

[0028] The beneficial effects of this application are: 1. Effectively solves the problem of scarce small sample data and improves the quality of sample augmentation - by generating augmented samples collaboratively by multiple generators, and combining the discriminant constraint of the discriminator with the similarity metric constraint of the evaluator, the model mode collapse is suppressed, ensuring that the augmented samples not only closely approximate the real sample distribution and have high realism, but also have sufficient diversity. This effectively makes up for the shortcoming of scarce real samples in extreme weather scenarios of power systems and provides high-quality data support for subsequent model training.

[0029] 2. Optimize model training synergy and improve training stability and efficiency. The joint training mechanism of alternating iteration of "discriminator → evaluator → generator" is adopted to achieve synergistic optimization of the discriminator, evaluator and generator. This solves the problems of the separation between adversarial training mechanism and training steps and the disconnect between the optimization of each module in the existing technology. It greatly improves the training convergence speed and stability of small sample generation model and reduces model training cost.

[0030] 3. Improve the accuracy of extreme weather scene recognition and enhance the power system's prevention and control capabilities—Construct a high-quality enhanced sample set through effectiveness verification to train the scene recognition model, effectively improving the model's feature extraction capability and classification accuracy for extreme weather scenes. This solves the problems of low recognition accuracy and poor generalization ability of traditional recognition models in small sample scenarios, enabling accurate recognition and early warning of extreme weather scenes, and providing reliable technical support for the safe and stable operation of the power system.

[0031] 4. Enhance the flexibility and applicability of technical solutions – Clarify that the generator can adopt parallel, series, or hierarchical connection structures, which can be flexibly selected according to the extreme weather data characteristics and training needs of different power systems, adapting to various power system application scenarios and expanding the application scope of technical solutions.

[0032] 5. Facilitates engineering implementation and promotion – By storing the method in the form of a computer program through a computer-readable storage medium, the technical solution can be stored, called, and ported, reducing the cost of technology implementation and promoting the large-scale application of this technology in the field of extreme weather prevention and control in power systems. It has significant practical value and promotion prospects. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the two-stage overall process of the small sample generation and identification method for extreme weather scenarios in power systems involved in this application; Figure 2 This is a schematic diagram of a structure involving multiple generators connected in parallel, as described in this application. Figure 3 This is a schematic diagram of a multi-generator structure connected in series, which is involved in this application. Figure 4 This is a schematic diagram of a multi-generator structure with hierarchical connections involved in this application; Figure 5 This is a schematic diagram illustrating the collaborative operation of the evaluator and discriminator in this application; Figure 6 This is a schematic diagram showing the distribution of four time parameters in the real extreme weather samples involved in this application; Figure 7 This is a schematic diagram comparing the distribution of generation results from different models involved in this application. Detailed Implementation

[0034] The technical solution of this application will be further described in detail below through embodiments and in conjunction with the accompanying drawings.

[0035] Example 1: A method for generating and identifying small samples of extreme weather scenarios in power systems based on a multi-generator-evaluator collaborative mechanism, comprising: S1. Obtain historical operating data of the power system and meteorological observation data, construct a real sample set of extreme weather scenarios, and preprocess the real sample set; S2. Construct a few-shot generation model by setting a discriminator, an evaluator independent of the discriminator, and at least two generators that work together through a preset connection structure. S3. The small sample generation model is trained through a joint training mechanism. During training, a discriminator is used for discrimination constraints, and an evaluator is used for similarity measurement constraints to suppress mode collapse. S4. Use the trained small sample generation model to generate an expanded sample set, and merge the expanded sample set with the real sample set to obtain an enhanced sample set; S5. Train the scene recognition model based on the enhanced sample set, and use the trained scene recognition model to identify extreme weather scenes of the samples to be identified, and output the recognition results.

[0036] In practical applications, the steps of using real sample sets and preprocessing include: The acquired historical power system operation data and meteorological observation data were filtered to construct an extreme weather scenario dataset; The extreme weather scenario dataset is sliced ​​to obtain the real sample set; Preprocessing is performed on the real sample set by at least one of missing value processing, outlier processing, and normalization processing. The preprocessed real sample set is divided into a training sample set and a test sample set. The training sample set is used to train the few-shot generation model, and the test sample set is used to validate the few-shot generation model.

