A method, device, equipment and medium for screening transient stability of Nk fault in power system

Abstract

The present application discloses a method, apparatus, equipment, and medium for screening transient stability of N‑k faults in a power system, relating to the field of power technology. The method comprises: using a preset simulation tool to perform transient time-domain simulation of the power system, adjusting the fault-related parameters of the power system to generate target long-term simulation data under different fault scenarios, swapping the processing order in the deep learning model, constructing an initial accident screening model based on the obtained improved learning model, relationship network, and comparison network, and respectively using the training set corresponding to the target long-term simulation data and the initial accident screening model to obtain the target accident screening model; when a target extreme event occurs, based on the current operating state of the power system and using the preset simulation tool, performing a short-term simulation, inputting the obtained input data into the target accident screening model to obtain an output result of transient stability or transient instability. This method realizes N‑k accident screening of large-scale power systems at low cost.

Description

Technical Field

[0001] The present invention relates to the field of electric power technology, and in particular to a method, device, equipment and medium for screening transient stability of Nk faults in an electric power system. Background Art

[0002] Accident analysis is a pre-disaster safety and stability analysis of potential power system accidents during extreme events (such as extreme weather, natural disasters, or other sudden failures). To improve the efficiency of accident analysis, accident screening techniques are often used before detailed analysis. However, in practice, transient stability issues are more valuable to system planners than thermal stability, voltage overshoot, and static stability issues due to their faster dynamic evolution, wider impact, and more severe consequences. They serve as a basis and reference for dispatchers to formulate contingency plans and control strategies for high-risk accidents. With the increasing complexity of power systems, especially under the influence of extreme weather, the probability of component failure increases significantly. Simultaneous or near-simultaneous failure of multiple components is common, and Nk accidents are likely to have more severe consequences than N-1 accidents. Therefore, Nk accident screening is crucial for improving the power system's ability to cope with low-probability, high-risk extreme events.

[0003] Currently, research on transient stability fault screening considering dynamic processes primarily focuses on model-driven and data-driven approaches. Time-domain simulation methods offer accurate results but are computationally time-consuming. To enhance the adaptability and interpretability of data-driven models in complex and volatile power grid scenarios and under massive amounts of data, a number of emerging architectures have been proposed. However, for large-scale power systems, labeling Nk training samples based on typical operating scenarios is computationally expensive, and existing research is unable to generalize to Nk unseen faults by labeling a subset of Nk samples.

[0004] As can be seen from the above, how to achieve Nk fault screening of large-scale power systems at low cost is an urgent problem to be solved. Summary of the Invention

[0005] In view of this, the purpose of the present invention is to provide a method, device, equipment and medium for screening transient stability of power system Nk faults, which can realize Nk accident screening of large-scale power systems at low cost. The specific scheme is as follows:

[0006] In a first aspect, the present application provides a method for screening transient stability of an Nk fault in a power system, comprising:

[0007] Using a preset simulation tool to perform transient time-domain simulation of the power system, and adjusting fault-related parameters of the power system during the simulation process to generate target long-term simulation data under different fault scenarios, and standardizing the target long-term simulation data to obtain a training set and a validation set;

[0008] Reversing the processing order of layer normalization and residual connection in the deep learning model to obtain an improved learning model, constructing an initial accident screening model based on the improved learning model, the preset learning method, the relationship network, and the comparison network, and training and validating the initial accident screening model using the training set and the validation set, respectively, to obtain a target accident screening model;

[0009] When a target extreme event occurs, a short-term simulation is performed based on the current operating state of the power system and using the preset simulation tool to generate input data, and the input data is input into the target accident screening model to obtain an output result of transient stability or transient instability.

[0010] Optionally, the method of performing transient time-domain simulation on the power system using a preset simulation tool and adjusting fault-related parameters of the power system during the simulation process to generate target long-term simulation data under different fault scenarios includes:

[0011] Using preset simulation software and preset simulation duration, a three-phase short circuit fault is set at any preset location on the transmission line of the power system to obtain low-order fault data of a single component fault;

[0012] Setting a three-phase short circuit fault at any preset location on the transmission line based on the preset simulation software and the preset simulation duration, and adjusting parameters of the number of line breaks, load level, fault location, and fault line of the power system during the simulation to obtain high-order fault data of simultaneous failure of multiple components;

[0013] The low-order fault data and the high-order fault data are used to obtain target long-term simulation data under different fault scenarios.

[0014] Optionally, the standardizing the target long-term simulation data to obtain a training set and a validation set includes:

[0015] determining input samples including node voltage magnitude, phase angle, active power, reactive power, and electrical frequency based on the target long-term simulation data;

[0016] Determining a maximum generator power angle difference corresponding to each of the input samples, and determining a sample category corresponding to the input sample based on the maximum generator power angle difference;

[0017] A training sample is determined based on the input sample and the corresponding sample category, and the training sample is standardized to obtain a standardized sample. The standardized sample is used to determine a training set and a validation set.

[0018] Optionally, determining the maximum generator power angle difference corresponding to each input sample, and determining the sample category corresponding to the input sample based on the maximum generator power angle difference, includes:

[0019] determining a maximum generator power angle difference corresponding to each of the input samples, and determining whether the maximum generator power angle difference exceeds a target power angle difference threshold;

[0020] If the maximum generator power angle difference exceeds a target power angle difference threshold, determining that the sample category of the input sample corresponding to the maximum generator power angle difference that exceeds the target power angle difference threshold is transient instability;

[0021] If the maximum generator power angle difference does not exceed the target power angle difference threshold, it is determined that the sample category of the input sample corresponding to the maximum generator power angle difference that does not exceed the target power angle difference threshold is transient stability.

[0022] Optionally, constructing an initial accident screening model based on the improved learning model, the preset learning method, the relationship network, and the comparison network includes:

[0023] Determining the standardized samples corresponding to the low-order fault data as a support set for meta-learning, and determining the standardized samples corresponding to the high-order fault data as a query set for meta-learning;

[0024] A meta-learning task is constructed based on the support set and the query set, and an initial accident screening model is constructed based on the meta-learning task, the improved learning model, the relational network and the comparison network.

