Source network load operation scene generation method, system and device under ice disaster disturbance based on transfer learning and storage medium

Abstract

The invention discloses a source network load operation scene generation method, system and device under ice disaster disturbance based on transfer learning and a storage medium, and relates to the technical field of power system operation and analysis, and the method comprises the steps: firstly constructing a simulation sample library based on a power grid structure, a fault type and an icing condition, and then integrating and standardizing source domain samples, dividing risk levels, and constructing and pre-training a least square loss function condition generative adversarial network model; a final standardized target domain sample set is generated, a target domain ice disaster scene generator is obtained through transfer learning and model fine tuning, and then an ice disaster disturbance operation scene is generated; the ice disaster disturbance operation scene can be accurately generated, an electric power department can be helped to master risks in advance, equipment maintenance, resource allocation and emergency preparation can be made accordingly, the influence of ice disasters on an electric power system is reduced, power supply stability is guaranteed, power grid planning and transformation can be guided, and the overall disaster resistance and market risk management capacity can be improved.

Description

Technical Field

[0001] This invention relates to the field of power system operation and analysis technology, and in particular to a method, system, device and storage medium for generating source-grid-load operation scenarios under ice storm disturbances based on transfer learning. Background Technology

[0002] Against the backdrop of large-scale integration of new energy sources and frequent extreme weather events, power grid scenario generation technology enhances system resilience by constructing extreme weather scenarios.

[0003] Existing methods are mainly divided into two categories: statistical methods, including Monte Carlo simulation, Latin hypercube sampling, Markov chain method, scenario tree method, and time series method. These methods rely on the statistical regularity of historical data and construct power grid operation scenarios by analyzing information such as the failure probability of distribution network components and power prediction errors. Artificial intelligence methods, mainly based on generative adversarial networks, variational autoencoders, and regularized flow models, utilize the unsupervised learning characteristics of deep learning to achieve diversified scenario construction and multi-parameter joint scenario generation by setting constraints and mapping potential spaces. However, existing research often focuses solely on the power source side, grid side, or load side, or only considers the simple correlation between the two, lacking systematic modeling of the coordinated dynamic response of "source-grid-load," making it difficult to simulate the complex changes of the power system under real disaster scenarios. Traditional statistical methods are based on fixed probability distributions or static temporal relationships for modeling, making it difficult to capture the dynamic evolution of weather systems, spatiotemporal interactions, and the complex fluctuation characteristics of renewable energy output. For example, when generating typhoon disaster scenarios, the failure to consider the temporal correlation and spatial propagation effects of meteorological elements leads to significant deviations between the generated fault scenarios and the actual disaster distribution. In multi-regional wind power output forecasting, insufficient capture of meteorological correlations between regions leads to forecast results that fail to accurately reflect the actual situation. Furthermore, these methods have limited ability to characterize the non-stationary characteristics of sudden increases and decreases in renewable energy power under extreme weather conditions, resulting in inaccurate risk assessments. Deep learning models rely on massive amounts of labeled data for training, while extreme weather data is often scarce and unevenly distributed, easily leading to overfitting in small-sample scenarios. When constructing wind power scenarios, if the training data lacks extreme condition samples such as typhoons and icing, the generated scenarios may not accurately reflect extreme power fluctuations, resulting in "model collapse." Moreover, the black-box nature of artificial intelligence models makes it difficult to interpret the generation logic, hindering the establishment of decision-making trust in scenarios with extremely high reliability requirements, such as grid dispatching. Under extreme weather conditions, the generated scenarios are easily decoupled from actual meteorological conditions, failing to meet practical application needs. Existing scenario generation methods often consider uncertainties on the power supply side or load side in isolation, neglecting the dynamic interaction among the three. Traditional scenario tree methods only expand scenarios based on load growth rates, failing to incorporate the feedback impact of grid component failures on power output and load demand. While joint prediction frameworks can simultaneously generate new energy and load scenarios, they do not consider the constraints of grid topology changes on power transmission, leading to discrepancies between the generated scenarios and the actual system operating state. Especially in icing disasters, existing methods do not couple the dynamic response of line icing growth and melting processes with power output and load demand, making it difficult to assess the cascading impact of the disaster. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by this invention is: addressing the shortcomings of existing power grid scenario generation technologies in terms of dynamic characterization, generalization ability, and source-grid-load collaborative modeling, by using a DSP-pre-trained least squares loss function conditional generative adversarial network model and fine-tuning it with a small number of icing samples, and incorporating a source-grid-load collaborative mechanism, to accurately generate power system operation scenarios under ice disaster disturbances to support risk assessment and scheduling decisions.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a method for generating source-grid-load operation scenarios under ice storm disturbances based on transfer learning, including:

[0008] Based on power grid structure data, typical fault type data, and icing disturbance conditions, a simulation sample library for power system operation scenarios is constructed.

[0009] Based on the power system operation scenario simulation sample library, source domain samples are integrated and standardized to obtain a standardized source domain sample set and its feature matrix.

[0010] Based on the standardized source domain sample set and its feature matrix, the risk level of the source domain samples is divided to obtain a standardized source domain sample set with risk level labels;

[0011] Based on a standardized source domain sample set with risk level labels, a least-squares loss function conditional generative adversarial network model is constructed and pre-trained.

[0012] Based on the standardized target domain sample set and its feature matrix of icing operation samples with risk level labels in the power system operation scenario simulation sample library, the final standardized target domain sample set is generated.

[0013] Based on the pre-trained least squares loss function conditional generative adversarial network model, combined with the final standardized target domain sample set, transfer learning and target domain model fine-tuning are performed to obtain the fine-tuned target domain ice disaster scene generator.

[0014] Based on the finely tuned target domain ice disaster scenario generator, an ice disaster disturbance operation scenario is generated.

[0015] As a preferred solution for a source-grid-load operation scenario generation method based on transfer learning under ice disaster disturbance, wherein:

[0016] The power system operation scenario simulation sample library constructed based on power grid structure data, typical fault type data, and icing disturbance conditions includes:

[0017] A simulation sample library of power system operation scenarios is constructed in the form of a three-dimensional matrix. The simulation sample library of power system operation scenarios consists of a sample set of extreme operation scenarios.

[0018] Based on power grid structure data, typical fault type data, and icing disturbance conditions, sample simulation work is carried out: Based on system safety requirements, the overload ratio increase limit of branch power flow and the constraint limit of node voltage are determined. Based on the predetermined operating mode, the output combination of wind turbines is changed under different new energy penetration rates for simulation. Based on the time-domain simulation waveform generated by the simulation, steady-state node voltage, branch power flow and frequency data are extracted, and the node injected power is further calculated to obtain the operating sample set a.

[0019] The beneficial effects of this preferred technical solution are as follows: Constructing a simulation sample library of power system operation scenarios represented in a three-dimensional matrix format allows for clearer and more accurate organization and management of samples, facilitating subsequent data processing and analysis. By conducting sample simulations based on power grid structure data, typical fault type data, and icing disturbance conditions, and considering the branch power flow overload ratio increase limit and node voltage constraint limit determined for system safety needs, simulations are performed by changing the output combination of wind turbines under different renewable energy penetration rates. This allows for comprehensive acquisition of power system operation samples under various operating conditions, providing a rich and accurate data foundation for subsequent research and contributing to a deeper understanding of the power system's operating characteristics under different circumstances.

[0020] As a preferred solution for a source-grid-load operation scenario generation method based on transfer learning under ice disaster disturbance, wherein:

[0021] The construction of the power system operation scenario simulation sample library based on power grid structure data, typical fault type data, and icing disturbance conditions also includes:

[0022] Set the same constraints as the simulation under normal weather conditions. Based on the predetermined operating mode, determine the typical fault types and write them into the fault text data file. Load them into the same electromechanical transient simulation project. After reasonably setting the maximum simulation parallel threads based on the simulation platform memory space, perform batch simulation. Extract the steady-state node voltage, branch power flow and frequency data from the time-domain simulation waveform generated by the simulation, and further calculate the node injection power to obtain the operating sample set b.

[0023] Based on a predetermined operating mode, simulations are performed in conjunction with icing disturbance conditions. Steady-state node voltage, branch power flow, and frequency data are extracted from the time-domain simulation waveforms generated by the simulation, and the node injected power is further calculated to obtain the operating sample set c.