[0037] In practical applications, the steps for generating the expanded sample set include: Random latent variables are generated through the latent space input module in the small sample generation model; Each generator generates augmented samples based on the random latent variable mapping, ensuring that the dimensions of the augmented samples are consistent with those of the real samples; The discriminator receives samples from the real sample set and generated expanded samples, determines the source of the samples, and applies discriminative constraints to the expanded samples to make them approximate the real samples and ensure that the two are distributed in a consistent manner. The evaluator imposes similarity measurement constraints on the expansion samples and on the expansion samples and the real samples; After completing the discriminant constraints and similarity measurement constraints, an expanded sample set is formed.

[0038] In practical applications, the discriminator performs the following steps for determining the constraints: After receiving samples from the real sample set and the generated expanded samples, the discriminator outputs the probability value that the sample belongs to the real distribution. Based on the probability values, a discriminator loss function and a generator adversarial loss function are constructed; By minimizing the discriminator loss function, the discriminator's ability to distinguish between the real sample set and the augmented sample set is improved. By minimizing the generator adversarial loss function, the generator is optimized to make the augmented sample distribution approximate the true sample distribution.

[0039] In practical applications, the steps of the evaluator in using cosine similarity to measure sample similarity include: Calculate the mean sample of the augmented sample set and the mean sample of the true sample set respectively; Based on two mean samples, calculate the mean cosine similarity between each augmented sample and the mean sample of the augmented sample set, and the mean cosine similarity between each real sample and the mean sample of the real sample set. Based on the deviation between the two sets of mean cosine similarity, a similarity score is determined to characterize the diversity between the augmented sample set and the real sample set. The generator is subjected to diversity constraints based on the similarity score, so that the diversity of the expanded sample set is close to that of the real sample set.

[0040] In practical applications, a joint loss function is constructed based on the generator adversarial loss function and the similarity score; By minimizing the joint loss function, the samples generated by each generator can approximate the real samples. The steps for jointly training the few-shot generation model are as follows: By fixing the parameters of each generator and evaluator, the parameters of the discriminator are updated using real and expanded samples to improve the discriminator's ability to distinguish between real and expanded samples. The parameters of the discriminator and each generator are fixed, and the parameters of the evaluator are updated according to the similarity scores between the output samples of each generator to enhance the evaluator's ability to represent the redundant patterns of the expanded samples. By fixing the discriminator and evaluator parameters, and updating multiple generator parameters based on the joint loss function, the generator can improve the authenticity of samples while reducing the redundancy of output. Repeat the parameter update steps in sequence, and continuously iterate and optimize the discriminator, evaluator and generators until the preset number of training rounds or the joint loss function converges.

[0041] In practical applications, latent variables are input into the trained small sample generation model, and each generator outputs candidate augmented samples. Candidate augmentation samples are selected or fused according to preset rules to obtain an augmentation sample set; Based on the aforementioned test sample set, the effectiveness of the expanded sample set is verified; The augmented samples that pass the validity verification are merged with the training sample set to obtain the enhanced sample set.

[0042] The steps for training a scene recognition model and performing extreme weather scene recognition include: Based on the enhanced sample set, the scene recognition model extracts features from the enhanced sample set; Based on the extracted features, prediction results are obtained for the corresponding samples, and these prediction results are used to distinguish extreme weather scenarios. Based on the enhanced sample set and the prediction results, the scene recognition model is trained using the cross-entropy loss function; Based on the trained scene recognition model, it is applied to the test sample set and the sample set to be recognized to output the extreme weather scene recognition results.

[0043] In practical applications, the scene recognition model is a convolutional neural network model, which includes at least two convolutional layers, at least two pooling layers, and at least one classification output layer; the connection structure is any one of parallel connection, serial connection, or hierarchical connection.

[0044] Figure 1 In this technical solution, GAN stands for Generative Adversarial Network, which is used to generate and expand extreme weather scene samples to solve the problem of small sample scarcity. In this technical solution, it is an improved multi-generator GAN. GNN stands for Graph Neural Network, which is used to extract power system topology features and power grid operation relationship features.

