Multi-extreme meteorological high-risk scene set generation method based on joint training generative adversarial network

Abstract

The invention relates to the technical field of energy meteorology and intelligent power grids, in particular to a multi-extreme meteorological high-risk scene set generation method, system and equipment based on a joint training generative adversarial network. The method comprises the following steps: constructing a physical information generative adversarial network framework comprising a generator, a discriminator, a predictor and a physical constraint module; designing a multi-objective loss function fusing adversarial loss, prediction loss, physical consistency loss and task performance loss; adopting a training strategy combining meta-learning initialization and incremental learning to jointly optimize parameters of the generator and the discriminator in stages; extreme risk scene data of specified disaster types, seasons and intensity grades are generated through condition vector control, and the extreme risk scene data are stored in a high-risk scene library after physical consistency verification. Through the method, a multi-extreme-weather high-risk scene set with statistical authenticity, physical rationality and task correlation can be directly generated, and the risk identification, scheduling optimization and toughness evaluation capabilities of the clean energy base under extreme weather conditions are remarkably improved; the problems of sample scarcity, model overfitting and lack of physical constraints in scene generation in the prior art are solved, efficient and automatic generation of a high-risk scene is realized, and reliable data support is provided for power grid toughness evaluation and scheduling decision of a clean energy base.

Description

Technical Field

[0001] This invention relates to the field of energy meteorology and smart grid risk management technology. It addresses the problem of constructing high-risk scenario sets for clean energy bases, mainly wind and photovoltaic, under extreme weather conditions such as extreme heat and no wind, extreme cold and no light, typhoons, cold waves, and floods. It proposes a direct generation method for high-risk scenario sets of multiple extreme weather conditions based on jointly trained generative adversarial networks, which significantly improves the performance of prediction models under extreme scenarios. Background Technology

[0002] The inherent randomness, intermittency, and volatility of clean energy output make it highly dependent on weather conditions, leading to profound changes in the operating characteristics of the power system. Frequent extreme weather events, such as typhoons, cold waves, extreme heat waves, and sandstorms, can easily trigger high-risk scenarios such as large-scale power output drops at clean energy bases, forced equipment shutdowns, and even partial grid disconnections, posing a serious threat to the safe, stable operation of the power system and reliable power supply.

[0003] Against this backdrop, constructing a high-quality set of scenarios that accurately reflects the operational risks of systems under extreme weather conditions has become a crucial prerequisite for conducting power grid resilience assessments, developing disaster prevention and mitigation plans, and optimizing dispatching and operation strategies. However, existing technologies face several core bottlenecks in constructing high-risk scenario sets:

[0004] 1. The scarcity and imbalance of samples in extreme scenarios are prominent issues: Extreme weather events are inherently low-probability events, and the statistical sample size based on historical observation data is severely insufficient. Furthermore, in training data-driven machine learning models, the number of samples under normal and extreme conditions differs by several orders of magnitude. This severe sample imbalance easily leads to overfitting of the model to normal scenarios, while resulting in poor predictive performance for high-risk scenarios. Consequently, the model's generalization ability is insufficient when facing extreme events in practical applications, leading to ineffective early warnings.

[0005] 2. The limitations of traditional scene generation methods are becoming increasingly apparent: Currently, mainstream methods can be divided into two categories: First, numerical simulation methods based on physical mechanisms, such as coupling meteorological models like WRF with power system simulation software. This method suffers from high computational costs and long simulation cycles, making it difficult to quickly generate massive amounts of scenes. Furthermore, its accuracy heavily depends on the accuracy of model parameter settings, and the interaction between the complex underlying surface and the atmospheric boundary layer introduces significant uncertainties, potentially leading to large deviations in simulation results. Second, data-driven methods based on statistical learning, such as traditional Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs). While these methods can learn the distribution of historical data and generate new samples, they are essentially a "black box" of statistical replication, and the physical plausibility of the generated scenes cannot be guaranteed.

