Power system transient stability assessment method based on space-time diagram convolutional network

Abstract

The invention provides an electric power system transient stability evaluation method based on a space-time diagram convolutional network, and belongs to the technical field of electric power system transient stability evaluation. In order to solve the technical problem of low efficiency and performance of an evaluation model caused by high-dimensional redundancy of transient stability characteristics of an existing power system, the technical scheme adopted by the invention is as follows: generating an operation scene as a simulation sample by utilizing a time domain simulation function of power system simulation software, and incorporating the simulation sample into a sample data set; processing the original data of the simulation sample by adopting a two-stage screening hybrid intelligent feature selection method; a subset of final data features of the power system is converted into graph structure data, training is performed based on a learning framework of a space-time graph convolution network, topological features of the graph structure data of the power system are captured by using a space-time graph convolution encoder, and transient stability of the power system is evaluated based on a multi-outlet classification mechanism. Establishing a transient stability evaluation model; the method is applied to transient stability evaluation of the power system.

Description

Technical Field

[0001] The present invention provides a power system transient stability assessment method based on a spatiotemporal graph convolutional network, belonging to the technical field of power system transient stability assessment. Background Art

[0002] With the rapid development of power systems and the continuous access of a large number of renewable energy sources, power systems are facing increasingly surging risks and challenges. The power systems are affected by various faults, which will lead to an increasing possibility of transient instability.

[0003] In the context of the development of new power systems, using effective feature selection methods to extract key response features that are strongly correlated with transient stability is of great significance for studying the transient instability mechanism and potential safety hazards of the system. However, the transient stability characteristics of power systems with a high proportion of renewable energy and DC access are currently high-dimensional and redundant, which will affect the efficiency and performance of data-driven evaluation models. Therefore, a new feature selection method is needed to solve the above problems. Summary of the Invention

[0004] In order to solve the technical problems existing in the background technology, the present invention adopts a technical solution: a power system transient stability assessment method based on spatiotemporal graph convolutional network, including the following assessment steps:

[0005] Step 1: Use the time-domain simulation function of the power system simulation software to generate operation scenarios as simulation samples and include them in the sample data set;

[0006] Step 2: A two-stage hybrid intelligent feature selection method is used to process the original data of the simulation samples to achieve feature dimensionality reduction. In the first stage, the random forest algorithm is used to preliminarily screen the data feature dimensions based on the importance of the features. In the second stage, the improved large cane rat algorithm is used to optimize the subset of data features.

[0007] Step 3: Convert a subset of the final power system data features into graph-structured data. Train the data using a spatiotemporal graph convolutional network (STGCN) learning framework. Use a STGCN encoder to capture the topological features of the power system graph data. Evaluate the transient stability of the power system based on a multi-export classification mechanism, and establish a transient stability assessment model.

[0008] Step 4: Use the phasor measurement unit in the wide-area measurement system to collect the required real-time measurement data, input the actual measurement data into the trained transient stability assessment model, and finally output the transient stability assessment results of the power system.

[0009] The specific method of step 1 is:

[0010] A power system model was built in the power system simulation software, and batch time-domain simulations were performed with different fault times, fault locations, and load levels. A data set was obtained through the simulations, and faults were randomly introduced on the transmission lines with a sampling interval of 10 milliseconds and a total sampling duration of 10 seconds.

[0011] Obtain power system operating variables. Sampled data includes the voltage amplitude and voltage phase angle of the generator bus, and the active power and reactive power of the generator bus.

[0012] Select a continuous sequence of 20 time steps in the sampled data as a feature , by all sample features Composition of the dataset ;

[0013] The transient stability state of the power system is defined as , 0 or 1, indicating that the system is transiently stable or transiently unstable, and the state of all samples Composing a label set ;

[0014] The sample feature With transient stability label Correspondingly, construct a sample set ;

[0015] The sample labels are determined based on the transient stability index, and an original data set containing a large amount of system operation status information is generated. The maximum power angle difference between any two generator sets after the system failure is selected as the basis for transient stability judgment, and the transient stability index of the power system is calculated. The calculation formula is:

[0016] ;

[0017] in, is the maximum power angle difference between the two generators at any time;

[0018] when When , the system determines that it is a transient stable sample and determines the sample label to be 1; otherwise, the system determines that it is a transient unstable sample and determines the sample label to be 0.