[0024] The beneficial effects of this preferred technical solution are as follows: By setting the same constraints as simulations under normal weather conditions and conducting batch simulations to obtain the operational sample set b, the operating state of the power system under typical fault conditions can be simulated, providing data support for analyzing the impact of faults on the system. Furthermore, by combining simulations with icing disturbance conditions to obtain the operational sample set c, the impact of ice storms on power system operation is considered, enriching the sample library and making it more reflective of the power system's operation under actual complex environments. This helps improve the accuracy and reliability of subsequent scenario generation.

[0025] As a preferred solution for a source-grid-load operation scenario generation method based on transfer learning under ice disaster disturbance, wherein:

[0026] The pre-trained least squares loss function-based conditional generative adversarial network model, combined with the final standardized target domain sample set, undergoes transfer learning and target domain model fine-tuning to obtain a fine-tuned target domain ice disaster scene generator, including:

[0027] The parameters of the source domain generator are transferred, and some parameters of the source domain generator are selected as the initial parameters of the target domain generator; the model parameters of the target domain generator are fine-tuned using the icing sample set;

[0028] Freeze the parameters of the generator that were successfully trained in the pre-training stage, use the power grid operation scenario samples under icy weather as the fine-tuning sample set and perform data preprocessing; set the number of icy samples to be selected each time the parameters are fine-tuned; randomly select a certain number of samples and their corresponding labels from the fine-tuning sample set to update the parameters; repeat the steps of selecting samples and updating parameters until the model is successfully trained based on the fine-tuning sample set.

[0029] As a preferred solution for a source-grid-load operation scenario generation method based on transfer learning under ice disaster disturbance, wherein:

[0030] The process of classifying source domain samples into risk levels based on a standardized source domain sample set and its feature matrix to obtain a standardized source domain sample set with risk level labels includes:

[0031] The normal scenario sample set, the fault scenario sample set, and the icing operation scenario sample set are integrated into the source domain sample set and then standardized to obtain the standardized source domain sample set and its feature matrix.

[0032] The sample features are dimensionality reduced to determine the optimal number of clusters. The optimal number of clusters is selected by comprehensively considering the sum of squared errors within clusters and the silhouette coefficient. If the results obtained by different evaluation methods are the same, the result is directly used as the optimal number of clusters; if they are different, the number of clusters corresponding to the largest comprehensive factor is selected as the optimal number of clusters. After determining the optimal number of clusters, the risk level of the samples in the operating scenario is classified.

[0033] As a preferred solution for a source-grid-load operation scenario generation method based on transfer learning under ice disaster disturbance, wherein:

[0034] The standardized target domain sample set and its feature matrix, based on the icing operation samples with risk level labels in the power system operation scenario simulation sample library, generate the final standardized target domain sample set, which includes:

[0035] The scenario represented by all power grid operation samples is taken as the source domain, and the scenario represented by power grid operation samples under icing weather is taken as the target domain.

[0036] A conditional generative adversarial network model based on the least squares loss function is constructed. The target domain sample set is standardized using ice-covered running samples with risk level labels for iterative training, and the model parameters are dynamically updated using the least squares loss function.

[0037] The beneficial effects of this preferred technical solution are as follows: by taking the scenarios represented by all power grid operation samples as the source domain and the scenarios represented by power grid operation samples under icing weather as the target domain, the source and target domains of transfer learning are clearly defined. A conditional generative adversarial network model based on the least squares loss function is constructed, and iterative training is performed using a standardized target domain sample set of icing operation samples with risk level labels. The model parameters are dynamically updated using the least squares loss function, enabling the model to better learn the features and distribution of the target domain samples, improving the model's performance in the target domain, and thus generating a final standardized target domain sample set that better reflects the actual situation.

[0038] As a preferred solution for a source-grid-load operation scenario generation method based on transfer learning under ice disaster disturbance, wherein:

[0039] The process of generating the final standardized target domain sample set, based on the standardized target domain sample set and its feature matrix of icing operation samples with risk level labels from the power system operation scenario simulation sample library, also includes:

[0040] Icing simulation and sample set extraction were performed. The same constraints as those used in the simulation and sample set extraction of different output combinations of wind turbines under normal conditions were set. Based on the predetermined operating mode, single or combined conditions such as wind power output reduction, partial line interruption and load demand increase were applied one by one for simulation. Steady-state node voltage, branch power flow and frequency data were extracted from the relevant graph files generated by the simulation, and the node injection power was further calculated to obtain the operating sample set.

[0041] The beneficial effects of this preferred technical solution are as follows: Icing simulation and sample set extraction are performed using the same constraints as those used in simulations and sample set extraction of different output combinations of wind turbines under normal conditions. Single or combined conditions such as reduced wind power output, partial line interruptions, and increased load demand are applied one by one for simulation, enabling a comprehensive simulation of various operating conditions that the power system may face during ice storms. Steady-state node voltage, branch power flow, and frequency data are extracted from the relevant graph files generated by the simulation, and node injection power is further calculated to obtain an operational sample set. This provides rich and accurate icing operation samples for the target domain sample set, helping to improve the quality of the final standardized target domain sample set, thereby enhancing the accuracy of generating ice storm disturbance operation scenarios.

[0042] Secondly, the present invention provides a source-grid-load operation scenario generation system based on transfer learning under ice storm disturbance, comprising:

[0043] The module for building a sample library of operating scenarios is used to construct a simulation sample library of power system operating scenarios based on power grid structure data, typical fault type data, and icing disturbance conditions.

[0044] The source domain sample standardization module is used to integrate source domain samples and standardize them based on the power system operation scenario simulation sample library to obtain a standardized source domain sample set and its feature matrix.

[0045] The source domain sample risk classification module is used to classify the risk level of source domain samples based on the standardized source domain sample set and its feature matrix, and obtain a standardized source domain sample set with risk level labels.

[0046] The pre-trained model building module is used to construct a least-squares loss function conditional generative adversarial network model based on a standardized source domain sample set with risk level labels, and to perform pre-training.

[0047] The target domain sample generation module is used to standardize the target domain sample set and its feature matrix based on the icing operation samples with risk level labels in the power system operation scenario simulation sample library, and generate the final standardized target domain sample set.

[0048] The target domain model fine-tuning module is used to perform transfer learning and target domain model fine-tuning based on the pre-trained least squares loss function conditional generative adversarial network model, combined with the final standardized target domain sample set, to obtain the fine-tuned target domain ice disaster scene generator.

[0049] The ice disaster scenario generation module is used to generate ice disaster disturbance operation scenarios based on the fine-tuned target domain ice disaster scenario generator.

[0050] Thirdly, the present invention provides an electronic device, comprising:

[0051] Memory and processor;

[0052] The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the one or more programs are executed by the one or more processors, the one or more processors implement the method for generating source-grid-load operation scenarios under ice disaster disturbance based on transfer learning as described in this invention.

[0053] Fourthly, the present invention provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the aforementioned method for generating source-grid-load operation scenarios under ice storm disturbances based on transfer learning.

[0054] The beneficial effects of this invention are as follows: The method for generating source-grid-load operation scenarios under ice storm disturbances based on transfer learning provided by this invention has significant practical application value. By constructing a simulation sample library of power system operation scenarios, comprehensively considering grid structure, typical fault types, and icing disturbance conditions, it fully covers the operation of the power system under different operating conditions, providing a rich and accurate data foundation for subsequent analysis. Dividing the source domain samples into risk levels and constructing a sample set with risk level labels helps power departments identify the risk level of different operation scenarios in advance, formulate targeted response strategies, and improve the safety and reliability of the power system. Pre-training a least-squares loss function conditional generative adversarial network model, and combining it with the final standardized target domain sample set for transfer learning and target domain model fine-tuning, can generate operation scenarios that better reflect the actual situation of ice storms. This allows power departments to conduct more accurate simulations and predictions before ice storms arrive, enabling them to prepare equipment maintenance, resource allocation, and emergency plan formulation in advance, reducing the impact of ice storms on the power system and ensuring the stability of power supply. The generated ice storm disturbance operation scenarios can be used to assess the weaknesses of the power system under ice storms, guiding the planning and transformation of the power grid. Based on the scenario analysis results, the power grid structure can be optimized, the anti-icing capacity of transmission lines can be enhanced, and power sources and loads can be rationally distributed, thereby improving the overall disaster resilience of the power system. Furthermore, this method can also support risk management in the electricity market, helping power companies better cope with market fluctuations caused by ice storms. Attached Figure Description

[0055] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 This is an overall flowchart of the method for generating source-grid-load operation scenarios under ice disaster disturbance based on transfer learning provided by the present invention.