[0006] 3. Spatiotemporal differences: Extreme weather characteristics vary significantly across different regions and seasons, and direct transfer of models can lead to a decline in generalization performance.

[0007] Therefore, given the practical challenges of a small number of existing samples and numerous uncertainties in model construction, how to combine the advanced ideas of joint training GANs with the physical constraints in the field of energy meteorology to develop a new method that can directly generate a set of high-risk, multi-extreme weather scenarios for clean energy bases that combines statistical realism, physical rationality, and task efficiency has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0008] This invention provides a method, system, and device for generating sets of high-risk multi-extreme weather scenarios based on jointly trained generative adversarial networks, which can effectively solve the problems in the background technology.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0010] A method for generating a set of high-risk multi-extreme weather scenarios based on jointly trained generative adversarial networks, the method comprising:

[0011] A physical information generative adversarial network framework is constructed, which includes a generator, a discriminator, a predictor, and a physical constraint module;

[0012] Design a multi-objective loss function, which includes adversarial loss, prediction loss, physical consistency loss, and task performance loss;

[0013] A training strategy combining meta-learning initialization and incremental learning is adopted to train the parameters of the network framework in stages;

[0014] Extreme risk scenarios of a specified type are generated through conditional control, and the generated scenarios are physically verified. Scenarios that pass verification are stored in a high-risk scenario library.

[0015] Furthermore, a framework for generative adversarial networks based on physical information is constructed, including:

[0016] The generator adopts a conditional generative adversarial network structure, with random noise vector and conditional vector as inputs and multivariate time series data as outputs;

[0017] The discriminator adopts a multi-scale one-dimensional convolutional neural network structure, which contains three parallel convolutional channels to process features at different time scales.

[0018] The predictor employs a physical information long short-term memory network, adding a physical model calculation channel to the traditional LSTM.

[0019] The physical constraint module transforms domain knowledge into optimization objectives through mathematical formulas, including power curve equations, the law of conservation of energy, and variable threshold range constraints.

[0020] Furthermore, the generator's condition vector includes disaster type code, seasonal label, and intensity level information, and achieves multi-dimensional condition control through a combination of one-hot encoding and numerical encoding;

[0021] The discriminator not only determines the authenticity of the data, but also outputs a physical constraint satisfaction vector to evaluate the rationality of the generated scene in multiple physical dimensions.

[0022] The predictor's input includes historical time-series data and intermediate variables calculated by the physical model, and its output is the predicted value for the next moment, ensuring that the temporal evolution of the generated scene conforms to physical laws.

[0023] Furthermore, a multi-objective loss function is designed, including:

[0024] The total loss function is a weighted sum of adversarial loss, prediction loss, physical consistency loss, and task performance loss, and its mathematical expression is:

[0025]

[0026] in, The loss is based on Wasserstein distance and includes gradient penalty. To predict the loss, the mean squared error is used to calculate the difference between the predicted value and the actual value; The physical consistency loss includes both equality constraint loss and inequality constraint loss. This represents the task performance loss, used to constrain the effectiveness of the generated samples in downstream prediction or risk assessment tasks. This is the loss weighting coefficient.

[0027] The adversarial loss is based on Wasserstein distance and supplemented by gradient penalty, which enhances training stability by constraining the gradient norm of the discriminant function to be close to 1.

[0028] The prediction loss is calculated using mean squared error to determine the difference between the predictor output value and the true value, ensuring the temporal dynamic consistency of the generated scene.

[0029] Furthermore, the physical consistency loss includes equality constraint loss and inequality constraint loss;

[0030] The equality constraint loss ensures that the generated data conforms to known physical laws, including verifying the consistency between the generated power and the theoretical power through the power curve equation;

[0031] The inequality constraint loss ensures that the generated variables are within a physically possible reasonable range, and generates a secondary penalty when the variables exceed the allowed threshold.

[0032] Furthermore, the task performance loss is calculated through a pre-trained downstream task model, guiding the generator to produce data that has direct value for power grid security assessment, scheduling decisions, and other business operations.