[0019] The specific method of step 2 is:

[0020] Step 2-1: Define the sample data set in step 1 in the first stage The eigenvectors are normalized and the data range is transformed to , get the normalized sample set :

[0021] ;

[0022] Where, It is The first sample Features and is the maximum and minimum characteristic value of the sample; is the characteristic mean of the sample;

[0023] In the first stage, the random forest algorithm is used to integrate multiple decision trees and use the contribution of the tree model to the feature when the node is split to evaluate the importance of the feature. The calculation formula for evaluating the importance of the feature is:

[0024] ;

[0025] Where, Indicates the Importance score of each feature; represents the number of decision trees in the forest; Represents a decision tree; Indicates the The classification accuracy of the decision tree under the original feature order; Indicates the Classification accuracy after randomly shuffling the order of features;

[0026] Fix the order of other features and only randomly rearrange the first The sample value of each feature is used to destroy the true association between the feature and the label. If the accuracy drops significantly after the disruption, it means that the feature is important for model prediction. All features are sorted according to the feature importance score, and the features ranked in the top 80% of importance are selected to complete the initial screening;

[0027] Step 2-2: Initialize the population in the second stage:

[0028] Initialize random generation search agents, each representing a possible feature solution, with a dimension equal to the number of features in the original dataset , the vector expression of the agent is:

[0029] ;

[0030] in, Indicates selection Features, Indicates no choice;

[0031] During each iteration, the position vector returned by each search agent The corresponding feature subset can be obtained Each feature subset will be used as the input of the learning model, and the K-nearest neighbor algorithm will be used as the subset classifier in the feature optimization process. Combined with the transient stability multidimensional time series data, the feature subset is obtained. The evaluation results guide the next iteration through the function score;

[0032] Step 2-3: Calculate the fitness of each individual and select the individual with the best fitness as the dominant individual , update it to the global optimal solution ,at this time ;

[0033] Using classification accuracy, Value calculation fitness index:

[0034] Use the current feature subset on the training set for training and prediction, and calculate the corresponding fitness value. The calculation formula is:

[0035] ;

[0036] Where, and is the performance indicator of transient stability assessment when using the selected feature subset, 、 and They are accuracy, and the weight parameter corresponding to the importance of the number of selected feature subsets, is the feature dimension corresponding to feature subset selection, It is the dimension after the initial screening;

[0037] Step 2-4: According to the random threshold With parameters Compare and update the feature subset:

[0038] The new position of the remaining population in the search space is determined based on the position of the dominant individual. If the fitness function value of another individual exceeds that of the optimal individual, the optimal individual is updated, and the positions of other individuals are adjusted based on the newly established optimal individual. The corresponding update calculation formula is:

[0039] ;

[0040] ;

[0041] ;

[0042] Where, is the current iteration number; is the maximum number of iterations; is the fitness function value of the dominant individual; is the disturbance factor, simulating the effect of abundant food sources; is the coefficient simulating the reduction of food sources; is the coefficient that encourages the population to move to other abundant food sources available in the breeding area; , is a random integer between 1 and 4;

[0043] When rand < ρ, a global search algorithm is used to update the position based on the current fitness. The calculation formula is:

[0044] ;

[0045] Where, Is a random value between 0 and 1; Is a dominant individual of Dimensional position;

[0046] Define the initial perturbation probability as high, and traverse the remaining dimensional features of the dominant individuals whose scores increase after removing a certain dimensional feature;

[0047] When rand ≥ ρ, select non-dominant individuals near the dominant individual for local search and fine-tune the position near the known good feature combination. The calculation formula is:

[0048] ;

[0049] Where, is a randomly selected non-dominant individual of Dimensional position; , make small adjustments;

[0050] Traverse the adjacent dimension features whose scores decrease after removing a certain dimension feature from the dominant individuals;

[0051] Step 2-5: Calculate the fitness of the feature subset after the updated position, adopt the update strategy of the exploration or development phase according to the fitness value, update the individual position again, and update the global optimal solution ;

[0052] Step 2-6: Repeat steps 2-4 to 2-5 until the maximum number of iterations is reached , determine the total number of iterations of the proposed algorithm times, execute the loop from to , starts the iterative process of transient stability key feature optimization algorithm, and selects new dominant individuals in the iterative process , and finally return to the global optimal solution , that is, the feature subset that removes redundant features.