[0057] Figure 2 This is a framework diagram for constructing the training sample set of the method for generating source-grid-load operation scenarios under ice disaster disturbance based on transfer learning provided by the present invention.

[0058] Figure 3 This is a schematic diagram of the improved IEEE 39-node system in a simulation example of the method for generating source-grid-load operation scenarios under ice disaster disturbance based on transfer learning provided by this invention.

[0059] Figure 4 This is a schematic diagram of the principal component contribution rate and cumulative contribution rate in a simulation example of the source-grid-load operation scenario generation method based on transfer learning provided by this invention.

[0060] Figure 5 These are the cluster error maps and profile coefficient maps within different K values ​​in a simulation example of the source-grid-load operation scenario generation method based on transfer learning under ice disaster disturbance provided by this invention.

[0061] Figure 6 This is a clustering diagram of the operation scenario data in a simulation example of the method for generating source-grid-load operation scenarios under ice disaster disturbance based on transfer learning provided by this invention.

[0062] Figure 7 This is a schematic diagram showing the proportion of the ice-covered operation scenario sample set under four types of risks in a simulation example of the source-grid-load operation scenario generation method based on transfer learning under ice disaster disturbance provided by the present invention.

[0063] Figure 8 This is a schematic diagram of the node voltage in a simulation example of the method for generating source-grid-load operation scenarios under ice disaster disturbance based on transfer learning provided by the present invention.

[0064] Figure 9 This is a schematic diagram of the active power of the branch in a simulation example of the method for generating source-grid-load operation scenarios under ice disaster disturbance based on transfer learning provided by the present invention. Detailed Implementation

[0065] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0066] Example 1, referring to Figure 1 This is the first embodiment of the present invention, which provides a method for generating source-grid-load operation scenarios under ice storm disturbances based on transfer learning, including:

[0067] S1: Based on power grid structure data, typical fault type data, and icing disturbance conditions, construct a simulation sample library for power system operation scenarios;

[0068] S2: Based on the power system operation scenario simulation sample library, integrate source domain samples and standardize them to obtain a standardized source domain sample set and its feature matrix;

[0069] S3: Based on the standardized source domain sample set and its feature matrix, the risk level of the source domain samples is divided to obtain a standardized source domain sample set with risk level labels;

[0070] S4: Based on a standardized source domain sample set with risk level labels, construct a least-squares loss function conditional generative adversarial network model and pre-train it;

[0071] S5: Based on the standardized target domain sample set and its feature matrix of the icing operation samples with risk level labels in the power system operation scenario simulation sample library, generate the final standardized target domain sample set;

[0072] S6: Based on the pre-trained least squares loss function conditional generative adversarial network model, combined with the final standardized target domain sample set, transfer learning and target domain model fine-tuning are performed to obtain the fine-tuned target domain ice disaster scene generator.

[0073] S7: Generate ice disaster disturbance operation scenarios based on the fine-tuned target domain ice disaster scenario generator.

[0074] It should be noted that, through steps S1-S7, LS-CGAN is first pre-trained using samples of all operating modes constructed based on DSP software to form a basic generative model. Then, parameters are transferred using a small number of power grid icing operation samples. The target generative model, obtained after fine-tuning the parameters, can generate power grid icing operation mode samples. Based on the LS-CGAN model and transfer learning method, the features of the source domain dataset can be learned. Only a small amount of power grid icing operation mode data and minor fine-tuning are needed to accurately generate operating mode samples that meet the requirements, providing data support for the formulation of power system operating modes.

[0075] Example 2, refer to Figures 1-2 As an embodiment of the present invention, based on the previous embodiment, a method for generating source-grid-load operation scenarios under ice disaster disturbance based on transfer learning is provided, including:

[0076] In this embodiment, the step S1 above, which involves constructing a power system operation scenario simulation sample library based on power grid structure data, typical fault type data, and icing disturbance conditions, includes:

[0077] The power system operation scenario simulation sample library consists of extreme operation scenario sample sets, represented by a three-dimensional matrix (NS, NI, 8). NS represents the number of samples, NI represents the total number of power grid nodes, and each sample contains eight features: injected active power, injected reactive power, node voltage amplitude, phase angle, maximum frequency deviation, steady-state frequency deviation, frequency settling time, and transient overvoltage parameters for all nodes in the power grid.

[0078] Specifically, based on power grid structure data (DAT / SWI files), typical fault type data (LSD files), and icing disturbance conditions (reduced wind power output / line interruption / load increase), the electromechanical transient calculation module (DSP-TRANSIENT VERSION:2.3.39.0 BUILD:2024.11.13) in the power system calculation and analysis software DSP (DynamicSimulation Package) was used to perform sample simulations, including determining the operating mode, simulating power grid operation under normal weather conditions and extracting sample set a, simulating power grid operation under typical fault conditions and extracting sample set b, and simulating power grid operation under icing weather conditions and extracting sample set c.

[0079] Furthermore, the operational mode is determined to include:

[0080] An electromechanical transient simulation project was created in the DSP, which included power flow calculation DAT files and stability calculation SWI files. The DAT files were used to construct power flow text data for different high penetration scenarios, and the SWI files were used to store system transient parameters and simulation output settings for the operating modes. This paper sets the output of node injected active power, injected reactive power, node voltage amplitude, phase angle, maximum frequency deviation, steady-state frequency deviation, frequency settling time and transient overvoltage information for all lines.

[0081] Simulation of different output combinations of wind turbine units under normal conditions and extraction of sample set a include:

[0082] For system safety reasons, it is necessary to first determine the overload ratio limit of branch power flow and the constraint limit of node voltage. Firstly, based on a defined operating mode, simulations are conducted by changing the output combination of wind turbines under different renewable energy penetration rates. The simulation-generated CHT chart file contains time-domain simulation waveforms of node voltage amplitude, phase angle, branch power flow, and frequency. Steady-state node voltage, branch power flow, and frequency data are extracted from the chart, and the node injection power is further calculated, thus obtaining the operating sample set a.

[0083] Typical fault simulation and sample set b extraction include:

[0084] The same constraints as those used in the simulation of different output combinations of wind turbines under normal conditions and in the extraction of sample set a were set. First, based on the determined operating mode, typical fault types were identified and written into a fault text data LSD file, which was then loaded into the same electromechanical transient simulation project. Second, based on the simulation platform's memory space, the maximum number of parallel simulation threads was set appropriately, and batch simulations were performed. The CHT diagram file generated by the simulation is a time-domain simulation waveform of node voltage amplitude, phase angle, branch power flow, and frequency. Steady-state node voltage, branch power flow, and frequency data were extracted from the diagram, and the node injected power was further calculated, thus obtaining the operating sample set b.

[0085] In another possible implementation, the sample set a can also be obtained through a Monte Carlo simulation-based method;

[0086] Specifically, for the output of wind turbines, a large number of different combinations of output values ​​are generated using a random number generator, based on their possible distribution range (for example, assuming that the output of wind turbines follows a certain probability distribution, such as normal distribution, Weibull distribution, etc.).

[0087] At the same time, other random factors that may affect the operation of the power grid, such as random load fluctuations, should be considered. The range and probability distribution of load fluctuations can be statistically determined based on historical load data, and different load values ​​can be randomly generated.

[0088] Based on a defined operating mode, the randomly generated wind turbine output combination and load value are input into the electromechanical transient calculation module of the power system calculation and analysis software DSP for simulation.

[0089] Each simulation records the generated CHT diagram file, which contains the time-domain simulation waveforms of node voltage magnitude, phase angle, branch power flow, and frequency.

[0090] Extract steady-state node voltage, branch power flow, and frequency data from each CHT plot file.

[0091] The node injection power is further calculated based on the extracted data. Each simulation yields a sample, which is then aggregated to form the running sample set a.

[0092] In another possible implementation, the sample set a can also be obtained through a scene generation algorithm;

[0093] Specifically, different scenarios are defined, such as different weather conditions (although icing is the primary consideration, it can be extended to the impact of other weather conditions on wind turbine output) and different changes in power grid topology (such as temporary maintenance of certain lines).