[0033] The downstream task model includes a short-term power prediction model or a system security assessment model. By minimizing the difference between the task model output and the expected target, the business usability of the generated scenario is optimized.

[0034] Furthermore, a training strategy combining meta-learning initialization and incremental learning is adopted, including:

[0035] The meta-learning initialization phase employs a meta-learning algorithm to find a good set of initialization parameters, enabling the model to quickly adapt to new types of extreme scene generation tasks.

[0036] During the joint training phase, the parameters of the generator and discriminator are optimized alternately. When the discriminator is fixed, the generator is jointly optimized through adversarial loss, physical consistency loss and task performance loss. When the generator is fixed, the discriminator is optimized only through adversarial loss.

[0037] The incremental learning phase employs a dynamic network expansion strategy. When new types of extreme event data are obtained, the model parameters are updated through an elastic weight consolidation method to avoid forgetting existing knowledge.

[0038] Furthermore, by using conditional controls, specific types of extreme risk scenarios can be generated, including:

[0039] By adjusting the disaster type code, seasonal label, and intensity level in the condition vector, the generator is driven to produce the target extreme scenario;

[0040] The generated scenario is subjected to automated physical verification, and the verification rules include power balance verification, energy conservation verification, and operational constraint verification.

[0041] The power balance verification requires that the deviation between the total power generation, load power, and grid loss be less than a set threshold.

[0042] The energy conservation verification requires that the changes in the system's energy input, output, and storage within a specific time period satisfy the conservation relationship.

[0043] The operational constraint verification requires all variables to be within the safe operating range, including parameters such as equipment output, bus voltage, and line power flow.

[0044] Furthermore, the verified scenarios will be stored in a high-risk scenario library, including:

[0045] A hierarchical storage structure is adopted, with the first level indexed by disaster type, the second level categorized by season, and the third level divided by intensity level;

[0046] Each scenario is labeled with key metadata, including characteristic parameters such as duration, maximum wind speed, minimum temperature, and maximum output reduction;

[0047] It supports rapid retrieval and retrieval based on multiple dimensions such as disaster type, season, and intensity level, and can be applied to power grid resilience assessment, disaster prevention plan formulation, and operation mode optimization.

[0048] Furthermore, the generator in the physical information generative adversarial network framework is implemented using a deep convolutional neural network, which generates multivariate time-series data step by step through deconvolutional layers and upsampling layers;

[0049] The discriminator employs a multi-scale feature fusion mechanism, extracting local and global features from time-series data using convolutional kernels of different sizes;

[0050] The predictor takes the theoretical power calculated by the physical model as an additional input and controls the fusion ratio of physical information and data-driven information through a gating mechanism.

[0051] A system for generating sets of high-risk, multi-extreme weather scenarios based on jointly trained generative adversarial networks, the system comprising:

[0052] Network construction unit, used to construct the physical information generation adversarial network framework;

[0053] A loss design unit is used to design the multi-objective loss function;

[0054] The training execution unit is used to implement the meta-learning initialization, joint training, and incremental learning strategies.

[0055] The scenario generation unit is used to generate and verify extreme risk scenarios and manage the high-risk scenario library;

[0056] A device for generating a set of high-risk multi-extreme weather scenarios based on a jointly trained generative adversarial network is used to implement the method for generating a set of high-risk multi-extreme weather scenarios based on a jointly trained generative adversarial network.

[0057] The technical solution of this invention can achieve the following technical effects:

[0058] By employing a joint training framework that integrates physical information embedding and task performance orientation, extreme risk scenarios with both statistical realism and physical plausibility are generated. A conditional generation mechanism is used to achieve on-demand generation of multiple types and levels of scenarios. Meta-learning and incremental learning strategies are used to enhance the model's generalization ability. A complete scenario verification and management system is established to provide high-quality data support for the grid safety assessment of clean energy bases.

[0059] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0060] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0061] Figure 1 This is a flowchart illustrating a method for generating a set of high-risk multi-extreme weather scenarios based on a jointly trained generative adversarial network.