[0045] Next, combined Figures 1 to 7 The following is a detailed supplementary explanation of the above-mentioned method for generating and identifying small samples of extreme weather scenarios in power systems based on a multi-generator-evaluator collaborative mechanism: S1. The steps of preprocessing the real sample set include: Step S1.1: Obtain historical operating data and historical meteorological observation data of the power system. The historical operating data includes at least load time-series data, and the meteorological observation data includes at least one or more of temperature, wind speed, and solar irradiance. The data may be sourced from the dispatching operation platform, meteorological monitoring platform, historical database, or other combined power and meteorological data sources, and aligned according to a unified time scale.

[0046] Step S1.2: Based on the preset extreme weather judgment rules, select the sample set corresponding to the target extreme weather scenario from the original historical data (including historical operational data and historical meteorological observation data) (i.e., construct an extreme weather scenario dataset). The extreme weather scenario includes, but is not limited to, any one or more of high temperature, rainstorm, cold wave, and strong wind; the judgment rules can be determined based on temperature threshold, rainfall threshold, wind speed threshold, abnormal radiation threshold, abnormal load fluctuation threshold, or a combination thereof.

[0047] Step S1.3: Slice the selected data into segments using a preset time window to construct extreme weather scenario samples (real sample set). Preferably, samples are constructed on a daily basis, and a single sample is represented as a multi-dimensional time-series vector at multiple consecutive moments; each sample contains at least four types of features: load, temperature, wind speed, and solar irradiance, thereby forming a multivariate time-series sample set for extreme weather scenario identification.

[0048] Step S1.4: Preprocess the constructed multivariate time series sample set. The preprocessing includes one or more of missing value processing, outlier processing, normalization or standardization processing to eliminate the influence of different units and improve the stability of subsequent model training.

[0049] Step S1.5: Divide the preprocessed sample set into a training sample set and a test sample set. The training sample set is used for subsequent small sample generation model training and recognition model training, while the test sample set is used to verify the generation effect and scene recognition effect. During the division process, maintain the consistency of labels and the integrity of time series for various extreme weather samples.

[0050] S2. The steps for building a small sample generation model include: Step S2.1: Construct the latent space input module. A random latent variable input space is defined to provide initial perturbation input for the generative model. The latent variables can be represented as random noise vectors, which follow a preset probability distribution, preferably a Gaussian or uniform distribution. By inputting the same or different latent variables into different generators, each generator can be driven to learn different extreme weather sample distribution patterns.

[0051] Step S2.2: Construct a multi-generator module. Set up at least two generators, preferably three, denoted as Generator 1, Generator 2, and Generator 3. Each generator maps latent variables to extreme weather scenario samples, outputting augmented samples with the same dimensions as the multivariate time-series samples constructed in Step 1. Each generator can use the same or different network parameter structures to learn different pattern regions in the real sample distribution.

[0052] Step S2.3: Construct a connection structure. Multiple generators are connected through the connection structure according to a preset cooperative method. The connection structure can be any one of parallel connection, series connection, or hierarchical connection. Wherein: 1) In parallel connection, each generator receives latent variable input independently and generates candidate samples separately to expand the coverage of different regions of the real sample distribution; 2) In the case of a series connection, the output of the previous generator serves as one of the inputs or intermediate condition information of the next generator, enabling the next generator to gradually correct and refine the results based on the previous generator, thereby improving the precision and diversity of the expanded samples. 3) In hierarchical connection, multiple lower-level generators first generate candidate samples respectively, and then the upper-level generators fuse or remap the candidate samples to construct a hierarchical sample generation process.

[0053] Step S2.4: Construct the discriminator module. A discriminator is configured to receive real samples and augmented samples, and output corresponding discrimination results to determine whether the input sample originates from the real sample distribution or the augmented sample distribution. The discriminator is used to constrain the augmented samples to approximate the real sample distribution, improving the realism and distribution consistency of the augmented samples.

[0054] Step S2.5: Construct the evaluator module. An evaluator is set up outside the multi-generator module. This evaluator is used to compare features and evaluate the similarity of the expanded samples output by different generators to measure the degree of output redundancy among the generators. The evaluator measures the differences between the expanded samples based on feature similarity to identify whether multiple generators have generated repetitive or highly similar patterns, thus providing a basis for subsequent diversity constraints.