[0094] Scene generation algorithms, such as Latin hypercube sampling (LHS), are used to generate a series of representative scenes. The LHS algorithm can uniformly sample points in a multidimensional parameter space, thereby obtaining scenes with different parameter combinations.

[0095] For each generated scenario, the corresponding parameters (such as wind turbine output and load in that scenario) are input into the electromechanical transient calculation module of the DSP for simulation.

[0096] Record the CHT diagram file generated in each simulation, extract steady-state node voltage, branch power flow and frequency data from it, calculate node injected power, and form the operating sample set a.

[0097] In this embodiment, step S2 above, based on the power system operation scenario simulation sample library, integrates source domain samples and standardizes them to obtain a standardized source domain sample set and its feature matrix, including:

[0098] The source domain sample set is constructed by integrating sample set a (normal scenario) and sample set b (failure scenario). The Z-score standardization method is used to eliminate dimensional differences in the source domain sample set. The calculation formula is as follows:

[0099]

[0100] Where, x ij z is the standardized value of the j-th indicator for the i-th sample; ij This is the original data; μ j and σ j denoted as the mean and standard deviation of the j-th indicator, respectively.

[0101] The standardized feature parameter matrix X is constructed as follows:

[0102]

[0103] Where a is the number of samples in the running scenario, and b is the number of feature factors.

[0104] In this embodiment, step S3 above, based on the standardized source domain sample set and its feature matrix, divides the source domain samples into risk levels to obtain a standardized source domain sample set with risk level labels, including:

[0105] To differentiate the safety risks among samples from different operating scenarios, it is necessary to further classify the risk levels of the samples. Before classifying, it is necessary to reduce the dimensionality of the sample features and determine the optimal number of clusters.

[0106] Specifically, principal component analysis is used for dimensionality reduction: the covariance matrix S of X is calculated as follows: ij ) b×b , represented as:

[0107]

[0108] in, The average value of the i-th eigenvector of matrix X; Let be the average value of the j-th eigenvector of matrix X.

[0109] Calculate the eigenvalues ​​λ of the covariance matrix S i The formula is as follows:

[0110] det(X-λE)=0

[0111] Calculate the eigenvalues ​​λ of the covariance matrix S i With the corresponding eigenvector a i And sort them in descending order of eigenvalues ​​to obtain b principal component components F. i :

[0112]

[0113]

[0114] Calculate the contribution rate ψ of the eigenvalues i Given the cumulative contribution rate ξ, we select principal components using the following formula:

[0115]

[0116] Where m represents the number of principal components selected.

[0117] In another possible implementation, the optimal number of clusters can be determined using the elbow method. The core principle of this method is to evaluate the clustering effect using the sum of squared errors (SSE) of the studied sample data. As the number of clusters (K) increases, the target data becomes clearer, the clustering effect improves, and the SSE gradually decreases. The rate of decrease in SSE exhibits a trend of first sharply decreasing and then stabilizing depending on the number of clusters. Specifically, when the number of clusters (K) is less than the actual number of clusters in the target data, the clustering density of each cluster increases sharply, causing the SSE to decrease rapidly. Conversely, when the number of clusters (K) equals the actual number of clusters, the clustering density of each cluster approaches its limit, and the SSE gradually stabilizes, eventually forming an elbow shape. The formula for the sum of squared errors (SSE) is as follows:

[0118]

[0119] Among them, C i Let p be the i-th cluster; p be the sample data points in the cluster; m be the number of clusters. irepresents the sample data points in the cluster; SSE is the sum of squared errors for all data.

[0120] In another possible implementation, the optimal number of clusters can be determined using the silhouette coefficient method. The silhouette coefficient is an indicator for evaluating clustering effectiveness, combining the clustering cohesion and separation to measure the quality of the clustering results. The silhouette coefficient ranges from -1 to 1, with a larger value indicating better clustering performance.

[0121] For each sample i, the silhouette coefficient s(i) is defined as:

[0122]

[0123] Where a(i) is the average distance (cohesion) from sample i to other samples in its cluster; b(i) is the minimum average distance (separation) from sample i to samples in other clusters, calculated as follows:

[0124]

[0125] For the entire dataset, the total silhouette coefficient S is the average silhouette coefficient of all samples:

[0126]

[0127] Where n is the total number of samples.

[0128] In this embodiment, the optimal number of clusters is selected based on a combination of the sum of the squared errors (SSE) and the silhouette coefficient (SC).

[0129] Specifically, if the optimal number of clusters determined by SC and SSE is the same, then that result is directly taken as the optimal number of clusters; if the results are different, then the number of clusters corresponding to the largest comprehensive factor is taken as the optimal number of clusters, as defined below:

[0130]

[0131] In the formula: The comprehensive factor when the optimal number of clusters is α; and ω1 and ω2 are the standardized intra-cluster SSE and SC when the optimal number of clusters is α, respectively; ω1 and ω2 are the corresponding weights.

[0132] After performing feature dimensionality reduction and determining the optimal number of clusters for the samples, this embodiment preferably uses the tk-means fast method to classify the risk levels of the samples in the operating scenario.

[0133] Specifically, tk-means introduces a t-mixture model (TMM), whose probability density function is:

[0134]

[0135] Among them, v k For the degree of freedom function, controlling the thickness of the distribution tail, when v k As v approaches infinity, the t-distribution degenerates into a Gaussian distribution; when v k When finite, its heavy-tailed property can better model outliers.

[0136] Based on indicator variable z and weight variable μ n Log-likelihood of complete data lnL c (ψ|x,μ,z) can be decomposed into two parts:

[0137] lnL c (ψ|x,μ,z)=lnL G (v|μ,z)+lnL N (μ,α|x,μ,z)

[0138] Among them, the gamma distribution likelihood lnL G Corresponding to μ n The gamma distribution likelihood is expressed as:

[0139]

[0140] Gaussian conditional distribution likelihood lnL N , represented as:

[0141]

[0142] The parameters are optimized alternately through the Expectation Maximization (EM) algorithm, which is specifically divided into the E process (expectation calculation) and the M process (parameter update).

[0143] The E-process (expectation calculation) includes:

[0144] Calculate sample x n The posterior probability of belonging to the k-th cluster, i.e., the posterior membership degree τ. nk :

[0145]

[0146] Where t k (·) is the t-distribution density function.

[0147] Calculate the latent variable μ n Expected μ nk :

[0148]

[0149] The weight decreases as the distance between the sample and the cluster center increases, in order to suppress the influence of outliers.

[0150] The M-process (parameter update) includes:

[0151] New cluster centers of tk-means Represented as:

[0152]

[0153] It should be noted that this update method involves all samples in the process of updating cluster centers, and introduces μ. nk Weighting reduces the contribution of samples far from the center.

[0154] covariance parameter α * The dispersion of the weighted sample is reflected as follows:

[0155]

[0156] degrees of freedom v * Iterative updates using approximate formulas:

[0157]

[0158] Furthermore, to improve computational efficiency, the computational complexity is simplified:

[0159] First, fix the degree of freedom parameter and set it to the value of the corresponding Cauchy distribution to avoid complex iterations.

[0160] Secondly, the covariance matrix is ​​degenerated into a scalar, at which point the cluster center update simplifies to:

[0161]

[0162] It should be noted that, despite the simplification, it is still achieved through τ nk Soft allocation is implemented, preserving the robustness of the algorithm.

[0163] It should also be noted that the loss function of the tk-means method is the log-squared loss function, which is expressed as:

[0164]

[0165] Compared to the squared loss function of the traditional k-means algorithm, the log-squared loss function is less sensitive to outliers and exhibits stronger robustness when dealing with noisy or outlier data.

[0166] The cluster center update of tk-means not only depends on the sample information within the current cluster, but also takes into account the global information of all samples. The introduction of this global information significantly reduces the algorithm's dependence on the selection of initial cluster centers, thereby improving the stability of the clustering results.

[0167] In another possible implementation, agglomerative hierarchical clustering can also be used when classifying the risk levels of source domain samples.