[0025] Optionally, the training and validating the initial accident screening model using the training set and the validation set respectively to obtain a target accident screening model includes:

[0026] Combining the support set and the query set to obtain an embedded feature, and determining a corresponding relationship score based on the embedded feature and using a relationship network in the initial accident screening model;

[0027] Converting the support set and the query set into one-dimensional vectors to obtain first flattened features and second flattened features corresponding to the support set and the query set, respectively;

[0028] determining a synergy score based on the first flattened feature and the second flattened feature and using a comparison network in the initial accident screening model, and determining a prediction result using the relationship score and the synergy score;

[0029] Determining a loss value using the prediction result and the sample categories in the training set, and updating the model parameters in the initial accident screening model based on the loss value to obtain a trained accident screening model;

[0030] The trained accident screening model is verified based on the verification set to obtain a target accident screening model.

[0031] Optionally, the verifying the trained accident screening model based on the verification set to obtain a target accident screening model includes:

[0032] Verifying the trained accident screening model based on the validation set to obtain a verified model, and performing a performance evaluation on the verified model using a preset confusion matrix and various performance evaluation indicators to obtain corresponding evaluation results;

[0033] Based on the evaluation results and using a grid search model, the model parameters of the verified model are adjusted to obtain a target accident screening model.

[0034] In a second aspect, the present application provides a low-cost Nk accident screening device for large-scale power systems, comprising:

[0035] A simulation module is used to perform transient time-domain simulation of the power system using a preset simulation tool, and to adjust fault-related parameters of the power system during the simulation process to generate target long-term simulation data under different fault scenarios, and to standardize the target long-term simulation data to obtain a training set and a validation set;

[0036] a target model determination module, configured to swap the processing order of layer normalization and residual connection in the deep learning model to obtain an improved learning model, construct an initial accident screening model based on the improved learning model, a preset learning method, a relationship network, and a comparison network, and train and verify the initial accident screening model using the training set and the validation set, respectively, to obtain a target accident screening model;

[0037] The output result determination module is used to generate input data by performing short-term simulation based on the current operating state of the power system and using the preset simulation tool when a target extreme event occurs, and input the input data into the target accident screening model to obtain an output result of transient stability or transient instability.

[0038] In a third aspect, the present application provides an electronic device, comprising:

[0039] Memory, used to store computer programs;

[0040] A processor is used to execute the computer program to implement the aforementioned power system Nk fault transient stability screening method.

[0041] In a fourth aspect, the present application provides a computer-readable storage medium for storing a computer program, wherein the computer program, when executed by a processor, implements the aforementioned power system Nk fault transient stability screening method.

[0042] This application uses a preset simulation tool to perform transient time-domain simulation on the power system, and adjusts the fault-related parameters of the power system during the simulation process to generate target long-term simulation data under different fault scenarios, and standardizes the target long-term simulation data to obtain a training set and a verification set; the processing order of the middle-layer normalization and residual connection in the deep learning model is swapped to obtain an improved learning model, and an initial accident screening model is constructed based on the improved learning model, the preset learning method, the relationship network and the comparison network, and the initial accident screening model is trained and verified using the training set and the verification set respectively to obtain a target accident screening model; when a target extreme event occurs, short-term simulation is performed based on the current operating state of the power system and using the preset simulation tool to generate input data, and the input data is input into the target accident screening model to obtain an output result of transient stability or transient instability.

[0043] As can be seen from the above, the present application uses a preset simulation tool to simulate the transient time domain of the power system, and adjusts the fault-related parameters to generate target long-term simulation data under different fault scenarios, which can cover as many fault conditions as possible. Then, the processing order in the deep learning model is changed to obtain an improved learning model. This improvement may make the model more suitable for handling the transient stability screening task of the power system, improve the model's processing ability and feature extraction effect for the input data, and then construct an initial accident screening model based on the improved learning model, the preset learning method, the relationship network and the comparison network. The initial accident screening model is trained and verified through the training set and the validation set, which can continuously optimize the model parameters, improve the accuracy and generalization ability of the model, and finally obtain a target accident screening model with good performance. In this way, when a target extreme event occurs, based on the current operating state of the power system and using the preset simulation tool to generate input data, and then input the input data into the target accident screening model, the output results of transient stability or transient instability can be quickly obtained, without the need to build an actual physical system for testing, saving a lot of time and cost, thereby helping operators to take timely measures to prevent the expansion of accidents. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0045] Figure 1 This is a flow chart of a method for screening transient stability of power system Nk faults disclosed in this application;

[0046] Figure 2 A schematic diagram of target long-term simulation data provided by this application;

[0047] Figure 3 A schematic diagram of an embedded module in a relationship network provided by this application;

[0048] Figure 4 A schematic diagram of an improved deep learning model provided in this application; Figure 4 (a) is the deep learning model before improvement; Figure 4 (b) is the improved deep learning model;

[0049] Figure 5 The accuracy of model training based on different datasets provided for this application; Figure 5 (a) is the model training accuracy based on dataset B; Figure 5 (b) is the model training accuracy based on dataset C; Figure 5 (c) is the model training accuracy based on dataset D;

[0050] Figure 6 A schematic diagram of a model evaluation index provided for this application;

[0051] Figure 7 The accuracy of the model based on different multiples of the training set size provided for this application;

[0052] Figure 8 A schematic diagram of time comparison provided for this application;

[0053] Figure 9 A flow chart of a specific method for screening transient stability of power system Nk faults provided in this application;

[0054] Figure 10 This is a schematic structural diagram of a power system Nk fault transient stability screening device disclosed in this application;

[0055] Figure 11This is a structural diagram of an electronic device disclosed in this application. DETAILED DESCRIPTION

[0056] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0057] At present, the research on transient stability accident screening considering dynamic processes is mainly carried out from two aspects: model-driven and data-driven. Although the time domain simulation method has accurate results, it is time-consuming to calculate. For large-scale power systems, the computational time cost of labeling Nk training samples according to typical operating scenarios is very high. Existing research cannot be applied to labeling some Nk samples and generalizing to Nk accidents that have never been seen. To this end, the present application provides a method for transient stability screening of Nk faults in power systems. When a target extreme event occurs, input data is generated based on the current operating state of the power system and using a preset simulation tool. The input data is then input into the target accident screening model. The output results of transient stability or transient instability can be quickly obtained without building an actual physical system for testing, saving a lot of time and cost, thereby helping operators to take timely measures to prevent the expansion of accidents.