[0055] Step S2.6: Construct a small sample generation model: Combine the latent space input module, multi-generator module, connection structure, discriminator module, and evaluator module to form a small sample generation model based on a multi-generator-evaluator collaborative mechanism. This small sample generation model is used to improve the diversity and distribution coverage of small samples generated for extreme weather scenarios by leveraging complementary learning among multiple generators and the redundancy suppression effect of the evaluator, while ensuring the authenticity of the expanded samples.

[0056] S3. Train the few-shot generation model using a joint training mechanism: Step S3.1: The step of generating the expanded sample set includes: Random latent variables are generated through the latent space input module in the small sample generation model; each generator generates expanded samples based on the mapping of the random latent variables, ensuring that the dimensions of the expanded samples are consistent with those of the real samples; the discriminator receives samples from the real sample set and the generated expanded samples, determines the source of the samples, and applies discriminative constraints to the expanded samples to approximate the real samples, ensuring that their distributions are consistent; the evaluator applies similarity measurement constraints between expanded samples and between expanded samples and real samples; after completing the discriminative constraints and similarity measurement constraints, an expanded sample set is formed.

[0057] Specifically: The random latent variables are input into multiple generators constructed in step S2 to obtain multiple expanded samples. Let the first... Let the generators be denoted as Its input latent variables are denoted as Then the first The augmented samples output by each generator can be represented as: (19) in, For the number of generators, Indicates the first An expanded sample of extreme weather scenarios output by a generator.

[0058] When multiple generators are connected in parallel, each generator independently receives latent variables and generates samples. When connected in series, the next generator receives the output information of the previous generator and makes progressive corrections. When connected in a hierarchical manner, the upper-level generator merges or remaps the output results of the lower-level generator.

[0059] Step S3.2: Construct a discriminator to challenge the training target. The discriminator performs the following discriminative constraints: After receiving samples from the real sample set and the generated expanded samples, the discriminator outputs the probability value that the sample belongs to the real distribution. Based on the probability value, a discriminator loss function and a generator adversarial loss function are constructed. By minimizing the discriminator loss function, the discriminator's ability to distinguish between the real sample set and the expanded sample set is improved. By minimizing the generator adversarial loss function, the generator is optimized so that the expanded sample distribution approximates the real sample distribution.

[0060] Specifically: real samples With expanded samples Simultaneously input discriminator The discriminator outputs the probability value of a sample belonging to the true distribution, thus constructing an adversarial training mechanism between the generator and the discriminator. The discriminator loss function can be defined as: (20) in, Represents the true sample distribution. This represents the distribution of latent variables.

[0061] Correspondingly, the adversarial loss function of the generator can be defined as: (twenty one) The discriminator loss is used to enhance the discriminator's ability to distinguish between real samples and augmented samples, and the generator adversarial loss is used to make the augmented samples approximate the distribution of real samples, thereby improving the authenticity of the augmented samples.

[0062] Step S3.3: Constructing a cosine similarity-based evaluator similarity measurement mechanism, wherein the evaluator uses cosine similarity to measure sample similarity, the steps of which include: Calculate the mean sample of the expanded sample set and the mean sample of the real sample set respectively; based on the two mean samples, calculate the mean cosine similarity between each expanded sample and the mean sample of the expanded sample set, and the mean cosine similarity between each real sample and the mean sample of the real sample set respectively; based on the deviation of the two sets of mean cosine similarities, determine the similarity score used to characterize the diversity of the expanded sample set and the real sample set; based on the similarity score, impose diversity constraints on the generator to make the diversity of the expanded sample set close to that of the real sample set.

[0063] Specifically: To characterize the redundancy between the outputs of different generators, samples generated by each generator are input into the evaluator for feature comparison. Preferably, cosine similarity is used as the similarity measure operator to highlight the similarity of the overall shape of multivariate time-series trajectories, while downplaying the impact of simple amplitude differences.

[0064] Let any two augmented samples and The vectorized representations are respectively and Then the cosine similarity between the two is defined as: (twenty two) in, This represents the dot product of two vectors. and These represent the norms of the corresponding vectors.

[0065] Since extreme weather scenario samples are essentially multivariate time-series trajectories, pattern collapse often manifests as different random inputs generating time-series curves with highly similar shapes. Therefore, cosine similarity can more effectively characterize the degree of repetition of expanded samples at the trajectory shape level.