[0062] Figure 2 Generate an adversarial network framework diagram for joint training;

[0063] Figure 3 Flowchart of meta-learning and incremental training strategies;

[0064] Figure 4 Generate and verify storage flowcharts for extreme scenarios. Detailed Implementation

[0065] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0066] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0067] according to Figure 1 A flowchart illustrating a method for generating a set of high-risk multi-extreme weather scenarios based on jointly trained generative adversarial networks. This application provides a method for generating a set of high-risk multi-extreme weather scenarios based on jointly trained generative adversarial networks, the method including:

[0068] S10: Collect multi-source meteorological and energy output data of the target area, and perform time alignment, cleaning, normalization and feature encoding on it;

[0069] S20: Construct a jointly trained generative adversarial network that includes a generator G, a discriminator D, a predictor P, and a physical constraint module M;

[0070] S30: Design of multi-objective loss function and physical constraint module;

[0071] S40: Training is performed using a joint optimization mechanism that combines meta-learning initialization with incremental learning;

[0072] S50: Generates extreme risk scenarios of a specified type through conditional control;

[0073] S60: The generated scene is checked by the physical verification module. If the verification is successful, it is stored in the high-risk scene library according to the code.

[0074] Specifically, the process begins by collecting and aggregating multi-source meteorological observation data, remote sensing data, and clean energy output and equipment operation data for the target area to construct a unified time-series sample set and conditional label set. Then, the raw data is processed, and conditional vectors are constructed using information such as disaster type, season, and intensity level for subsequent model calls. Next, a physically-enhanced generative adversarial network (GAN) system is built, consisting of a generator, discriminator, predictor, and physical constraint module. The generator takes random noise and conditional vectors as input and generates multivariate time-series samples through deep convolution and recurrent units. The discriminator simultaneously scores the sample authenticity and physical consistency under a multi-scale convolutional structure. The predictor embeds physical model calculation channels into a traditional time-series prediction network to maintain the physical rationality of the time-series evolution. The physical constraint module transforms knowledge in areas such as power curves, energy conservation, and variable boundaries into differentiable penalty terms and feeds them back to the training phase. Subsequently, meta-learning is used to initialize the model, enabling it to obtain a parameter starting point with rapid adaptability. Finally, joint training integrates adversarial loss, prediction loss, and physical consistency. Losses and task performance losses are weighted and jointly optimized. An alternating update strategy is adopted to form adversarial and collaborative training dynamics between the generator and discriminator, while improving the predictor's time-series prediction capability in the generation-discrimination closed loop. Then, after the model stabilizes, an incremental learning mechanism is used to introduce new disaster types or new seasonal samples. Network expansion and regularization are used to avoid forgetting between disaster types and maintain the model's continuous learning capability. Next, in the generation stage, conditional vectors are used to control the generation of extreme weather time-series samples under specified disaster types, and the consistency between the generated power data and the theoretical power calculated by the physical model is verified. Only scenarios that meet power balance, energy conservation, and operational constraints can be included in the database. The verified scenarios are hierarchically labeled and stored. A high-risk scenario database is established according to disaster type, season, and intensity level, and metadata is added to each scenario to support multi-dimensional retrieval and task invocation. Furthermore, the scenario database is linked with downstream power grid simulation, scheduling optimization, and resilience assessment systems to support disaster prevention and mitigation decision-making and operational optimization, thereby realizing a closed-loop workflow from data acquisition to scenario generation and then to engineering application.

[0075] The technical solution of this invention emphasizes the dual embedding of physical information and task orientation throughout the entire process to solve the problem that statistical learning alone cannot guarantee physical rationality. Meta-learning and incremental learning ensure the generalization and sustainable learning capabilities across disaster types and regions, so that the generated high-risk scenarios not only meet the diversity of statistical features but also have the constraints of engineering usability. This provides a reliable data foundation and model support for risk identification, resilience assessment and scheduling decisions of clean energy bases under extreme weather conditions.