[0058] See also Figure 1 As shown, an embodiment of the present invention discloses a method for screening transient stability of a power system Nk fault, comprising:

[0059] Step S11: Use a preset simulation tool to perform transient time-domain simulation on the power system, and adjust the fault-related parameters of the power system during the simulation process to generate target long-term simulation data under different fault scenarios, and standardize the target long-term simulation data to obtain a training set and a validation set.

[0060] In this embodiment, based on the power system (such as the topology and load level of the IEEE39 node system) and using a preset simulation tool to simulate transient processes under different fault scenarios, time domain waveform data containing electrical quantities such as node voltage amplitude, phase angle, active / reactive power, and electrical frequency are generated to obtain target long-term simulation data; the preset simulation tool can be Powerfactory time domain simulation software. In a specific embodiment, Figure 2A schematic diagram of target long-term simulation data provided for this embodiment. The fault occurs on a transmission line that does not include a transformer branch. The total simulation time is 10 seconds, and the simulation step is 0.02s, that is, 50 data points are collected per second, for a total of 500 time points. However, this embodiment records data from 20 sampling points from 0 to 0.2s at intervals of 1, that is, it focuses on the transient initial stage after the fault occurs. Then, a three-phase short circuit fault is set to occur at 10%, 50% or 90% of the transmission line at 0.02s, and the fault duration is 0.15s. The active and reactive power demands of the load vary evenly in the range of 90% to 110% with a step size of 10%, that is, the load level can be set to 90%, 100%, and 110% to obtain low-order fault data of a single component fault, namely, data set A. The fault clearing strategy is to cut off the fault line. Then, three-phase short-circuit faults occur simultaneously at two locations at 10%, 50%, or 90% of the transmission line, ignoring the occurrence of two faults on the same line, to obtain data set B; three-phase short-circuit faults are set at 50% of each line, and the three fault points are triggered simultaneously to obtain data set C; three-phase short-circuit faults are set at 50% of each line, and the four fault points are triggered simultaneously to obtain data set D; 3000, 3000, and 7000 non-repeating combinations are randomly selected from the N-2, N-3, and N-4 combinations of the line, respectively, and the specific fault locations are set to random numbers in the line length interval (1%, 100%). All fault points are triggered simultaneously to form data sets E, F, and G, which are used as test sets respectively.

[0061] Specifically, the preset simulation tool is used to perform transient time domain simulation on the power system, and the fault-related parameters of the power system are adjusted during the simulation process to generate target long-term simulation data under different fault scenarios, including: using the preset simulation software and the preset simulation time to set a three-phase short circuit fault at any preset position on the transmission line of the power system to obtain low-order fault data of a single component failure; based on the preset simulation software and the preset simulation time, a three-phase short circuit fault is set at any preset position on the transmission line, and during the simulation process, the number of line breaks, load level, fault location, and fault line of the power system are adjusted to obtain high-order fault data of simultaneous failure of multiple components; and the low-order fault data and the high-order fault data are used to obtain target long-term simulation data under different fault scenarios.

[0062] It is understandable that after obtaining the target long-term simulation data, input samples including node voltage amplitude, phase angle, active power, reactive power, and electrical frequency are determined based on data sets B, C, and D in the target long-term simulation data, and the maximum generator power angle difference corresponding to each input sample is determined. The formula corresponding to the maximum generator power angle difference is:

[0063] ;

[0064] in, is the maximum generator power angle difference; is the simulation duration; For the moment No. The power angle of the generator; For the moment No. The power angle of each generator; the power angle is the phase difference between the generator potential and the system bus voltage, reflecting the operating status of the generator and the stability of the system. In one specific embodiment, if the maximum generator power angle difference exceeds 360 degrees, the sample category of the input sample corresponding to the maximum generator power angle difference exceeding the target power angle difference threshold is determined to be transient instability; if the maximum generator power angle difference does not exceed 360 degrees, the sample category of the input sample corresponding to the maximum generator power angle difference not exceeding the target power angle difference threshold is determined to be transient stability.

[0065] Specifically, the determination of the maximum generator power angle difference corresponding to each input sample, and the determination of the sample category corresponding to the input sample based on the maximum generator power angle difference, includes: determining the maximum generator power angle difference corresponding to each input sample, and judging whether the maximum generator power angle difference exceeds a target power angle difference threshold; if the maximum generator power angle difference exceeds the target power angle difference threshold, judging that the sample category of the input sample corresponding to the maximum generator power angle difference exceeding the target power angle difference threshold is transient instability; if the maximum generator power angle difference does not exceed the target power angle difference threshold, judging that the sample category of the input sample corresponding to the maximum generator power angle difference not exceeding the target power angle difference threshold is transient stability. It is worth mentioning that the target power angle difference threshold can be adjusted according to actual conditions and is not specifically limited here.

[0066] Furthermore, after obtaining the sample category corresponding to the input sample, a training sample is determined based on the input sample and the corresponding sample category, and the training sample is normalized using the z-score to obtain a normalized sample. The corresponding formula is as follows:

[0067] ;

[0068] in, is a feature of a certain dimension of the training sample; is the mean of all data of a single feature (such as voltage amplitude); In a specific embodiment, after obtaining the standardized samples, the training set and the validation set are evenly divided into two parts according to the ratio of 1:7 and 1:19, respectively.

[0069] Specifically, the target long-term simulation data is standardized to obtain a training set and a validation set, including: determining input samples including node voltage amplitude, phase angle, active power, reactive power and electrical frequency based on the target long-term simulation data; determining the maximum generator power angle difference corresponding to each input sample, and determining the sample category corresponding to the input sample based on the maximum generator power angle difference; determining training samples based on the input samples and the corresponding sample categories, and standardizing the training samples to obtain standardized samples, and using the standardized samples to determine the training set and the validation set.

[0070] Step S12: swap the processing order of layer normalization and residual connection in the deep learning model to obtain an improved learning model, construct an initial accident screening model based on the improved learning model, the preset learning method, the relationship network and the comparison network, and use the training set and the verification set to train and verify the initial accident screening model respectively to obtain a target accident screening model.