[0066] Step S3.4: Construct the similarity score SS. Details are as follows: To comprehensively assess the diversity structure of the expanded sample set, a similarity score SS is introduced, which is constructed as follows: First, calculate the mean sample of the expanded sample set: (twenty three) in, Indicates the number of samples to be expanded. Indicates the first An expanded sample.

[0067] Then, the mean cosine similarity between each augmented sample and the mean sample of the augmented sample set is calculated: (twenty four) Similarly, the mean of the real sample set is calculated for the real sample set. and its mean cosine similarity: (25) (26) in, Indicates the actual number of samples. Indicates the first One real sample.

[0068] Furthermore, the deviation between the augmented sample set and the real sample set in terms of overall diversity structure is defined as: (27) To penalize both over-concentration and unreasonable divergence of the expanded sample, the absolute value of the above deviations is taken to obtain the final similarity score: (28) in, The smaller the value, the closer the overall diversity structure of the expanded sample set is to the real sample set; The larger the value, the more severe the distributional bias in the expanded sample set compared to the true sample set. In this application, when... When the value is too large, it indicates that different latent variables have failed to generate a sufficiently rich sample pattern, and thus can be regarded as a high degree of pattern collapse.

[0069] Step S3.5: Based on the generator adversarial loss function and the similarity score, construct a joint loss function as follows: To simultaneously ensure the authenticity and diversity of the augmented samples, a diversity constraint term based on similarity scores is introduced on top of the generator adversarial loss to construct a joint loss function. Preferably, the first... The joint loss function of the generators can be expressed as: (29) in, This is a balancing coefficient used to adjust the weight between the authenticity constraint and the diversity constraint. This is a similarity penalty term constructed based on the evaluator output or similarity score.

[0070] Preferably, the similarity penalty term can take the following form: (30) Alternatively, the average similarity of the pairwise similarities between the output samples from different generators can be used: (31) By minimizing the joint loss function mentioned above, multiple generators can approximate the real sample distribution while avoiding generating highly similar samples, thereby reducing the probability of mode collapse.

[0071] Step S3.6: Jointly train the few-shot generation model, that is, alternately and iteratively update the parameters of each module, as follows: The discriminator, evaluator, and multiple generators are jointly trained using an alternating iterative optimization approach, specifically including: 1) Fix the parameters of multiple generators and evaluators, and update the discriminator parameters using real samples and expanded samples to improve the discriminator's ability to distinguish between real samples and expanded samples; 2) Fix the parameters of the discriminator and multiple generators, and update the evaluator parameters according to the cosine similarity or similarity score between the output samples of different generators to enhance its ability to represent redundant patterns of expanded samples. 3) Fix the parameters of the discriminator and evaluator, and update the parameters of multiple generators according to the joint loss function, so that the generator can improve the authenticity of the samples while reducing the redundancy of the output; 4) Repeat the above steps to continuously iterate and optimize the discriminator, evaluator and generators until the preset number of training rounds or the loss function convergence condition is reached.

[0072] Step S3.7: Obtain the trained few-shot generation model. After multiple rounds of iterative training, a trained few-shot generation model based on a multi-generator-evaluator collaborative mechanism is obtained. This model can generate extended samples of extreme weather scenarios with realism, diversity, and strong distribution coverage under limited extreme weather sample conditions, providing data support for subsequent scene recognition.

[0073] Step S4: The step of generating an expanded sample set using the trained small sample generation model and fusing the expanded sample set with the real sample set to obtain an enhanced sample set is as follows: Step S4.1: Input the latent variables into the trained small sample generation model, and each generator will output candidate augmented samples: The random latent variables are input into a trained few-sample generation model based on a multi-generator-evaluator collaborative mechanism, and multiple generators output extreme weather scenarios to expand the samples. Let the input latent variables be... The trained few-shot generation model is Then the expanded sample can be expressed as: (32) in, This represents the generated extended extreme weather samples. When using a multi-generator structure, multiple generators can output candidate extended samples respectively, and the generated results can be selected or merged according to preset rules to obtain the final extended sample set.