[0076] Furthermore, constructing a standardized time-series dataset includes:

[0077] Historical meteorological data for the target area was collected from multiple sources, including meteorological observation stations, satellite remote sensing, and numerical weather prediction models. This data included variables such as wind speed, temperature, irradiance, humidity, and air pressure. Corresponding clean energy output data, such as wind power and photovoltaic power, were also acquired. Data collection needed to cover at least 10 years of historical records to ensure the inclusion of samples from various extreme weather events. After collection, the data underwent preprocessing operations such as time alignment, outlier removal, missing value imputation, and standardization. Conditional variables such as season, disaster type, and intensity level were then encoded using one-heat or numerical encoding to form a standardized time-series dataset.

[0078] Furthermore, the framework for constructing a physical information generative adversarial network mainly includes a generator, a discriminator, a predictor, and a physical constraint module:

[0079] The generator is used to generate multivariate time-series data based on a random noise vector and a conditional vector. Its structure employs a conditional generative adversarial network (GAN) architecture. The inputs are a random noise vector z and a conditional vector c, and the output is multivariate time-series data of length k. The conditional vector contains disaster type, seasonal label, and intensity level information, mathematically represented as: , where z is the noise vector, c is the condition vector, and X̂ is the generated multivariate time series data.

[0080] The discriminator is used to determine the authenticity and physical plausibility of the input data. It employs a multi-scale one-dimensional convolutional neural network structure, containing three parallel convolutional channels, corresponding to short-term local fluctuation characteristics, medium-term change trends, and long-term seasonal characteristics, respectively. Its discrimination process is as follows: The discriminator generates two sub-vectors at the output: one outputs the result of determining whether the data is true or false, and the other outputs a vector indicating whether the physical constraints are satisfied. It is used to evaluate the physical plausibility of the generated scene and also to evaluate the performance of the generated data in downstream tasks. Its loss function is:

[0081]

[0082] Where n represents the number of training samples. It is the i-th real data of length k. Indicates that the input is The output of the discriminator.

[0083] The predictor learns the temporal evolution patterns between meteorological variables. It employs a Physical Information Long Short-Term Memory (PI-LSTM) network structure, adding a physical model computation channel to the traditional LSTM. Through a gating mechanism, it fuses the calculated values ​​from the theoretical power equation with historical time-series characteristics to predict the meteorological state or power output level at the next moment. The predictor's input includes historical time-series data and intermediate variables calculated by the physical model; the output is the predicted value for the next moment. Its network structure is represented as follows: , where Φ(·) is the physical model calculation function.

[0084] Furthermore, a loss function system integrating multiple optimization objectives is designed to balance the joint training process among the generator, discriminator, and predictor, including:

[0085] The total loss function is composed of the mean-weighted sum of the following four parts:

[0086]

[0087] in, , , To balance the weighting coefficients of various loss factors, multiple optimizations are achieved in terms of statistical authenticity, physical rationality, and task applicability.

[0088] The adversarial loss component, primarily used to improve the training stability of the generative adversarial network, employs a Wasserstein distance-based approach supplemented with gradient penalty, calculated as follows:

[0089]

[0090] in, This represents the average output score of the discriminator D for all real samples x in a batch. The discriminator prefers this value to be as large as possible. This indicates that the discriminator D evaluates all generated samples in a batch. The average of the output scores is what the discriminator wants to be as small as possible; This represents the gradient penalty term, used to constrain the gradient norm of the discriminant function to be close to 1, thereby enhancing training stability. , It is a random number uniformly sampled between 0 and 1, ∇ represents the gradient, and ‖⋅‖2 represents the L2 norm. It is the weight coefficient of the gradient penalty term.

[0091] For the prediction loss component, this part is mainly used to constrain the deviation between the predictor output and the true value, and its average value is calculated using the mean squared error (MSE):

[0092]

[0093] This computational predictor is based on historical sequences. Predicted next time value , and the actual next time value The batch average of the squared differences between the predicted and actual time series is used to maintain consistency between the temporal evolution of the generated scene and the actual data by comparing the differences between the predicted results and the actual time series.