[0071] In this embodiment, the order of processing mid-layer normalization and residual connections in the Transformer model (i.e., deep learning model) is swapped to obtain an improved learning model. The standardized samples corresponding to the low-order fault data are determined as the support set for meta-learning, and the standardized samples corresponding to the high-order fault data are determined as the query set for meta-learning. Since the basic unit of meta-learning is a task, and each task consists of a support set and a query set, a meta-learning task is constructed based on the support set and the query set. An initial accident screening model is constructed based on the meta-learning task, the improved learning model, the relational network, and the comparison network.

[0072] Figure 3 This is a schematic diagram of an embedding module in a relationship network provided in this embodiment, where the samples in the query set Q are Samples in the support set S After processing by the embedding module, the embedding module uses the improved Transformer learning model to generate embedding features and For each busbar there is a corresponding two-dimensional tensor As input, the two-dimensional tensor is then linearly transformed through the attention head to obtain the corresponding query (Query, Q), key (Key, K) and value (Value, V) matrices; then the attention weight is determined by the dot product of Q and K, and the scaled dot product attention mechanism is used to determine the new two-dimensional tensor, that is, the target two-dimensional tensor, and then the target two-dimensional tensor is reshaped by dimensionality reduction to obtain a one-dimensional tensor ;in, represents the learnable parameters in the embedding module. Similarly, we get The corresponding one-dimensional tensor , concatenate the above two one-dimensional vectors to obtain the final output tensor of the embedding module.

[0073] Furthermore, by swapping the processing order of layer normalization and residual connection in the deep learning model and removing the position encoding module, an improved learning model is obtained. Figure 4 A schematic diagram of an improved deep learning model provided in this embodiment; Figure 4 (a) It can be seen that in the deep learning model before the improvement, the embedded feature vector needs to be added to the position code first, and then input into the multi-head attention structure, and then the residual connection is performed before the layer normalization. However, performing layer normalization on the mixture of the original input and the sub-layer output may cause excessive smoothing of the key information of the original input; Figure 4 (b) Directly skip the position encoding and addition layer to avoid the position encoding interfering with the original features, directly perform layer normalization on the embedded feature vector, then determine the attention output result, and finally directly superimpose the original input and the normalized sub-layer output through the residual connection, which can ensure that the unnormalized original is losslessly transmitted to the subsequent network and prevent the transition smoothing of important features by layer normalization. In summary, the process of the deep learning model before the improvement is as follows: embedding feature vectors, position encoding and addition layers, multi-head attention structure, layer normalization, feedforward network, layer normalization; among them, the deep learning model before the improvement does not have explicit residuals. Although the residual logic of the input and attention output results is implicit, it is smoothed by layer normalization. The process of the improved deep learning model is as follows: embedding feature vectors, layer normalization, multi-head attention, residual connection, layer normalization, feedforward network, residual connection; among them, the improved deep learning model retains the mutation characteristics in the original features. The internal data processing process of TransformerBlock in the improved learning model is as follows:

[0074] ;

[0075] ;

[0076] in, It is the intermediate result after multi-head attention mechanism and layer normalization processing; is layer normalization; Input data for the multi-head attention mechanism The processing results; The final output of the Transformer Block in the improved learning model; For the intermediate results of the feedforward neural network After obtaining the improved learning model, a meta-learning task is constructed based on the support set and the query set, and an initial accident screening model is constructed based on the meta-learning task, the improved learning model, the relational network, and the comparison network.

[0077] Specifically, the constructing of the initial accident screening model based on the improved learning model, the preset learning method, the relational network and the comparative network includes: determining the standardized samples corresponding to the low-order fault data as the support set for meta-learning, and determining the standardized samples corresponding to the high-order fault data as the query set for meta-learning; constructing a meta-learning task based on the support set and the query set, and constructing an initial accident screening model based on the meta-learning task, the improved learning model, the relational network and the comparative network.

[0078] In this embodiment, after obtaining the final output tensor of the embedding module, i.e., the embedded feature, the embedded feature is input into the relationship module in the relationship network. The relationship module consists of two fully connected layers and outputs a scalar between 0 and 1, representing and The similarity between them is used to get the relationship score. For a query set of NK, input and the corresponding support set , the relationship scores are as follows:

[0079] ;

[0080] in, for and The relationship score between is an activation function, which represents about Perform normalization operation; It is a combination operation, which here means splicing or integration; For the relationship module; and They are and The corresponding one-dimensional tensor.

[0081] It is understandable that after obtaining the relationship score, a comparison network is designed to distinguish whether low-order faults that are transiently stable in a single fault scenario will cause transient instability when combined into high-order faults. The comparison network consists of three fully connected layers. Based on the first flattened features and the second flattened features corresponding to the support set and the query set, respectively, the nonlinear relationship between the features is determined and learned to obtain a synergy effect score. The formula for the synergy effect score is as follows:

[0082] ;

[0083] in, assigning a score to the synergistic effect; To compare network functions; is the first flattening feature; The second flattened feature is represented by the relationship score and the synergy score. The relationship score and the synergy score are added to determine the final prediction result. The initial accident screening model is then forward trained using the training set, and the training set loss is backpropagated to optimize the model's internal parameters.

[0084] Specifically, the initial accident screening model is trained and verified using the training set and the verification set respectively to obtain a target accident screening model, including: combining the support set and the query set to obtain embedded features, and determining corresponding relationship scores based on the embedded features and using the relationship network in the initial accident screening model; converting the support set and the query set into one-dimensional vectors to obtain first flattened features and second flattened features corresponding to the support set and the query set respectively; determining a synergy score based on the first flattened features and the second flattened features and using the comparison network in the initial accident screening model, and determining a prediction result using the relationship score and the synergy score; determining a loss value using the prediction result and the sample category in the training set, and updating the model parameters in the initial accident screening model based on the loss value to obtain a trained accident screening model; and verifying the trained accident screening model based on the verification set to obtain a target accident screening model.