[0053] The specific method of step 3 is:

[0054] The spatiotemporal graph convolutional network uses graph convolutional networks, gated linear units, and gated convolutional neural networks to capture the structural characteristics of the power grid space and the temporal variation characteristics of the transient process data volume, respectively, to extract transient stability features;

[0055] Introducing multiple classification exit nodes at different depths in the network, adding a multi-branch early exit classification module, evaluating the prediction confidence of each exit node, and stopping the reasoning process if the confidence exceeds a predefined threshold, and outputting the current prediction;

[0056] According to the transient stability feature set obtained in step 2 after the two-stage hybrid intelligent feature screening process, the transient stability feature set is converted into graph structure data, and a transient stability assessment model based on the spatiotemporal graph convolutional network is trained.

[0057] The beneficial effects of the present invention compared with the prior art are as follows: the present invention provides a method for transient stability assessment of power systems based on spatiotemporal graph convolutional networks, which mainly utilizes the time domain simulation function of the power system simulation software to generate simulation results of operating scenarios as samples, and realizes feature dimensionality reduction through a two-stage screening hybrid intelligent feature selection method, thereby improving the efficiency and accuracy of model stability assessment, and converts sample data into graph structure data. The spatiotemporal graph convolutional encoder is used to capture system topology features, constructs a multi-dimensional feature-driven integrated model, and uses the trained model for transient stability assessment of power systems; the intelligent feature selection scheme based on two-stage hybrid provided by the present invention can further realize feature subset optimization on the basis of the system's initial screening features, construct a data set with a moderate and rich number of sample features, and improve the efficiency and performance of high-dimensional time series feature selection. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] The present invention will be further described below with reference to the accompanying drawings:

[0059] Figure 1 This is a flowchart of the steps of the power system transient stability assessment method based on feature selection and spatiotemporal graph convolution of the present invention;

[0060] Figure 2 This is a flowchart of the steps of the GCRA optimization algorithm involved in an embodiment of the present invention;

[0061] Figure 3 This is a structural diagram of an improved spatiotemporal graph convolutional network model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0062] like Figures 1 to 3As shown, the power system transient stability assessment method based on the spatiotemporal graph convolutional network provided by the present invention mainly uses the time domain simulation function of the power system simulation software to generate simulation results of the operation scenario as samples, realizes feature dimensionality reduction through a two-stage screening hybrid intelligent feature selection method, uses the random forest algorithm to preliminarily screen the feature dimensions according to the importance of the features, and uses the large cane rat optimization algorithm to optimize the feature subset based on the preliminary screening, selects the optimal feature set from it, and reduces the original feature dimension; then, the power system sample data is further converted into graph structure data, and a learning framework based on the spatiotemporal graph convolutional network is used, and a spatiotemporal graph convolutional encoder is adopted to capture the system topology features; a multi-export mechanism is used to achieve efficient and reliable data classification; finally, after constructing a multi-dimensional feature-driven integrated model based on offline training, the trained model is used to complete the online transient stability assessment.

[0063] To achieve the above-mentioned power system transient stability assessment goal, the present invention provides a power system transient stability assessment method based on a spatiotemporal graph convolutional network, which specifically includes the following steps:

[0064] Step 1: Use the time domain simulation function of the power system simulation software to generate simulation results of the operation scenario as samples and include them in the sample data set;

[0065] Specifically, a power system model is built in the simulation software, and different fault times, fault locations and load levels are set for batch time domain simulation to obtain power system operation variables, including node voltage amplitude and phase angle, generator active power and reactive power. The sampled data is used as the feature quantity , all sample features Composition of the dataset The transient stability state of the power system is defined as , 0 or 1, indicating that the system is transiently stable or transiently unstable. All sample states Composing a label set .sample With transient stability label Correspondingly, construct a sample set .