[0094] As for the physical consistency loss, this part is mainly used to ensure that the generated data satisfies energy conservation and physical boundary conditions, including equality constraint loss and inequality constraint loss:

[0095]

[0096] Equivalent loss : Ensure that the generated data follows known physical laws, such as power curves.

[0097]

[0098] in, , These represent the generated wind power and solar power data, respectively. , The theoretical power, calculated from the generated meteorological data (wind speed w, irradiance GHI) based on the physical model, is the batch average of the squared difference between the generated power and the theoretical power.

[0099] Inequality loss By imposing a secondary penalty on out-of-bounds variables, we ensure that the generated variables are within a physically possible and reasonable range.

[0100]

[0101] When variable w exceeds its maximum allowed value or below the minimum allowable value When this happens, a penalty will be imposed. ... This indicates that other variables, such as temperature and irradiance, can be added as constraints; the max(0, ⋅) function ensures that a loss only occurs when a variable goes out of bounds, and obtains the batch average of the squares of all out-of-bounds penalty terms.

[0102] Regarding the performance loss of the task, this part is mainly used to ensure that the generated data can effectively improve the performance of downstream tasks, such as power prediction or security assessment:

[0103]

[0104] Where T(⋅) represents a pre-trained downstream task model; It is the downstream task model that generates data The predicted output; It is the expected target output that meets actual business needs. This item calculates the batch average of the squared differences between the downstream task model's predicted output of the generated data and the target output, aiming to guide the generator to produce data that is useful for downstream tasks.

[0105] Furthermore, a joint optimization mechanism combining meta-learning initialization and incremental learning is employed for training, including:

[0106] The process consists of four main stages: meta-learning initialization, joint training, incremental learning, and scenario generation and verification.

[0107] In the meta-learning initialization phase, the network is pre-trained using the Model-Independent Meta-Learning (MAML) algorithm. Through meta-task learning on various extreme weather events, such as typhoons, high temperatures, cold waves, and torrential rains, the model acquires rapid adaptability across different disaster types. The goal of this phase is to find a set of initial parameters. This allows the model to achieve stable performance with only a small number of samples when facing new types of extreme scenarios, mathematically expressed as:

[0108]

[0109] Where θ represents the initial parameters of the model; This represents the i-th task sampled from the task distribution; For the model in Losses; The model represents the The parameters are updated after k steps. This process achieves parameter sharing between tasks through gradient updates between the outer and inner layers, giving the initialized model a good starting point for generalization and significantly reducing the convergence time in the subsequent joint training phase.

[0110] Furthermore, during the joint training phase, an adversarial balance is achieved by alternately optimizing the parameters of the generator, discriminator, and predictor, including:

[0111] During training, the discriminator parameters are first fixed, and the generator and predictor are updated to simultaneously minimize adversarial loss, physical consistency loss, and task performance loss. Then, the generator parameters are fixed, and the discriminator is independently optimized to maximize the adversarial loss, thereby enhancing its ability to distinguish between real and fake samples. The parameter update process can be represented as follows:

[0112]

[0113]

[0114] in These are the generator parameters; These are the parameters for the discriminator; The learning rate of the generator; The learning rate of the discriminator; This indicates that the gradient is calculated with respect to the generator parameters; This indicates that the gradient is calculated with respect to the discriminator parameters; , These are the weights for physical consistency loss and task performance loss, respectively.

[0115] Through repeated, alternating optimization, the generator's output samples gradually approximate the real data distribution while maintaining consistency under physical constraints. After training stabilizes, the discriminator and predictor reach an equilibrium, and the generator can autonomously generate multidimensional time-series data that conforms to specific weather types and intensity levels.