[0085] Furthermore, the grid search model is used to optimize the model's accuracy by using its optimal hyperparameters. The hyperparameters of this embodiment are set as follows: the number of iterations is 500, the batch size is 200, the learning rate is 0.0005, the number of layers of the improved learning model Transformer layer is 3, the hidden layer of the feedforward neural network is set to 4 times the input dimension of the upper layer, the hidden layer dimension of the relational network fully connected layer is 15 dimensions, the hidden layer dimension of the comparison network fully connected layer is 300 and 30 dimensions, and a dropout with a probability of 0.5 is set. It is worth mentioning that the hyperparameter settings can also be adjusted accordingly according to actual conditions. Then, the trained accident screening model is verified and its performance is evaluated. After the trained accident screening model is verified using the verification set to obtain the verified model, based on the preset confusion matrix , and use accuracy, F1 score, and FNR value to evaluate the performance of the verified model. The calculation formulas for accuracy, F1 score, and FNR value are as follows:

[0086] ;

[0087] ;

[0088] ;

[0089] in, 、 、 、 They represent the number of samples that are correctly predicted to be stable, predicted to be unstable when they are actually stable, predicted to be stable when they are actually unstable, and correctly predicted to be unstable; is the accuracy rate; is the F1 score; is the FNR value.

[0090] In this embodiment, after obtaining the evaluation results including the accuracy rate, the F1 score, and the FNR value, the model parameters of the verified model are adjusted based on the evaluation results to obtain a target accident screening model. Specifically, verifying the trained accident screening model based on the validation set to obtain the target accident screening model includes: verifying the trained accident screening model based on the validation set to obtain a verified model, and performing a performance evaluation on the verified model using a preset confusion matrix and various performance evaluation indicators to obtain a corresponding evaluation result; and adjusting the model parameters of the verified model based on the evaluation result and using a grid search model to obtain the target accident screening model.

[0091] Step S13: When a target extreme event occurs, a short-term simulation is performed based on the current operating state of the power system and using the preset simulation tool to generate input data, and the input data is input into the target accident screening model to obtain an output result of transient stability or transient instability.

[0092] In this embodiment, when a target extreme event occurs, input data to be processed is generated based on the current operating state of the power system using a pre-set simulation tool. The standardized voltage amplitude, phase angle, injected active power, and reactive power data corresponding to the processed input data are determined as input data and applied to the target accident screening model to obtain transient stability and transient instability identification results. The target extreme event includes natural disasters, such as transient crisis scenarios caused by multiple power system faults, whose core characteristics are low probability, high risk, and strong nonlinearity. To verify the effectiveness of this embodiment, training and testing were performed on datasets B, C, and D, respectively. Furthermore, to compare the generalization capabilities of this embodiment's target accident screening model with existing deep learning models using small sample sizes, conventional data-driven architectures used in current research, such as graph convolutional networks (GCNs), support vector machines (SVMs), convolutional neural networks (CNNs), and transformers (deep learning models based on self-attention mechanisms), were selected in their optimal configurations as comparison models. The target accident screening model and the comparison model of this embodiment are trained on data sets B, C, and D, where: Figure 5 (a) is the model training accuracy corresponding to data set B, Figure 5 (b) is the model training accuracy corresponding to dataset C; Figure 5 (c) is the model training accuracy corresponding to the dataset D; the accuracy change of the validation set during the training process is shown in Figure 5 As shown, the converged model performs well on the test set in terms of various evaluation indicators. Figure 6 As shown. Figure 5 It can be seen that the training set (i.e., T-MLCF training set) and the validation set (i.e., T-MLCF validation set) of the target accident screening model of this embodiment tend to converge at the 100th epoch, and all four models show slight overfitting. Figure 6 It can be seen that the three indicators of Acc, IF1, and IFNR of the target accident screening model of this embodiment are better than those of other comparison models on the three test sets representing N-2, N-3, and N-4.

[0093] It is understandable that due to the difficulty in enumerating Nk faults in the power system, for small sample solutions, sample size is the key to reflecting model performance and reducing simulation costs. Therefore, we retrained the model based on datasets A, B, C, and D at 0.8, 2, and 6 times the size of the selected training set, and tested it on datasets E, F, and G. The accuracy rates are as follows: Figure 7As shown. The three data sets all show that when the number of training set samples is between 0.8 and 2 times the original size, the model performance does not change significantly; when the number of training set samples is 6 times the original size, the accuracy of the three data sets all reaches more than 99%. This shows that the target accident screening model of the model in this embodiment can maintain competitiveness under small sample and sufficient sample conditions. The model time consumption is divided into offline time and online time. The offline time includes the time for generating the training set and the training time, and the online time includes the time for short-term simulation to generate the samples to be evaluated and the online evaluation time. In order to reflect the time efficiency of the target accident screening model, it is compared with the best Transformer in the comparison model to observe the time taken for the accuracy achieved by the training data set B, C, and D tests to be consistent with the target accident screening model. Time comparison Figure 8 As shown in the figure, due to the huge space of Nk fault samples and the high computational cost of time-domain simulation, the time difference is mainly reflected in the time consumption of simulation sample labeling. It can be seen that the target accident screening model (i.e., TMLCF model) has a significant time advantage, and the advantage becomes stronger with the increase of the fault order k.

[0094] As can be seen from the above, the present application uses a preset simulation tool to simulate the transient time domain of the power system, and adjusts the fault-related parameters to generate target long-term simulation data under different fault scenarios, which can cover as many fault conditions as possible. Then, the processing order in the deep learning model is changed to obtain an improved learning model. This improvement may make the model more suitable for handling the transient stability screening task of the power system, improve the model's processing ability and feature extraction effect for the input data, and then construct an initial accident screening model based on the improved learning model, the preset learning method, the relationship network and the comparison network. The initial accident screening model is trained and verified through the training set and the validation set, which can continuously optimize the model parameters, improve the accuracy and generalization ability of the model, and finally obtain a target accident screening model with good performance. In this way, when a target extreme event occurs, based on the current operating state of the power system and using the preset simulation tool to generate input data, and then input the input data into the target accident screening model, the output results of transient stability or transient instability can be quickly obtained, without the need to build an actual physical system for testing, saving a lot of time and cost, thereby helping operators to take timely measures to prevent the expansion of accidents.