[0168] Specifically, each sample in the standardized source domain sample set is treated as a separate cluster, at which point the number of clusters equals the number of samples. A cluster distance metric is chosen, such as single link (distance between the closest samples in two clusters), full link (distance between the farthest samples in two clusters), or average link (average distance between all sample pairs in two clusters). The distances between all clusters are calculated; the two closest clusters are identified and merged into a new cluster. The number of clusters is then reduced by 1. The steps of calculating inter-cluster distances and merging clusters are repeated until a preset number of clusters is reached or other stopping conditions are met. Based on the final clustering results, a risk level label is assigned to each cluster.

[0169] In another possible implementation, split hierarchical clustering can also be used when classifying the risk level of source domain samples;

[0170] Specifically, the standardized source domain sample set is treated as a whole as a cluster; a suitable cluster is selected for splitting, which can be chosen based on factors such as the size of the cluster and the dispersion of the samples; a splitting method, such as the k-means algorithm, is used to split the selected cluster into two or more sub-clusters; the steps of selecting and splitting clusters are repeated until the preset number of clusters is reached or other stopping conditions are met; based on the final clustering results, a risk level label is assigned to each cluster.

[0171] In this embodiment, step S4 above, which involves constructing a least-squares loss function conditional generative adversarial network model based on a standardized source domain sample set with risk level labels and performing pre-training, includes:

[0172] It should be noted that transfer learning uses existing knowledge to assist in learning new tasks. Essentially, it analyzes the characteristics of source domain data, extracts knowledge, and applies it to a target domain with similar data characteristics, addressing the problem of scarce or disparate data in the target scenario. Compared to other machine learning methods, transfer learning does not require consistent distribution of training and test data, has a lower dependence on large amounts of labeled data, and can achieve model transfer applications across different tasks, demonstrating strong adaptability.

[0173] In transfer learning, the source task and the target task are represented by subscripts S and T, respectively, and D s and D TThe source domain and the target domain are respectively defined. By combining the feature vector space X with the probability distribution function P(X), their respective neighborhoods are constructed, i.e.:

[0174]

[0175] Let Y s Let Y be the label vector space of the source domain. t Let f be the label vector space of the target domain. s f is the mapping function of the source domain. t Let T be the mapping function of the target domain, then the tasks T in the source domain and the target domain are... s and T t They can be described as follows:

[0176]

[0177] Assuming the source and target domains have the same feature distributions but different label spaces, the goal of transfer learning is to learn the mapping function f. t :X t →X s In D T The expected error is minimized on X, and X is satisfied. t =X s Y t =Y s and P(Y) t |X t )≠P(Y s |X s ).

[0178] The Least Squares Conditional Generative Adversarial Network (LS-CGAN) model, building upon traditional GANs, effectively addresses common GAN training issues like vanishing gradients and mode collapse by introducing a least squares objective function. Its loss function employs a smoother, non-saturating gradient least squares loss function. Compared to the cross-entropy loss function used in original GANs, the least squares loss function provides additional gradients by penalizing samples far from the decision boundary. This optimizes the matching between generated samples and the real data distribution, improves the distributional discrepancy between generated and real samples, and ultimately enhances the quality of generated samples, increasing the stability of the training process.

[0179] The framework of LS-CGAN consists of a generator (G) and a discriminator (D). Let the input of the generator (G) be random noise z ~ p. zThe generator takes the real data x or the generated sample G(z|c) as input and the conditional information c as input, and outputs the generated sample G(z|c). The generator aims to generate samples that are as realistic as possible by learning and transforming random noise, in order to fool the discriminator. The discriminator (D) takes the real data x or the generated sample G(z|c) as input and the conditional information c as input, and outputs the probability value D(x|c) or D(G(z|c)|c). The discriminator aims to accurately distinguish between real data and generated samples. The conditional information c is usually added to the input of the generator and discriminator through an embedding layer or direct concatenation. Mathematically, the loss function of LS-CGAN is based on the least squares loss (MSE), which aims to minimize the mean squared error between real samples and generated samples. The loss functions of the discriminator and generator are as follows:

[0180]

[0181] Let *a* represent the generated pseudo-data, *b* represent the real data, and *d* represent the probability that G wants D to believe the pseudo-data; this value is, to some extent, equivalent to the decision boundary. Extensive experimental verification shows that when *b* = *d*, setting *b* = *d* = 1 and *a* = 0 allows the generator and discriminator to achieve optimal performance.

[0182]

[0183] Where x is the real sample, z is random noise, and c is conditional information; D(x|c) is the discriminator's output on the real sample x and condition c; G(z|c) is the sample generated by the generator based on the noise z and condition c; D(G(z|c)|c) is the discriminator's output on the generated sample G(z|c) and condition c; E represents the mathematical expectation, p data (x) is the distribution of the real data, p z (z) represents the distribution of random noise. This loss function reflects the discriminator's goal of outputting a probability value as close to 1 as possible for real data and as close to 0 as possible for generated samples, achieved by minimizing the mean squared error. The generator G aims to make the discriminator's output for generated samples approach the target value, making it difficult for the discriminator to distinguish between generated and real samples.

[0184] The discriminator and generator are trained using a minimax game approach. In the early stages of training, the samples generated by the generator may differ significantly from the real data, which the discriminator can easily identify. As training progresses, the generator continuously optimizes its parameters, producing increasingly realistic samples, while the discriminator adjusts its parameters accordingly to improve its discrimination ability. In this game-like learning process, both improve together until a Nash equilibrium is reached. At Nash equilibrium, the generator's output is completely identical to the original data distribution; at this point, the discriminator can no longer accurately distinguish between real and generated samples, achieving an ideal balance.

[0185] During the pre-training phase, the generator attempts to generate simulated power system operation scenario samples based on the rich data features in the source domain, while the discriminator strives to distinguish these generated samples from the real samples in the source domain. Through continuous iterative training, the model gradually learns the general patterns and feature representations in the source domain data, thereby constructing a basic model for scenario generation that can accurately reflect the operating state of the power system under various weather conditions.

[0186] The specific steps are as follows:

[0187] Step 4.1: Select the total training sample set s s The data is preprocessed.

[0188] Step 4.2: Combine the noise vector and label vector as input to the generator in the LS-CGAN model, and use the total training sample set s s Combined with label vector Y s As input to the discriminator.

[0189] Step 4.3: Set the number of samples n selected for each parameter pre-training session. s Randomly initialize the generator and discriminator parameters θ.

[0190] Step 4.4: Start training, starting from the total training sample set s s Randomly select n s Each sample and its corresponding label Update parameter θ.

[0191] Step 4.5: Repeat step 4.4 until based on the total training sample set s s The model was successfully trained.

[0192] After building the basic model, it is also necessary to construct a training sample set as its input. The framework for constructing the training sample set is as follows: Figure 2 As shown.

[0193] In this embodiment, the step S5 above, which generates the final standardized target domain sample set based on the standardized target domain sample set and its feature matrix of icing operation samples with risk level labels in the power system operation scenario simulation sample library, includes:

[0194] It should be noted that for the generation of ice storm disturbance scenarios, the scenario represented by all power grid operation samples is used as the source domain, including the power transmission status of each transmission line, voltage fluctuations and frequency status of power grid nodes, etc., to comprehensively reflect the operating status of the power system. The scenario represented by the power grid operation samples under icing weather is used as the target domain, and the sample data in the target domain reflects the unique operating characteristics of the power system under the influence of icing disasters.

[0195] The framework for generating ice storm disturbance scenarios consists of two stages: basic model construction and parameter migration.

[0196] In the basic model building phase, the core task is to construct a basic model with generalization capabilities through a "pre-training" process. First, a Conditional Generative Adversarial Network (LS-CGAN) based on the least squares loss function is built. A large-scale training sample set is fed into the LS-CGAN model for iterative training. The model parameters are dynamically updated using the least squares loss function, which guides the generator and discriminator parameters towards the optimal solution by minimizing the squared error between the predicted and true values. After multiple rounds of iterative training, the trained LS-CGAN model can generalize to a certain extent for generating all running scenarios, laying the foundation for subsequent generation of ice storm disturbance scenarios.

[0197] Perform icing simulation and extract sample set c, including:

[0198] The same constraints as those used in the simulation of different output combinations of wind turbines under normal conditions and the extraction of sample set a are set. First, based on the determined operating mode, single or combined conditions such as wind power output reduction, partial line interruption, and increased load demand are applied one by one for simulation. The CHT chart file generated by the simulation is a time-domain simulation waveform of node voltage amplitude, phase angle, branch power flow, and frequency. Steady-state node voltage, branch power flow, and frequency data are extracted from the chart, and the node injected power is further calculated to obtain the operating sample set c.