[0116] Furthermore, in the incremental learning phase, a strategy combining dynamic network expansion and regularization constraints is adopted to address the introduction of new types of extreme event data, including:

[0117] When a new disaster category or a new meteorological distribution sample is detected, the model adds a task-specific branch while maintaining the original structure, and prevents the disaster type forgetting effect through parameter regularization. Its parameter update can be expressed as:

[0118]

[0119] in, These are the original parameters of the model; The updated parameters; ε is the parameter adjustment amount; ε is the regularization constraint coefficient, which ensures that new data learning does not forget old knowledge, while constraining the parameter adjustment range to prevent overfitting to new data and achieve long-term stable knowledge accumulation.

[0120] Furthermore, during the scenario generation and verification phase, specified types of extreme risk scenarios are generated through conditional control, including:

[0121] After joint training, the generative adversarial network can directly generate extreme weather risk scenarios with specified disaster type, season, and intensity level after inputting a random noise vector z and a conditional vector c.

[0122]

[0123] in Code the type of disaster. For seasonal labels, Indicating intensity level.

[0124] The generated samples enter the physical verification module for rationality verification, including three aspects: power balance verification, energy conservation verification, and operational constraint verification.

[0125] Power balance verification requires that the deviation between the system's power generation, load power, and grid losses at any given time does not exceed a set threshold. The specific formula is as follows:

[0126]

[0127] in, Total power generation Total load power, For system network loss power, This is the tolerance threshold.

[0128] The energy conservation verification targets a time interval. To ensure a balance between system energy input and output, the following requirements must be met:

[0129]

[0130] in, For input power, For output power, To constrain constants;

[0131] Operational constraint verification checks whether variables such as wind speed, temperature, irradiance, power output, and voltage are all within the safe operating range:

[0132]

[0133] in, For each physical quantity, Lower bound of variable The upper limit of the variable;

[0134] Once the generated data meets the above physical constraints, it can be considered a physically reasonable sample of a high-risk meteorological scenario.

[0135] Furthermore, the verified scenarios will be stored in a high-risk scenario library, including:

[0136] A hierarchical management approach is adopted, indexed according to disaster type, season, and intensity level. The first level index is the disaster category, such as typhoon, cold wave, extreme heat, rainstorm, and sandstorm; the second level index is the season label, divided into four categories: spring, summer, autumn, and winter; the third level index is the intensity level, ranging from level I to level IV.

[0137] Each scenario sample is accompanied by metadata information upon being added to the database, including indicators such as maximum wind speed, minimum temperature, maximum output reduction, duration, and frequency of occurrence, facilitating subsequent retrieval and retrieval. The high-risk scenario database supports multi-dimensional tag queries and can directly serve business applications such as power grid resilience assessment, disaster prevention and mitigation plan development, and operation mode optimization.

[0138] Furthermore, the jointly trained generative adversarial network system in this embodiment can achieve highly parallel training and automated optimization under large-scale sample conditions. During training, batch normalization and adaptive learning rate adjustment mechanisms are employed to dynamically control the gradient update magnitude and prevent pattern collapse and gradient vanishing problems. Simultaneously, an experience replay buffer is introduced to periodically mix new and old samples for training, maintaining the stability of the sample distribution. Through these methods, the model can ultimately generate a set of high-risk, multi-extreme weather scenarios that possess statistical realism, physical plausibility, and task relevance, providing high-quality input for grid resilience assessment of clean energy bases.

[0139] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely exemplary illustrations of the application as defined herein, and are to be considered as covering any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Thus, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A method for generating a set of high-risk multi-extreme weather scenarios based on jointly trained generative adversarial networks, characterized in that, The method includes: Collect multi-source historical meteorological data and corresponding clean energy output data for the target area; A joint training generative adversarial network framework is constructed, which includes a generator, a discriminator, a predictor, and a physical constraint module; Design a multi-objective loss function consisting of adversarial loss, prediction loss, physical consistency loss, and task performance loss; A training strategy combining meta-learning initialization and incremental learning is adopted to train the parameters of the network framework in stages; Input random noise vector z and condition vector c, and generate multivariate time series extreme weather scene data through a generator; The generated data will be verified for power balance, energy conservation, and variable constraints by the physical consistency verification module. Verified data is stored in a high-risk scenario database, forming a collection of high-risk extreme weather scenarios categorized by disaster type and intensity level.