[0074] Step S4.2: Select or merge candidate augmented samples according to preset rules to obtain an augmented sample set: The number of expanded samples is determined based on the size and category distribution of the original extreme weather sample set, as well as the requirements of downstream identification tasks. Preferably, the number of expanded samples is greater than the number of original extreme weather samples to alleviate the problems of sample scarcity and category imbalance; more preferably, the number of expanded samples is set to a preset fixed value so as to compare with different generation methods under the same expansion scale.

[0075] Step S4.3: Based on the test sample set, perform validity verification on the expanded sample set. The validity verification includes one or more of the following: distribution consistency verification, sample diversity verification, and value rationality verification.

[0076] in: 1) Distribution consistency check is used to determine whether the expanded sample is consistent with the actual extreme weather sample in terms of overall distribution; 2) Sample diversity verification is used to determine whether there is excessive concentration or high degree of duplication among the expanded samples; 3) Value rationality verification is used to determine whether the characteristics such as load, temperature, wind speed and solar irradiance in the expanded sample meet the preset physical constraints or engineering experience range.

[0077] Expanded samples that do not meet the preset verification conditions can be removed or regenerated.

[0078] Step S4.4: Construct the enhanced sample set, as follows: The augmented samples that pass the validity check are merged with the training sample set to construct the enhanced sample set. Let the real sample set be denoted as... The expanded sample set is denoted as Then enhance the sample set It can be represented as: (33) in, Used for subsequent scene recognition model training.

[0079] Step S4.5: Output the augmented sample set: The completed enhanced sample set is output and used as the input data basis for subsequent scene recognition models to improve the learning ability, classification accuracy and generalization ability of scene recognition models for minority class samples.

[0080] S5. Train a scene recognition model based on the enhanced sample set, and use the trained scene recognition model to identify extreme weather scenes in the samples to be identified, and output the recognition results: Step S5.1: Construct a scene recognition model: Construct a classification model for extreme weather scene recognition. Preferably, the classification model is a convolutional neural network model, used to extract local temporal features and cross-variable correlation features from multivariate time-series samples. The convolutional neural network model includes at least two convolutional layers, at least two pooling layers, and at least one fully connected output layer.

[0081] Step S5.2: Input the enhanced sample set and perform feature extraction, as follows: The enhanced sample set obtained in step 4 is input into the scene recognition model. Let the first part of the enhanced sample set be... Each sample is represented as The feature representation extracted by the classification model can then be denoted as: (34) in, This represents the feature extraction mapping function in the scene recognition model. This represents the high-dimensional feature representation of the corresponding sample.

[0082] Step S5.3: Output the scene category results, as follows: Feature representation The input to the classification output layer yields predictions of which samples belong to different weather scene categories. Let the classification output layer be denoted as... Then we have: (35) in, Indicates the first The predicted category label or category probability output for each sample. When the scene recognition task is a binary classification task, the prediction result is used to distinguish between the target extreme weather scene and the non-target scene; when the scene recognition task is a multi-class classification task, the prediction result is used to distinguish between different types of extreme weather scenes.

[0083] Step S5.4: Train the scene recognition model, as follows: The scene recognition model is trained under supervision using samples from the augmented sample set and their corresponding labels. Preferably, the cross-entropy loss function is used as the classification loss function, denoted as: (36) in, Indicates the number of training samples. Indicates the number of categories. Indicates the first Does the sample belong to the ? The true label of the class, This represents the corresponding predicted probability. By minimizing the classification loss function, the parameters of the scene recognition model are updated, enabling the model to learn discriminative features to enhance the sample set.

[0084] Step S5.5: Output extreme weather scene recognition results. Apply the trained scene recognition model to the enhanced sample set or the sample set to be recognized, and output the corresponding extreme weather scene recognition results. The recognition results may include one or more of the following: sample category label, category probability, recognition accuracy, precision, and recall, to characterize the application effect of the method in extreme weather scene recognition in power systems.

[0085] Step S5.6: Preferably, the convolutional neural network model includes two convolutional layers, each followed by a two-dimensional max-pooling layer with a kernel size of 3. The pooled feature map is expanded and input into a linear layer, outputting the classification result corresponding to the extreme weather scenario. All comparative recognition experiments use the same convolutional neural network architecture to ensure the comparability of recognition results under different augmentation sample schemes.