[0199] In this embodiment, the pre-trained least squares loss function conditional generative adversarial network model in step S6, combined with the final standardized target domain sample set, is used for transfer learning and fine-tuning of the target domain model to obtain the fine-tuned target domain ice disaster scene generator, including:

[0200] In the parameter transfer phase, the "fine-tuning" process aims to adapt the pre-trained model to the specific task of generating ice disaster disturbance scenarios. First, parameter transfer is performed on the generator model in the source domain, selecting some parameters from the source domain generator as the initial parameters for the target domain generator. These initial parameters contain general features and patterns inherent in the source domain data, providing a good starting point for training the target domain model. Subsequently, the model parameters of the target domain generator are fine-tuned using a small set of icing samples. In this way, the target domain generator can quickly learn the unique features of ice disaster scenarios while inheriting knowledge from the source domain, thereby achieving the task of generating ice disaster disturbance scenarios.

[0201] It should be noted that choosing fine-tuning as the core method of transfer learning has significant advantages: it eliminates the need for large-scale model retraining, greatly saving time and costs; the pre-trained model in the source domain is usually trained on a large-scale dataset, and the features it learns can expand the feature dimension of the target domain dataset, effectively improving the robustness of the model in the target task; the operation is relatively simple, only requiring a focus on the data features and requirements of the target task, and making appropriate adjustments to the model, reducing the difficulty and complexity of model optimization.

[0202] Considering that the target domain is the scenario represented by the power grid operation samples under icing weather, which has a small amount of sample data, and the source domain is the scenario represented by all power grid operation samples with abundant data, the frozen training mode is selected for transfer learning.

[0203] Specifically, in the fine-tuning phase, the trained base model is finely tuned using a limited sample of power grid operation scenarios under icing weather conditions. During the fine-tuning process, the model focuses on the differences between the source and target domains. By making targeted and subtle adjustments to the model parameters, the model can accurately capture the unique operating modes of the power system under icing disasters, optimize the matching degree between the generated samples and the distribution of real data under icing weather conditions, significantly improve the distribution differences between the generated samples and real icing samples, and thus generate a realistic ice disaster disturbance scenario.

[0204] The specific steps are as follows:

[0205] Step 6.1: Freeze the parameters of the generator that was successfully trained in the pre-training phase.

[0206] Step 6.2: Use the power grid operation scenario samples under icing weather as the fine-tuning sample set s T The data is preprocessed.

[0207] Step 6.3: Set the number n of icing samples selected for each parameter fine-tuning. T .

[0208] Step 6.4: Begin training from the fine-tuning sample set ST Randomly select n T Each sample and its corresponding label Update parameter θ.

[0209] Step 6.5: Repeat step 6.4 until the result is based on the fine-tuned sample set S. T The model was successfully trained.

[0210] In another possible implementation, successful model training can be determined in the following way:

[0211] During the training of a least-squares loss function conditional generative adversarial network (LS-CGAN), the loss function values ​​of the generator and discriminator change continuously with each training epoch. When the loss function values ​​of the generator and discriminator no longer fluctuate significantly within a certain number of epochs, but instead stabilize and reach a relatively small value, the model training can be considered close to successful.

[0212] In another possible implementation, successful model training can be determined in the following way:

[0213] Based on experience or previous experiments, set a reasonable threshold for the loss function. When the loss function values ​​of both the generator and discriminator decrease below this threshold, the model can be considered successfully trained. For example, in multiple experiments, it was found that when the generator's loss value is below 0.1 and the discriminator's loss value is between 0.2 and 0.3, the generated samples are of higher quality. In this case, these two values ​​can be used as the threshold for judging successful training.

[0214] In another possible implementation, transfer learning and fine-tuning of the target domain model can be performed in multiple stages: the fine-tuning process is divided into several stages, each using a different learning rate and training strategy. In the early stages, a larger learning rate is used to allow the model to quickly adapt to the general features of the target domain; in the later stages, a smaller learning rate is used for fine-tuning to avoid overfitting.

[0215] Specifically, the first stage involves fine-tuning some parameters of the target domain generator using a relatively large learning rate (e.g., 0.01). This stage allows for rapid adjustment of the model's parameters, enabling it to initially adapt to the characteristics of power grid operation under icing conditions. The intermediate stage involves gradually reducing the learning rate (e.g., to 0.001) to continue fine-tuning the model. At this point, the model begins to learn the features of the target domain more deeply, while avoiding instability caused by excessive parameter updates. The final stage involves using an even smaller learning rate (e.g., 0.0001) for fine-tuning the model. This stage primarily involves subtle adjustments to the model's parameters to optimize the distribution differences between generated samples and real icing samples. After each stage, the change in the loss function value is used to determine whether to adjust the learning rate or proceed to the next stage, until the model is successfully trained.

[0216] In another possible implementation, when performing transfer learning and fine-tuning the target domain model, feature matching-based fine-tuning can also be performed. During fine-tuning, not only is the loss function of the model-generated samples considered, but a feature matching mechanism is also introduced to make the generator-generated samples closer to real icy samples at the feature level. The distance between the generated samples and the real samples in certain feature spaces can be calculated and added as an additional loss term to the original loss function.

[0217] Specifically, a pre-trained feature extractor (e.g., an intermediate layer of a convolutional neural network trained on the source domain) is used to extract features from both the power grid operation scenario samples and the generated samples under icy weather conditions. The distance between the generated and real samples in the feature space (e.g., Euclidean distance, cosine similarity, etc.) is calculated and used as the feature matching loss. The feature matching loss is combined with the original least squares loss function to obtain the joint loss function. During fine-tuning, the joint loss function is optimized simultaneously. Following the aforementioned fine-tuning steps (e.g., randomly selecting samples, updating parameters, etc.), the target domain generator is trained using the joint loss function until the model is successfully trained.

[0218] In this embodiment, the step S7 above, which generates an ice disaster disturbance operation scenario based on the fine-tuned target domain ice disaster scenario generator, includes:

[0219] A least-squares loss function conditional generative adversarial network model was chosen as the base model. s and X T The eight characteristic quantities are: active power injected into the nodes, reactive power injected into the nodes, node voltage magnitude, phase angle, maximum frequency deviation, steady-state frequency deviation, frequency settling time, and transient overvoltage, respectively, in the source and target domains; P(X s ) and P(X T The numbers T represent the probability distributions of the eight features in the source and target domains, respectively. sThe goal is to generate samples in the source domain that are infinitely similar to all the original running samples; T t The objective is to generate an ice storm disturbance scenario within the target domain; Y s and Y t These are the sample risk labels in the source and target domains, respectively; f s For the source domain basic model LS-CGAN, f t This is the LS-CGAN model after transfer to the target domain.

[0220] Input random noise and target domain risk labels into the fine-tuned target domain ice disaster scene generator, and output ice disaster disturbance scene samples.

[0221] Example 3: The above is an illustrative scheme of the source-grid-load operation scenario generation method based on transfer learning under ice disaster disturbance in this embodiment. It should be noted that the technical solution of the source-grid-load operation scenario generation system based on transfer learning under ice disaster disturbance is based on the same concept as the technical solution of the above-described source-grid-load operation scenario generation method based on transfer learning. Details not described in detail in the technical solution of the source-grid-load operation scenario generation system based on transfer learning under ice disaster disturbance in this embodiment can be found in the description of the technical solution of the above-described source-grid-load operation scenario generation method based on transfer learning under ice disaster disturbance.

[0222] This embodiment also provides a source-grid-load operation scenario generation system based on transfer learning under ice disaster disturbance, including:

[0223] The module for building a sample library of operating scenarios is used to construct a simulation sample library of power system operating scenarios based on power grid structure data, typical fault type data, and icing disturbance conditions.

[0224] The source domain sample standardization module is used to integrate source domain samples and standardize them based on the power system operation scenario simulation sample library to obtain a standardized source domain sample set and its feature matrix.

[0225] The source domain sample risk classification module is used to classify the risk level of source domain samples based on the standardized source domain sample set and its feature matrix, and obtain a standardized source domain sample set with risk level labels.