[0066] According to the transient stability index, the sample labels are determined and the original data set containing a large amount of system operation status information is generated. The transient stability evaluation index of the power system is constructed as follows:

[0067] ;

[0068] in, is the maximum power angle difference between the two generators at any time; when When , the system is judged to be a transient stable sample; otherwise, the system is transiently unstable.

[0069] Step 2: A two-stage hybrid intelligent feature screening method is used to reduce the dimensionality of the simulated data. The random forest algorithm is used to preliminarily screen the feature dimensions based on their importance. Based on this initial screening, the improved Greater Cane Rat Algorithm (GCRA) is used to optimize the feature subsets. The appropriate set of features is selected from the measured data, reducing the original feature dimensionality and improving the efficiency and accuracy of the model stability assessment.

[0070] Specifically, to address the current problems with selecting key response features during transient stability of power systems, the continued growth of wind and solar power generation and the high proportion of power electronic equipment deployed in new power grid operating characteristics will lead to a gradual increase in the number of features reflecting system stability. To obtain key response features that are strongly correlated with stability and applicable to new power systems from ultra-high-dimensional features, this paper proposes a two-stage hybrid intelligent feature screening method to select an appropriate set of feature quantities from measurement data, reduce the original feature dimensionality, and establish a dataset with rich samples and fast computation.

[0071] The specific principles of the two-stage hybrid intelligent feature screening method are as follows:

[0072] Step 2-1: For the sample set in step 1 The eigenvectors are normalized and the data range is transformed to , get the normalized sample set :

[0073] ;

[0074] Where, It is The first sample Features and is the maximum and minimum characteristic value of the sample; is the characteristic mean of the sample;

[0075] In the first stage, random forest is used to integrate multiple decision trees and use the tree model's contribution to the feature when the node is split to evaluate the importance. The calculation formula of feature importance is expressed as follows:

[0076] ;

[0077] Where, Indicates the Importance score of each feature; represents the number of decision trees in the forest; Represents a decision tree; Indicates the The classification accuracy of the decision tree under the original feature order; Indicates the Classification accuracy after randomly shuffling the order of features;

[0078] Fix the order of other features and only randomly rearrange the first The sample value of each feature is disrupted to destroy the true association between the feature and the label; if the accuracy drops significantly after the disruption, it means that the feature is crucial to the model prediction. All features are sorted according to the feature importance score, and the features ranked in the top 80% of importance are selected to complete the initial screening.

[0079] Step 2-2: In the second stage, swarm intelligence optimization algorithm is used to optimize feature subsets;

[0080] The GCRA algorithm is a meta-heuristic algorithm obtained by mathematically modeling the foraging and mating behavior of large cane rats in nature. It uses an improved optimization algorithm to optimize the feature subset of transient stable samples. First, the population is initialized: the initialization randomly generates search agents, each representing a possible feature solution, with a dimension equal to the number of features in the original dataset ; The agent's vector representation is:

[0081] ;

[0082] in Indicates selection Features, Indicates no choice;

[0083] During each iteration, the position vector returned by each search agent The corresponding feature subset can be obtained Each feature subset will be used as the input of the learning model, and the K-nearest neighbor algorithm (kNN) will be used as the subset classifier in the feature optimization process. Combined with the transient stability multidimensional time series data, the feature subset is obtained. The evaluation results are used to guide the next iteration through the function score.

[0084] Step 2-3: Calculate the fitness of each individual and select the dominant individual. Define the fitness function to evaluate the pros and cons of each feature subset. The individual with the best fitness is selected as the dominant individual. , update it to the global optimal solution ,at this time ;

[0085] Using classification accuracy, The fitness index is calculated by using the current feature subset on the training set for training and prediction, and the corresponding fitness value is calculated. The fitness function is defined as follows:

[0086] ;

[0087] Where, and is the performance indicator of transient stability assessment when using the selected feature subset, 、 and They are accuracy, and the weight parameter corresponding to the importance of the number of selected feature subsets, is the feature dimension corresponding to feature subset selection, It is the dimension after initial screening.