[0095] It can be seen from the above embodiments that the present application realizes efficient screening of transient stability accidents based on the meta-learning framework and the target accident screening model of the improved Transformer. Therefore, the process of realizing efficient screening of transient stability accidents based on the meta-learning framework and the target accident screening model of the improved Transformer is described.

[0096] See also Figure 9As shown, the embodiment of the present invention discloses a specific method for screening transient stability of power system Nk fault, including:

[0097] In this embodiment, the long-term labeled small sample data is equivalent to low-order fault data, and mainly focuses on low-order fault scenarios, such as N-1 faults. These data are derived from historical simulations, actual power grid accident replays, and other work, and each piece of data is clearly labeled with transient stability or instability labels; the preset fault set includes the types of faults that the power system may face, from simple N-1 faults to complex Nk faults. In the process of using the preset simulation tool to perform transient time domain simulation of the power system, the fault line location factor is taken into account. For example, faults are set at different locations such as 10%, 50%, and 90% of the transmission line, and the combination of the number of faults is also included, including The simulation included complex scenarios such as single-line faults and simultaneous faults of multiple lines. The load was considered to fluctuate between 90% and 110% of the rated load, including random fluctuations, to simulate the complex load conditions of the actual power grid. The fault occurred on a transmission line excluding the transformer branch. The location could be random or set at a fixed ratio (e.g., 10%, 50%, and 90%). The fault duration was set to 0.15s. The target long-term simulation data was focused on critical transient periods, such as the 0-0.2s period. Data was recorded at intervals of 1, for a total of 20 sampling points.

[0098] Next, the target long-term simulation data is standardized to obtain the corresponding training set and validation set. The order of layer normalization and residual connection in the traditional Transformer is swapped, with the residual connection in front, to ensure that the original input signal can be transmitted to the subsequent network without loss, to avoid excessive smoothing of important features by layer normalization, so that the original features in the fault transient process (such as voltage drop, frequency mutation, etc.) can be better preserved, and the position encoding module is removed to obtain an improved learning model. The embedding module receives the standardized target long-term simulation data; the input of the relationship module is the embedded features of the low-order fault data and the high-order fault data, and the relationship score is determined based on the embedded features and the relationship network. The input of the comparison module is the first flattened feature corresponding to the low-order fault data and the second flattened feature corresponding to the high-order fault data, and the comparison module outputs the synergistic effect score of the response , the relationship score output by the relationship network and the synergy score output by the comparison network are added and fused to obtain the final output result.

[0099] Furthermore, during the model training phase, the model continuously evaluates the degree of match between the output results and the true label. The true label is determined by the maximum generator power angle difference. When the maximum generator power angle difference is greater than the target power angle difference threshold, the system is considered transiently unstable; otherwise, it is transiently stable. The performance of the trained accident screening model (i.e., the trained T-MLCF transient stability accident screening model) is then measured by calculating indicators such as accuracy, F1 score, and FNR value. If these indicators do not meet the preset accuracy requirements or there is overfitting, the model parameters are adjusted in the opposite direction. These parameters include the weights of the improved Transformer, the parameters of the fully connected layers of the relational network and the comparison network, etc., and then the embedding, relational / comparison operations, and other processes are re-performed, with continuous iterative optimization until the model performance is optimized, thus obtaining the target accident screening model.

[0100] When a target extreme event occurs, a short-term simulation is performed based on the current operating state of the power system and using the preset simulation tool to generate input data, and the input data is input into the target accident screening model to obtain an output result of transient stability or transient instability. If the output result is transient stability, it means that the current power grid has not yet encountered the risk of instability under the target extreme event, and the subsequent continuous monitoring of the power grid operation status can ensure the stable operation of the power grid; if the output result is transient instability, a detailed accident analysis link is carried out to deeply analyze the fault development process and explore key instability factors, such as which low-order fault combinations have produced synergistic instability effects, and which nodes' electrical quantity changes after the fault occur are the key to causing instability. Based on these analyses, emergency control strategies are formulated and timely measures are taken to ensure the safety of the power grid and avoid the expansion of accidents.

[0101] As can be seen from the above, the present application simulates the transient time domain of the power system and adjusts the fault-related parameters to generate target long-term simulation data under different fault scenarios, and then exchanges the processing order in the deep learning model to obtain an improved learning model, and constructs an initial accident screening model based on the improved learning model, the preset learning method, the relationship network and the comparison network. The initial accident screening model is trained and verified through the training set and the verification set, and the model parameters are continuously optimized to obtain a target accident screening model with good performance. When a target extreme event occurs, input data is generated based on the current operating state of the power system and using a preset simulation tool, and the target accident screening model is used to quickly obtain the output results of transient stability or transient instability. In this way, when facing a large-scale power system, accurate transient stability accident screening can be performed, which effectively improves the ability of the power system to cope with complex faults and ensure transient stability.

[0102] Accordingly, see Figure 10 As shown, the present application also provides a power system Nk fault transient stability screening device, comprising:

[0103] A simulation module 11 is configured to perform transient time-domain simulation of the power system using a preset simulation tool, and to adjust fault-related parameters of the power system during the simulation process to generate target long-term simulation data under different fault scenarios, and to standardize the target long-term simulation data to obtain a training set and a validation set;

[0104] a target model determination module 12, configured to swap the processing order of layer normalization and residual connection in the deep learning model to obtain an improved learning model, construct an initial accident screening model based on the improved learning model, a preset learning method, a relationship network, and a comparison network, and train and verify the initial accident screening model using the training set and the validation set, respectively, to obtain a target accident screening model;

[0105] The output result determination module 13 is used to generate input data by performing short-term simulation based on the current operating state of the power system and using the preset simulation tool when a target extreme event occurs, and input the input data into the target accident screening model to obtain an output result of transient stability or transient instability.