[0086] Next, referring to the accompanying drawings, the technical effects of the improved technology involved in this application will be further explained in detail: like Figure 6 As shown, the real extreme weather samples exhibit specific temporal distribution characteristics across four dimensions: load, temperature, wind speed, and solar irradiance. Specifically, the load and temperature curves show relatively clear intraday variation patterns, wind speed distribution is highly volatile, while solar irradiance exhibits a relatively concentrated unimodal variation. These distributions collectively constitute the multivariate temporal pattern of the real extreme weather samples.

[0087] like Figure 7 As shown, comparing the expanded samples generated by the method of this application with real samples and samples generated by other generative models reveals that when using a traditional single-generator model, the expanded samples tend to concentrate in a narrow distribution range, especially for key variables such as load, making it difficult to fully cover the variation range of the real samples and exhibiting obvious pattern concentration. In contrast, the small-sample generation method based on the multi-generator-evaluator collaborative mechanism described in this application generates expanded samples whose distributions in multiple dimensions such as load, temperature, wind speed, and solar irradiance can better approximate the distribution of the real samples, and while maintaining a consistent overall trend, it also demonstrates richer sample diversity.

[0088] Furthermore, by Figure 7As can be seen, the expanded samples generated in this application can not only cover the high-probability areas in the real sample distribution, but also form a reasonable expansion in the marginal areas that are less covered by the real samples, thereby improving the ability of the expanded samples to represent the distribution of real extreme weather scenarios. This shows that this application can effectively reduce the model collapse phenomenon and improve the diversity and distribution coverage of the expanded samples through collaborative learning among multiple generators and the constraint effect of the evaluator on similar samples.

[0089] therefore, Figure 6 and Figure 7 This demonstrates that the present application can effectively maintain the distribution characteristics of real extreme weather samples under small sample conditions, reducing the problems of single sample patterns and insufficient distribution coverage in traditional generation methods, thus verifying the effectiveness of the present application in small sample generation for extreme weather scenarios.

[0090] Example 2: A computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the method for generating and identifying small samples of extreme weather scenarios in power systems based on a multi-generator-evaluator collaborative mechanism as described in the example.

[0091] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations may be made to the above embodiments. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for generating and identifying small samples of extreme weather scenarios in power systems based on a multi-generator-evaluator collaborative mechanism, characterized in that, include: Historical power system operation data and meteorological observation data are acquired to construct a real sample set of extreme weather scenarios, and the real sample set is preprocessed. A few-shot generation model is constructed by setting up a discriminator, an evaluator independent of the discriminator, and at least two generators that work together through a pre-defined connection structure. The few-shot generation model is trained through a joint training mechanism. During training, a discriminator is used for discrimination constraints, and an evaluator is used for similarity measurement constraints to suppress mode collapse. An expanded sample set is generated using the trained small sample generation model, and the expanded sample set is fused with the real sample set to obtain an enhanced sample set. The scene recognition model is trained based on the enhanced sample set, and the trained scene recognition model is used to identify extreme weather scenes of the samples to be identified, and the recognition results are output.

2. The method for generating and identifying small samples of extreme weather scenarios in power systems based on a multi-generator-evaluator collaborative mechanism according to claim 1, characterized in that, The steps of using a real sample set and preprocessing it include: The acquired historical power system operation data and meteorological observation data were filtered to construct an extreme weather scenario dataset; The extreme weather scenario dataset is sliced ​​to obtain the real sample set; Preprocessing is performed on the real sample set by at least one of missing value processing, outlier processing, and normalization processing. The preprocessed real sample set is divided into a training sample set and a test sample set. The training sample set is used to train the few-shot generation model, and the test sample set is used to validate the few-shot generation model.

3. The method for generating and identifying small samples of extreme weather scenarios in power systems based on a multi-generator-evaluator collaborative mechanism according to claim 2, characterized in that, The steps for generating the augmented sample set include: Random latent variables are generated through the latent space input module in the small sample generation model; Each generator generates augmented samples based on the random latent variable mapping, ensuring that the dimensions of the augmented samples are consistent with those of the real samples; The discriminator receives samples from the real sample set and generated expanded samples, determines the source of the samples, and applies discriminative constraints to the expanded samples to make them approximate the real samples and ensure that the two are distributed in a consistent manner. The evaluator imposes similarity measurement constraints on the expansion samples and on the expansion samples and the real samples; After completing the discriminant constraints and similarity measurement constraints, an expanded sample set is formed.