[0088] Step 2-4: Update the feature subset according to the random threshold With parameters The algorithm enters the exploration or development stage;

[0089] Individuals search for food in their territory. The dominant individual retains the trajectory information, and the remaining individuals adjust their positions according to the dominant individual. The new position of the remaining population in the search space is determined based on the position of the dominant individual. If the fitness function value of another individual exceeds the value of the optimal individual, the optimal individual is updated, and the positions of the other individuals are adjusted based on the newly established optimal individual. The corresponding update calculation formula includes:

[0090] ;

[0091] ;

[0092] ;

[0093] Where, is the current iteration number; is the maximum number of iterations; is the fitness function value of the dominant individual; is the disturbance factor, simulating the effect of abundant food sources; is the coefficient simulating the reduction of food sources; is the coefficient that encourages the population to move to other abundant food sources available in the breeding area; , is a random integer between 1 and 4.

[0094] Exploration phase (rand < ρ): In the exploration phase, the algorithm enters a global search and changes positions based on the current fitness:

[0095] ;

[0096] Where, Is a random value between 0 and 1; Is a dominant individual of Dimension position, defining the initial perturbation probability is high, and traversing the remaining dimensional features whose scores increase after removing a certain dimensional feature from the dominant individuals.

[0097] Development phase (rand ≥ ρ): Inheriting the dominant features and introducing small-scale perturbations to perform local search, the algorithm begins to fine-tune around known good feature combinations:

[0098] ;

[0099] Where, is a randomly selected non-dominant individual of Dimensional position; , make small-scale adjustments, and traverse the adjacent dimension features whose scores are reduced after removing a certain dimension feature from the dominant individuals.

[0100] Step 2-5: Evaluation and Update: Calculate the fitness of the feature subset after the updated position, adopt the update strategy of the exploration or development phase according to the fitness value, update the individual position again, and update the global optimal solution .

[0101] Step 2-6: Iteration and termination: Repeat steps 2-4 to 2-5 until the maximum number of iterations is reached , determine the total number of iterations of the proposed algorithm times, execute the loop from to , the iterative process of transient stability key feature optimization algorithm begins. In the process, new dominant individuals are selected , and finally return to the global optimal solution , that is, the feature subset that removes redundant features.

[0102] Step 3: Convert the final subset of the power system data into graph-structured data and train it using a spatiotemporal graph neural network (SGNN) learning framework. A spatiotemporal graph convolutional encoder is used to capture system topology features. A multi-outlet classification mechanism is then applied to transient stability assessment of the power system, enabling dynamic decision-making during the inference process.

[0103] In step 3, the spatiotemporal graph convolution operation is applied through the spatiotemporal graph convolution network to extract temporal and spatial features from the input graph data. The classification module also uses spatiotemporal graph convolution for feature extraction, and then combines it with a multi-export mechanism, which enables the model to produce classification results from multiple intermediate layers and terminate the calculation early when the classification confidence exceeds the specified threshold, thereby improving reasoning efficiency.

[0104] Specifically, based on the processed data set, a learning framework based on the spatiotemporal graph convolutional network is trained to extract the relationship between the key features of the power system and the corresponding stability labels; a multi-export classification mechanism is used to introduce multiple exit nodes at different depths in the network, and the model evaluates the prediction confidence of each exit point. If the confidence exceeds the predefined threshold, the reasoning process stops and the current prediction is output; the transient stability assessment model will be used for online transient stability assessment, and the improved spatiotemporal graph convolutional network model structure involved is as follows: Figure 3 shown.

[0105] The spatial structure of the power grid represents spatial characteristics, while the temporal changes in the amount of data during transient processes represent temporal characteristics. A spatiotemporal graph convolutional network (GCN), a gated linear unit (GLU), and a gated convolutional neural network (GatedCNN) are used to capture these spatial and temporal characteristics, respectively, enabling accurate extraction of transient stability features.