[0106] As can be seen from the above, the present application uses a preset simulation tool to simulate the transient time domain of the power system, and adjusts the fault-related parameters to generate target long-term simulation data under different fault scenarios, which can cover as many fault conditions as possible. Then, the processing order in the deep learning model is changed to obtain an improved learning model. This improvement may make the model more suitable for handling the transient stability screening task of the power system, improve the model's processing ability and feature extraction effect for the input data, and then construct an initial accident screening model based on the improved learning model, the preset learning method, the relationship network and the comparison network. The initial accident screening model is trained and verified through the training set and the validation set, which can continuously optimize the model parameters, improve the accuracy and generalization ability of the model, and finally obtain a target accident screening model with good performance. In this way, when a target extreme event occurs, based on the current operating state of the power system and using the preset simulation tool to generate input data, and then input the input data into the target accident screening model, the output results of transient stability or transient instability can be quickly obtained, without the need to build an actual physical system for testing, saving a lot of time and cost, thereby helping operators to take timely measures to prevent the expansion of accidents.

[0107] In some specific implementations, the simulation module 11 may specifically include:

[0108] a low-order fault data determining unit, configured to set a three-phase short circuit fault at any preset location on a transmission line of the power system using preset simulation software and a preset simulation duration, so as to obtain low-order fault data of a single component fault;

[0109] a low-order fault data determining unit, configured to set a three-phase short circuit fault at any preset location on the transmission line based on the preset simulation software and the preset simulation duration, and adjust parameters of the number of line breaks, load level, fault location, and fault line of the power system during the simulation process to obtain high-order fault data of simultaneous failure of multiple components;

[0110] The target simulation data determining unit is used to obtain target long-term simulation data under different fault scenarios using the low-order fault data and the high-order fault data.

[0111] In some specific implementations, the simulation module 11 may specifically include:

[0112] an input sample determination unit, configured to determine input samples including node voltage amplitude, phase angle, active power, reactive power, and electrical frequency based on the target long-term simulation data;

[0113] a maximum power angle difference determining unit, configured to determine a maximum generator power angle difference corresponding to each input sample, and determine a sample category corresponding to the input sample based on the maximum generator power angle difference;

[0114] The standardized sample determination unit is used to determine training samples based on the input samples and corresponding sample categories, and to perform standardization processing on the training samples to obtain standardized samples, and to determine a training set and a validation set using the standardized samples.

[0115] In some specific implementations, the simulation module 11 may specifically include:

[0116] a maximum power angle difference judgment unit, configured to determine a maximum generator power angle difference corresponding to each input sample, and to judge whether the maximum generator power angle difference exceeds a target power angle difference threshold;

[0117] a first sample category determination unit, configured to determine, if the maximum generator power angle difference exceeds a target power angle difference threshold, that the sample category of the input sample corresponding to the maximum generator power angle difference exceeding the target power angle difference threshold is transient instability;

[0118] The second sample category determination unit is configured to determine that the sample category of the input sample corresponding to the maximum generator power angle difference that does not exceed the target power angle difference threshold is transient stability if the maximum generator power angle difference does not exceed the target power angle difference threshold.

[0119] In some specific implementations, the target model determination module 12 may specifically include:

[0120] a support set determining unit, configured to determine the standardized samples corresponding to the low-order fault data as a support set for meta-learning, and to determine the standardized samples corresponding to the high-order fault data as a query set for meta-learning;

[0121] An initial screening model determination unit is used to construct a meta-learning task based on the support set and the query set, and to construct an initial accident screening model based on the meta-learning task, the improved learning model, the relational network and the comparative network.

[0122] In some specific implementations, the target model determination module 12 may specifically include:

[0123] a relationship score determination unit, configured to combine the support set and the query set to obtain an embedded feature, and determine a corresponding relationship score based on the embedded feature and using a relationship network in the initial accident screening model;

[0124] a flattened feature determining unit, configured to convert the support set and the query set into one-dimensional vectors to obtain a first flattened feature and a second flattened feature corresponding to the support set and the query set, respectively;

[0125] a prediction result determining unit, configured to determine a synergy effect score based on the first flattened feature and the second flattened feature and using a comparison network in the initial accident screening model, and determine a prediction result using the relationship score and the synergy effect score;

[0126] a model parameter updating unit, configured to determine a loss value using the prediction result and the sample categories in the training set, and to update the model parameters in the initial accident screening model based on the loss value to obtain a trained accident screening model;

[0127] A model verification unit is used to verify the trained accident screening model based on the verification set to obtain a target accident screening model.

[0128] In some specific implementations, the target model determination module 12 may specifically include:

[0129] a performance evaluation unit, configured to verify the trained accident screening model based on the verification set to obtain a verified model, and to perform a performance evaluation on the verified model using a preset confusion matrix and various performance evaluation indicators to obtain a corresponding evaluation result;

[0130] A model parameter adjustment unit is used to adjust the model parameters of the verified model based on the evaluation result and using a grid search model to obtain a target accident screening model.

[0131] Furthermore, the embodiment of the present application also discloses an electronic device, Figure 11 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content in the diagram should not be considered as any limitation on the scope of use of this application. The electronic device 20 may specifically include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 is used to store a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the power system Nk fault transient stability screening method disclosed in any of the aforementioned embodiments. In addition, the electronic device 20 in this embodiment may specifically be an electronic computer.

[0132] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and the external device. The communication protocol it follows is any communication protocol that can be applied to the technical solution of this application and is not specifically limited here; the input and output interface 25 is used to obtain external input data or output data to the outside world. Its specific interface type can be selected according to specific application needs and is not specifically limited here.

[0133] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or CD, etc. The resources stored thereon can include an operating system 221, a computer program 222, etc., and the storage method can be temporary storage or permanent storage.

[0134] The operating system 221 is used to manage and control the hardware devices and computer program 222 on the electronic device 20, and can be Windows Server, Netware, Unix, Linux, etc. In addition to including a computer program capable of implementing the power system Nk fault transient stability screening method executed by the electronic device 20 disclosed in any of the aforementioned embodiments, the computer program 222 can further include computer programs capable of implementing other specific tasks.

[0135] Furthermore, this application discloses a computer-readable storage medium for storing a computer program. When executed by a processor, the computer program implements the aforementioned power system Nk fault transient stability screening method. The specific steps of this method can be found in the corresponding contents disclosed in the aforementioned embodiments and will not be further described here.

[0136] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from the other embodiments. Reference can be made to the descriptions of the identical or similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple, and the relevant parts can be referred to the descriptions of the methods.

[0137] Professionals may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the above description has generally described the components and steps of each example according to their functions. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0138] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein may be implemented directly using hardware, a software module executed by a processor, or a combination of the two. The software module may be placed in random access memory (RAM), internal memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.