[0226] The pre-trained model building module is used to construct a least-squares loss function conditional generative adversarial network model based on a standardized source domain sample set with risk level labels, and to perform pre-training.

[0227] The target domain sample generation module is used to standardize the target domain sample set and its feature matrix based on the icing operation samples with risk level labels in the power system operation scenario simulation sample library, and generate the final standardized target domain sample set.

[0228] The target domain model fine-tuning module is used to perform transfer learning and target domain model fine-tuning based on the pre-trained least squares loss function conditional generative adversarial network model, combined with the final standardized target domain sample set, to obtain the fine-tuned target domain ice disaster scene generator.

[0229] The ice disaster scenario generation module is used to generate ice disaster disturbance operation scenarios based on the fine-tuned target domain ice disaster scenario generator.

[0230] This embodiment also provides an electronic device applicable to the method for generating source-grid-load operation scenarios under ice disaster disturbances based on transfer learning, including:

[0231] The system includes a memory and a processor. The memory stores computer-executable instructions, and the processor executes these instructions to implement the source-grid-load operation scenario generation method based on transfer learning proposed in the above embodiments.

[0232] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the source-grid-load operation scenario generation method based on transfer learning proposed in the above embodiments.

[0233] The storage medium proposed in this embodiment and the method for generating source-grid-load operation scenarios under ice disaster disturbance based on transfer learning proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0234] Example 4, refer to Figures 3-9 Tables 1-5 illustrate one embodiment of the present invention, providing a method for generating source-grid-load operation scenarios under ice disaster disturbances based on transfer learning. To verify the beneficial effects of the present invention, scientific demonstration is conducted through simulation experiments.

[0235] To create a high-proportion renewable energy integration scenario, in the IEEE 39-node system, the traditional turbines at nodes 30, 32, and 36 are replaced with wind turbines. The system structure is as follows: Figure 3 As shown.

[0236] To find the optimal dimension to describe the features of the sample set and improve clustering performance and efficiency, PCA dimensionality reduction was first performed on the sample set of the running scenario containing 8 features. After principal component analysis, the contribution rate of each principal component and the cumulative contribution rate are as follows: Figure 4 As shown, with a cumulative contribution rate threshold of 95%, three principal components are retained.

[0237] Using the method of this invention, cluster error maps (SSE maps) and silhouette coefficient maps (SC maps) with different K values ​​are obtained, such as... Figure 5As shown, SSE is the sum of squared Euclidean distances from the sample points within a cluster to the center point corresponding to the current K value. It reflects the sample density within the cluster; the smaller the value, the denser the samples. SSE decreases as the K value increases because the sample division is finer and the clusters are more compact. It continues to decrease until the K value equals the number of samples, reaching its minimum. However, a smaller SSE does not necessarily mean a more reasonable K value. The silhouette coefficient SC takes values ​​[-1, 1], with values ​​closer to 1 indicating better classification, as shown in Table 1.

[0238] Table 1. In-cluster errors and silhouette coefficients for clustering with different K values.

[0239]

[0240] From Table 1 and Figure 5 It can be seen that the SSE decreases significantly when K<4 in the intra-cluster error graph, and the decrease shrinks significantly after K>4, indicating that K has exceeded the reasonable value. The benefit of continuing to increase K will decrease sharply. Therefore, the SSE graph shows that the optimal K value is 4. In the silhouette coefficient graph, the SC value is closest to 1 when K=3. Therefore, the optimal K value in the SC graph is 3.

[0241] If the optimal number of clusters K determined by SSE and SC is the same, it is directly taken as the optimal result; if the two results are different, the number of clusters with the largest comprehensive factor is taken as the optimal result. Given the difference between the optimal K values ​​given by SSE and SC in the above results, it is necessary to standardize the SSE and SC values.

[0242] The calculation shows that the comprehensive factor is 11.3892 when K=4 and 2.0146 when K=3. Obviously, the optimal number of clusters is K=4.

[0243] Therefore, the number of clusters for the sample set was set to 4. After clustering, the data was visualized using the three principal components as three-dimensional coordinate axes. The results are shown below. Figure 6 .

[0244] In this embodiment, the LS-CGAN model is built using Python, and trained on a GPU via the PyTorch framework using CUDA parallel computing. The computing platform configuration is: Intel(R) Xeon(R) GOLD5218R 2.10GHz CPU, 128GB RAM, and NVIDIA 308010G GPU. The network structure parameters for building the LS-CGAN model are shown in Tables 2 and 3.

[0245] Table 2 LS-CGAN Network Generator Structure

[0246]

[0247]

[0248] Table 3. Structure of the LS-CGAN Network Discriminator

[0249]

[0250] After pre-training, the first three layers of the generator are frozen to retain the ability to generate general features at the lower levels. Their parameters are then transferred and used as the initial parameters for the generator in the fine-tuning stage. Next, a small number of icy samples are used to fine-tune the model, employing a least-squares loss function, combined with an Adam optimizer and an exponentially decaying learning rate to improve convergence and performance. The data loader batch size is set to 32, the two decay factors of the Adam optimizer are 0.5 and 0.999, the learning rates of the generator and discriminator are 0.0001 and 0.0002 respectively, and the weight decay coefficient is 0.0002.

[0251] After obtaining a sample set of icing operation scenarios through simulation, the distance between each sample and each cluster center is calculated. Samples are then assigned to their corresponding risk levels based on the closest distance, and the proportion of each risk level is statistically analyzed. The results are as follows: Figure 7 As shown.

[0252] from Figure 7 As can be seen, the highest proportion of the sample was for icing operation scenarios with level 3 risk, accounting for 48.71%; followed by level 4 risk, accounting for 28.86%; and level 2 and level 1 risks accounted for 15.14% and 7.29%, respectively.

[0253] Analysis of the risk classification proportions of the icing sample set revealed that the highest proportion of icing operation scenarios were at risk level 3, and high-risk icing scenarios received more attention from operators. Therefore, it is necessary to generate such high-risk samples specifically.

[0254] The LS-CGAN model, trained via parameter transfer, uses Level III risk labels and 500 sets of noise as input. The generated scenarios show the node voltages and branch active power flow as follows: Figure 8 and Figure 9 As shown.

[0255] Figure 8 The distribution characteristics of node voltage exceedances across four risk levels are presented. Different colored areas in the figure correspond to different risk levels; the darker the color, the higher the risk, the more severe the icing disaster, and the higher the system risk level. In addition, some nodes, due to their distance from the power source, have larger voltage deviations and are more prone to voltage exceedances in icing disaster scenarios; while nodes closer to the power source have smaller voltage deviations and relatively better voltage stability.

[0256] Figure 9The distribution of active power risk in the generated scenario is displayed, with different risk levels distinguished by color. The fluctuation characteristics of the active power risk distribution in each branch are obvious: the higher the risk level, the more significant the fluctuation. Some lines have high active power, indicating that their active power exceeds the limit severely, while some lines have zero active power, indicating that the icing disaster has a significant impact on the power grid and that there are line outages. As the risk level decreases, the fluctuation decreases, the active power exceeding the limit and the outage situation improves, indicating that the impact of the icing disaster on the power grid is weakened.

[0257] To verify the advantages of transfer learning in generating small-sample icing disaster scenarios, risk labels of levels I, II, and IV were horizontally spliced ​​with 500 groups of noise as input. Icing scenarios were generated by considering and not considering transfer learning. The distance from each sample to the cluster center was then calculated and assigned to the risk level corresponding to the nearest cluster center. The results are shown in Tables 4 and 5. The original sample set refers to a small number of icing operation scenario samples used in the fine-tuning stage.

[0258] Table 4. Statistics on the proportion of generated icing risk scenarios (considering transfer learning)

[0259]

[0260]

[0261] Table 5. Statistics on the proportion of generated icing risk scenarios (excluding transfer learning).

[0262]

[0263] As shown in Table 4, in the icing scenarios generated using transfer learning, the model significantly improved the generation rate of corresponding scenarios under the guidance of risk labels I, II, and IV compared to the original sample set, increasing by 23.4%, 21.2%, and 14.6%, respectively; while the generation rate under the guidance of risk label III was similar to that of the original sample set. Therefore, the icing scenario generation based on transfer learning can effectively learn the data distribution of the original sample set and generate a large number of corresponding risk samples in a targeted manner, reducing the impact of insufficient training samples on model training.