2. The method according to claim 1, characterized in that, The framework for constructing a jointly trained generative adversarial network includes: The generator G adopts a conditional generative adversarial network structure, with random noise vector z and conditional vector c as inputs and multivariate time series data as outputs. The discriminator D adopts a multi-scale one-dimensional convolutional neural network structure to determine the authenticity of data and output a physical constraint satisfaction vector. The predictor P employs a physical information long short-term memory network. Its inputs include historical time-series data and intermediate variables calculated by the physical model, and its output is the meteorological variable or energy output at the next moment. The physical constraint module M is used to calculate the deviation between the generated data and the physical laws and feed it back to the total loss function.

3. The method according to claim 2, characterized in that, The condition vector contains disaster type code, seasonal label, and intensity level information.

4. The method according to claim 1, characterized in that, The multi-objective loss function is a weighted sum of the adversarial loss, prediction loss, physical consistency loss, and task performance loss, and its expression is: in, The adversarial loss is based on Wasserstein distance and includes gradient penalty. To predict the loss, the mean squared error is used to calculate the difference between the predicted value and the actual value; The physical consistency loss includes both equality constraint loss and inequality constraint loss. This represents the task performance loss, used to constrain the effectiveness of the generated samples in downstream prediction or risk assessment tasks. This is the loss weighting coefficient.

5. The method according to claim 4, characterized in that, The physical consistency loss includes equality constraint loss and inequality constraint loss; The equality constraint loss This is used to ensure that the generated power data conforms to the physical power curve relationship: in, , These represent the generated wind power and solar power data, respectively. , The theoretical power, calculated from the generated meteorological data (wind speed w, irradiance GHI) based on the physical model, is the batch average of the squared difference between the generated power and the theoretical power. The inequality constraint loss This is used to ensure that the generated variables are within a physically reasonable range of values: in This indicates variables such as wind speed, temperature, and irradiance; ... indicates that other variables can be added similarly.

6. The method according to claim 1, characterized in that, The training strategy includes: In the meta-learning initialization phase, a meta-learning algorithm is used to initialize the model, enabling the model to adapt quickly to various extreme weather tasks. During the joint training phase, the discriminator parameters are fixed and the generator is optimized by combining the adversarial loss, physical consistency loss, and task performance loss. Then, the generator parameters are fixed and the discriminator is optimized by the adversarial loss. The authenticity and physical rationality of the generated data are improved through alternating optimization. During the incremental learning phase, when new types of extreme event data are obtained, a dynamic network expansion strategy is adopted to update the model parameters, and regularization constraints are applied to avoid the disaster amnesia effect caused by disaster migration.

7. The method according to claim 1, characterized in that, The physical verification of the generated scenario includes at least one of power balance verification, energy conservation verification, and operational constraint verification.

8. The method according to claim 1, characterized in that, The high-risk scenario database is stored hierarchically according to disaster type, season and intensity level, and supports tag-based retrieval and task invocation.

9. A system for generating sets of high-risk multi-extreme weather scenarios based on jointly trained generative adversarial networks, characterized in that: The system is used to implement the method as described in any one of claims 1 to 8, and includes: The data acquisition unit is used to acquire multi-source meteorological observation data and energy output data; Network construction unit, used to construct the physical information generation adversarial network framework; A loss design unit is used to design the multi-objective loss function; The training execution unit is used to execute the meta-learning initialization, joint training, and incremental learning strategies. The feature processing unit is used to clean, standardize, and extract time-series features from the raw data. The scene generation unit is used to generate target scene data based on the noise vector z and the condition vector c; The verification storage unit is used to perform physical constraint and power balance verification on extreme risk scenarios, and at the same time saves the set of high-risk extreme weather scenarios that have passed the verification.

10. A device for generating sets of high-risk multi-extreme weather scenarios based on jointly trained generative adversarial networks, characterized in that, The device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method as described in any one of claims 1 to 8.

Images