[0139] Finally, it should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article, or device comprising the element.

[0140] The above is a detailed introduction to the technical solution provided by the present application. Specific examples are used herein to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core idea. At the same time, for those skilled in the art, according to the ideas of the present application, there may be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.

Claims

1. A method for screening transient stability of power system Nk fault, characterized in that: include: Using a preset simulation tool to perform transient time-domain simulation of the power system, and adjusting fault-related parameters of the power system during the simulation process to generate target long-term simulation data under different fault scenarios, and standardizing the target long-term simulation data to obtain a training set and a validation set; Reversing the processing order of layer normalization and residual connection in the deep learning model to obtain an improved learning model, constructing an initial accident screening model based on the improved learning model, the preset learning method, the relationship network, and the comparison network, and training and validating the initial accident screening model using the training set and the validation set, respectively, to obtain a target accident screening model; When a target extreme event occurs, a short-term simulation is performed based on the current operating state of the power system and using the preset simulation tool to generate input data, and the input data is input into the target accident screening model to obtain an output result of transient stability or transient instability.

2. The method for screening transient stability of power system Nk fault according to claim 1, characterized in that: The method of using a preset simulation tool to perform transient time-domain simulation on the power system and adjusting fault-related parameters of the power system during the simulation process to generate target long-term simulation data under different fault scenarios includes: Using preset simulation software and preset simulation duration, a three-phase short circuit fault is set at any preset location on the transmission line of the power system to obtain low-order fault data of a single component fault; Setting a three-phase short circuit fault at any preset location on the transmission line based on the preset simulation software and the preset simulation duration, and adjusting parameters of the number of line breaks, load level, fault location, and fault line of the power system during the simulation to obtain high-order fault data of simultaneous failure of multiple components; The low-order fault data and the high-order fault data are used to obtain target long-term simulation data under different fault scenarios.

3. The method for screening transient stability of power system Nk fault according to claim 2, characterized in that: The standardization of the target long-term simulation data to obtain a training set and a validation set includes: determining input samples including node voltage magnitude, phase angle, active power, reactive power, and electrical frequency based on the target long-term simulation data; Determining a maximum generator power angle difference corresponding to each of the input samples, and determining a sample category corresponding to the input sample based on the maximum generator power angle difference; A training sample is determined based on the input sample and the corresponding sample category, and the training sample is standardized to obtain a standardized sample. The standardized sample is used to determine a training set and a validation set.

4. The method for screening transient stability of power system Nk fault according to claim 3, characterized in that: The determining of the maximum generator power angle difference corresponding to each input sample, and determining the sample category corresponding to the input sample based on the maximum generator power angle difference, includes: determining a maximum generator power angle difference corresponding to each of the input samples, and determining whether the maximum generator power angle difference exceeds a target power angle difference threshold; If the maximum generator power angle difference exceeds a target power angle difference threshold, determining that the sample category of the input sample corresponding to the maximum generator power angle difference that exceeds the target power angle difference threshold is transient instability; If the maximum generator power angle difference does not exceed the target power angle difference threshold, it is determined that the sample category of the input sample corresponding to the maximum generator power angle difference that does not exceed the target power angle difference threshold is transient stability.

5. The method for screening transient stability of power system Nk fault according to claim 3, characterized in that: The initial accident screening model is constructed based on the improved learning model, the preset learning method, the relationship network and the comparison network, including: Determining the standardized samples corresponding to the low-order fault data as a support set for meta-learning, and determining the standardized samples corresponding to the high-order fault data as a query set for meta-learning; A meta-learning task is constructed based on the support set and the query set, and an initial accident screening model is constructed based on the meta-learning task, the improved learning model, the relational network and the comparison network.

6. The method for screening transient stability of power system Nk fault according to claim 5, characterized in that: The initial accident screening model is trained and verified using the training set and the verification set respectively to obtain a target accident screening model, including: Combining the support set and the query set to obtain an embedded feature, and determining a corresponding relationship score based on the embedded feature and using a relationship network in the initial accident screening model; Converting the support set and the query set into one-dimensional vectors to obtain first flattened features and second flattened features corresponding to the support set and the query set, respectively; determining a synergy score based on the first flattened feature and the second flattened feature and using a comparison network in the initial accident screening model, and determining a prediction result using the relationship score and the synergy score; Determining a loss value using the prediction result and the sample categories in the training set, and updating the model parameters in the initial accident screening model based on the loss value to obtain a trained accident screening model; The trained accident screening model is verified based on the verification set to obtain a target accident screening model.

7. The method for screening transient stability of power system Nk fault according to claim 6, characterized in that: The verifying the trained accident screening model based on the verification set to obtain a target accident screening model includes: Verifying the trained accident screening model based on the validation set to obtain a verified model, and performing a performance evaluation on the verified model using a preset confusion matrix and various performance evaluation indicators to obtain corresponding evaluation results; Based on the evaluation results and using a grid search model, the model parameters of the verified model are adjusted to obtain a target accident screening model.

8. A power system Nk fault transient stability screening device, characterized in that: include: A simulation module is used to perform transient time-domain simulation of the power system using a preset simulation tool, and to adjust fault-related parameters of the power system during the simulation process to generate target long-term simulation data under different fault scenarios, and to standardize the target long-term simulation data to obtain a training set and a validation set; a target model determination module, configured to swap the processing order of layer normalization and residual connection in the deep learning model to obtain an improved learning model, construct an initial accident screening model based on the improved learning model, a preset learning method, a relationship network, and a comparison network, and train and verify the initial accident screening model using the training set and the validation set, respectively, to obtain a target accident screening model; The output result determination module is used to generate input data by performing short-term simulation based on the current operating state of the power system and using the preset simulation tool when a target extreme event occurs, and input the input data into the target accident screening model to obtain an output result of transient stability or transient instability.

9. An electronic device, characterized in that: include: Memory, used to store computer programs; A processor is configured to execute the computer program to implement the power system Nk fault transient stability screening method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that Used to store a computer program, wherein when the computer program is executed by a processor, the power system Nk fault transient stability screening method according to any one of claims 1 to 7 is implemented.

Images