4. The method for generating and identifying small samples of extreme weather scenarios in power systems based on a multi-generator-evaluator collaborative mechanism according to claim 3, characterized in that, The discriminator performs the following steps to determine the constraints: After receiving samples from the real sample set and the generated expanded samples, the discriminator outputs the probability value that the sample belongs to the real distribution. Based on the probability values, a discriminator loss function and a generator adversarial loss function are constructed; By minimizing the discriminator loss function, the discriminator's ability to distinguish between the real sample set and the augmented sample set is improved. By minimizing the generator adversarial loss function, the generator is optimized to make the augmented sample distribution approximate the true sample distribution.

5. The method for generating and identifying small samples of extreme weather scenarios in power systems based on a multi-generator-evaluator collaborative mechanism according to claim 4, characterized in that, The steps of the evaluator in using cosine similarity to measure sample similarity include: Calculate the mean sample of the augmented sample set and the mean sample of the true sample set respectively; Based on two mean samples, calculate the mean cosine similarity between each augmented sample and the mean sample of the augmented sample set, and the mean cosine similarity between each real sample and the mean sample of the real sample set. Based on the deviation between the two sets of mean cosine similarity, a similarity score is determined to characterize the diversity between the augmented sample set and the real sample set. The generator is subjected to diversity constraints based on the similarity score, so that the diversity of the expanded sample set is close to that of the real sample set.

6. The method for generating and identifying small samples of extreme weather scenarios in power systems based on a multi-generator-evaluator collaborative mechanism according to claim 5, characterized in that, Based on the generator adversarial loss function and the similarity score, a joint loss function is constructed; By minimizing the joint loss function, the samples generated by each generator can approximate the real samples. The steps for jointly training the few-shot generation model are as follows: By fixing the parameters of each generator and evaluator, the parameters of the discriminator are updated using real and expanded samples to improve the discriminator's ability to distinguish between real and expanded samples. The parameters of the discriminator and each generator are fixed, and the parameters of the evaluator are updated according to the similarity scores between the output samples of each generator to enhance the evaluator's ability to represent the redundant patterns of the expanded samples. By fixing the discriminator and evaluator parameters, and updating multiple generator parameters based on the joint loss function, the generator can improve the authenticity of samples while reducing the redundancy of output. Repeat the parameter update steps in sequence, and continuously iterate and optimize the discriminator, evaluator and generators until the preset number of training rounds or the joint loss function converges.

7. The method for generating and identifying small samples of extreme weather scenarios in power systems based on a multi-generator-evaluator collaborative mechanism according to claim 3, characterized in that, The latent variables are input into the trained small sample generation model, and each generator outputs candidate augmented samples. Candidate augmentation samples are selected or fused according to preset rules to obtain an augmentation sample set; Based on the aforementioned test sample set, the effectiveness of the expanded sample set is verified; The augmented samples that pass the validity verification are merged with the training sample set to obtain the enhanced sample set.

8. The method for generating and identifying small samples of extreme weather scenarios in power systems based on a multi-generator-evaluator collaborative mechanism according to any one of claims 1-7, characterized in that, The steps for training a scene recognition model and performing extreme weather scene recognition include: Based on the enhanced sample set, the scene recognition model extracts features from the enhanced sample set; Based on the extracted features, prediction results are obtained for the corresponding samples, and these prediction results are used to distinguish extreme weather scenarios. Based on the enhanced sample set and the prediction results, the scene recognition model is trained using the cross-entropy loss function; Based on the trained scene recognition model, it is applied to the test sample set and the sample set to be recognized to output the extreme weather scene recognition results.

9. The method for generating and identifying small samples of extreme weather scenarios in power systems based on a multi-generator-evaluator collaborative mechanism according to any one of claims 1-7, characterized in that, The scene recognition model is a convolutional neural network model, which includes at least two convolutional layers, at least two pooling layers, and at least one classification output layer; The connection structure can be any one of the following: parallel connection, series connection, or hierarchical connection.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the method for generating and identifying small samples of extreme weather scenarios in power systems based on a multi-generator-evaluator collaborative mechanism as described in any one of claims 1-9.

Images