[0264] As shown in Table 5, when generating icing scenarios without transfer learning, the risk proportion trend of the generated samples is similar to that of the original sample set, regardless of the input risk label and noise. Furthermore, the proportion of Level I and Level II risk samples, which are relatively low in the original sample set, is even lower when generated under the guidance of their corresponding labels, decreasing by 1.4% and 2.8%, respectively. This is because insufficient training samples and other factors can easily lead to model overfitting. An overfitted model may overemphasize common features in the training data while ignoring other risk level features—since Level I and Level II samples are extremely rare, the model struggles to learn their features, resulting in a risk proportion in the generated samples that is more biased towards the frequently occurring Level III and Level IV risks in the training data.

[0265] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for generating source-grid-load operation scenarios under ice storm disturbance based on transfer learning, characterized in that, include: Based on power grid structure data, typical fault type data, and icing disturbance conditions, a simulation sample library for power system operation scenarios is constructed. Based on the power system operation scenario simulation sample library, source domain samples are integrated and standardized to obtain a standardized source domain sample set and its feature matrix. Based on the standardized source domain sample set and its feature matrix, the risk level of the source domain samples is divided to obtain a standardized source domain sample set with risk level labels; Based on a standardized source domain sample set with risk level labels, a least-squares loss function conditional generative adversarial network model is constructed and pre-trained. Based on the standardized target domain sample set and its feature matrix of icing operation samples with risk level labels in the power system operation scenario simulation sample library, the final standardized target domain sample set is generated. Based on the pre-trained least squares loss function conditional generative adversarial network model, combined with the final standardized target domain sample set, transfer learning and target domain model fine-tuning are performed to obtain the fine-tuned target domain ice disaster scene generator. Based on the finely tuned target domain ice disaster scenario generator, an ice disaster disturbance operation scenario is generated.

2. The method for generating source-grid-load operation scenarios under ice disaster disturbance based on transfer learning as described in claim 1, characterized in that, The power system operation scenario simulation sample library constructed based on power grid structure data, typical fault type data, and icing disturbance conditions includes: A simulation sample library of power system operation scenarios is constructed in the form of a three-dimensional matrix. The simulation sample library of power system operation scenarios consists of three types of operation scenario sample sets. Based on power grid structure data, typical fault type data, and icing disturbance conditions, sample simulation work is carried out: Based on the system safety requirements, the overload ratio increase limit of branch power flow and the node voltage constraint limit are determined. Based on the predetermined operating mode, the output combination of wind turbines is changed under different new energy penetration rates for simulation. Steady-state node voltage, branch power flow and frequency data are extracted from the time-domain simulation waveform generated by the simulation, and the node injection power is further calculated to obtain the operating sample set a.

3. The method for generating source-grid-load operation scenarios under ice disaster disturbance based on transfer learning as described in claim 2, characterized in that, The power system operation scenario simulation sample library constructed based on power grid structure data, typical fault type data, and icing disturbance conditions also includes: Set the same constraints as the simulation under normal weather conditions. Based on the predetermined operating mode, determine the typical fault types and write them into the fault text data file. Load them into the same electromechanical transient simulation project. After reasonably setting the maximum simulation parallel threads based on the simulation platform memory space, perform batch simulation. Extract the steady-state node voltage, branch power flow and frequency data from the time-domain simulation waveform generated by the simulation, and further calculate the node injection power to obtain the operating sample set b. Based on a predetermined operating mode, simulations are performed in conjunction with icing disturbance conditions. Steady-state node voltage, branch power flow, and frequency data are extracted from the time-domain simulation waveform generated by the simulation, and the node injection power is further calculated to obtain the operating sample set c.

4. The method for generating source-grid-load operation scenarios under ice storm disturbance based on transfer learning as described in claim 3, characterized in that, The pre-trained least squares loss function-based conditional generative adversarial network model, combined with the final standardized target domain sample set, undergoes transfer learning and target domain model fine-tuning to obtain a fine-tuned target domain ice disaster scene generator, including: The parameters of the source domain generator are transferred, and some parameters of the source domain generator are selected as the initial parameters of the target domain generator; the model parameters of the target domain generator are fine-tuned using the icing sample set; Freeze the parameters of the generator that were successfully trained in the pre-training stage, use the power grid operation scenario samples under icy weather as the fine-tuning sample set and perform data preprocessing; set the number of icy samples to be selected each time the parameters are fine-tuned; randomly select a certain number of samples and their corresponding labels from the fine-tuning sample set to update the parameters; repeat the steps of selecting samples and updating parameters until the model is successfully trained based on the fine-tuning sample set.

5. The method for generating source-grid-load operation scenarios under ice disaster disturbance based on transfer learning as described in claim 4, characterized in that, The process of classifying source domain samples into risk levels based on a standardized source domain sample set and its feature matrix to obtain a standardized source domain sample set with risk level labels includes: The normal scenario sample set, the fault scenario sample set, and the icing operation scenario sample set are integrated into the source domain sample set and then standardized to obtain the standardized source domain sample set and its feature matrix. The sample features are dimensionality reduced to determine the optimal number of clusters. The optimal number of clusters is selected by comprehensively considering the sum of squared errors within clusters and the silhouette coefficient. If the results obtained by different evaluation methods are the same, the result is directly used as the optimal number of clusters; if they are different, the number of clusters corresponding to the largest comprehensive factor is selected as the optimal number of clusters. After determining the optimal number of clusters, the risk level of the samples in the operating scenario is classified.

6. The method for generating source-grid-load operation scenarios under ice disaster disturbance based on transfer learning as described in claim 5, characterized in that, The standardized target domain sample set and its feature matrix, based on the icing operation samples with risk level labels in the power system operation scenario simulation sample library, generate the final standardized target domain sample set, which includes: The scenario represented by all power grid operation samples is taken as the source domain, and the scenario represented by power grid operation samples under icing weather is taken as the target domain. A conditional generative adversarial network model based on the least squares loss function is constructed. The target domain sample set is standardized using ice-covered running samples with risk level labels for iterative training, and the model parameters are dynamically updated using the least squares loss function.

7. The method for generating source-grid-load operation scenarios under ice disaster disturbance based on transfer learning as described in claim 6, characterized in that, The process of generating the final standardized target domain sample set, based on the standardized target domain sample set and its feature matrix of icing operation samples with risk level labels from the power system operation scenario simulation sample library, also includes: Icing simulation and sample set extraction were performed. The same constraints as those used in the simulation and sample set extraction of different output combinations of wind turbines under normal conditions were set. Based on the predetermined operating mode, single or combined conditions such as wind power output reduction, partial line interruption and load demand increase were applied one by one for simulation. Steady-state node voltage, branch power flow and frequency data were extracted from the relevant graph files generated by the simulation, and the node injection power was further calculated to obtain the operating sample set.

8. A source-grid-load operation scenario generation system based on transfer learning under ice storm disturbance, using the method described in any one of claims 1 to 7, characterized in that, include: The module for building a sample library of operating scenarios is used to construct a simulation sample library of power system operating scenarios based on power grid structure data, typical fault type data, and icing disturbance conditions. The source domain sample standardization module is used to integrate source domain samples and standardize them based on the power system operation scenario simulation sample library to obtain a standardized source domain sample set and its feature matrix. The source domain sample risk classification module is used to classify the risk level of source domain samples based on the standardized source domain sample set and its feature matrix, and obtain a standardized source domain sample set with risk level labels. The pre-trained model building module is used to construct a least-squares loss function conditional generative adversarial network model based on a standardized source domain sample set with risk level labels, and to perform pre-training. The target domain sample generation module is used to standardize the target domain sample set and its feature matrix based on the icing operation samples with risk level labels in the power system operation scenario simulation sample library, and generate the final standardized target domain sample set. The target domain model fine-tuning module is used to perform transfer learning and target domain model fine-tuning based on the pre-trained least squares loss function conditional generative adversarial network model, combined with the final standardized target domain sample set, to obtain the fine-tuned target domain ice disaster scene generator. The ice disaster scenario generation module is used to generate ice disaster disturbance operation scenarios based on the fine-tuned target domain ice disaster scenario generator.

9. An electronic device, characterized in that, include: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, which, when executed by the processor, implement the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores computer-executable instructions that, when executed by a processor, implement the steps of the method according to any one of claims 1 to 7.

Images