[0106] Building on the spatiotemporal graph convolutional network, a multi-branch early exit classification module is added. This module introduces multiple classification exit nodes at different depths in the network, enabling dynamic decision-making during inference. Specifically, the model evaluates the prediction confidence of each exit point. If the confidence exceeds a predefined threshold, the inference process stops and the current prediction is output.

[0107] According to the transient stability feature set obtained in step 2 after two-stage hybrid intelligent feature screening, it is converted into graph structure data for training the transient stability assessment model based on spatiotemporal graph convolutional network.

[0108] Step 4: Finally, the required measurement data is obtained online using the Power Management Unit (PMU), input into the trained transient stability assessment model, and the transient stability assessment results are output;

[0109] Specifically, the present invention is based on an improved spatiotemporal graph convolutional network. When applied online, the phasor measurement unit (PMU) in the wide-area measurement system is used to collect the required real-time measurement data, and the actual measurement data is input into the trained transient stability assessment model. The model will simultaneously output the assessment results, thereby obtaining the transient stability assessment results of the power system and completing the online transient stability assessment.

[0110] In an embodiment of the present invention, the superiority of the present invention is verified by testing the power system described in step 1, such as Figure 1 As shown in the figure, the fitness effect of the optimization process of the method of the present invention is compared with that of other common optimization algorithms. The proposed hybrid intelligent feature selection method based on double screening can effectively achieve feature dimensionality reduction; the initial screening stage provides feature validity measurement values ​​and preliminary dimensionality reduction, which improves the efficiency of subsequent subset optimization. The intelligent optimization algorithm of the cane rat can fully consider the time-varying characteristics of the features and has higher classification accuracy; the evaluation indicators are defined as missing alarm rate (MAR), false alarm rate (FAR) and accuracy rate (ACC), among which To evaluate the correct stable sample, To estimate the stable sample of errors, To evaluate the unstable samples of errors, To evaluate the correct unstable sample, the specific calculation formula is:

[0111] ;

[0112] ;

[0113] ;

[0114] The above method is compared with other optimization algorithms on three price accuracy indicators, as shown in Table 1 below:

[0115]

[0116] Table 1 Comparison of evaluation indicators of various optimization methods

[0117] It can be seen from the experimental data in Table 1 that the method of the present invention can better balance the number of feature selections and the accuracy of classification in the selection of key features for transient stability of power systems. Therefore, the key features of transient stability selected by the method have better advantages in application and transient stability assessment.

[0118] The evaluation indicators are selected to test the classification performance of the transient stability assessment model based on SGNN. The specific calculation formula is:

[0119] ;

[0120] ;

[0121] ;

[0122] Where: Precision The ratio of correctly evaluated stable samples to all evaluated stable samples; recall The ratio of correctly evaluated stable samples to all actual stable samples; Defined as precision and recall The harmonic mean of The value range of is (0,1), The closer it is to 1, the better the classification performance of the TSA model.

[0123] In order to demonstrate the significant classification performance of the transient stability assessment model based on SGNN in this invention, it was tested and compared with other deep learning models. Three other methods, namely, fully convolutional network (FCN), gated recurrent unit (GRU), and long short-term memory network (LSTM), were tested using the same input samples. and The comparison results in terms of values ​​are shown in Table 2 below.

[0124]

[0125] Table 2 Experimental results of transient stability model

[0126] As can be seen from Table 2, the TSA model based on SGNN has more significant classification performance than other models. 、 The respective rates are as high as 98.78% and 98.69%, which shows that the classification performance meets the actual needs and is in line with the purpose of the present invention.

[0127] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for transient stability assessment of power systems based on spatiotemporal graph convolutional networks, characterized by: The assessment steps include: Step 1: Use the time-domain simulation function of the power system simulation software to generate operation scenarios as simulation samples and include them in the sample data set; Step 2: A two-stage hybrid intelligent feature selection method is used to process the original data of the simulation samples to achieve feature dimensionality reduction. In the first stage, the random forest algorithm is used to preliminarily screen the data feature dimensions based on the importance of the features. In the second stage, the improved large cane rat algorithm is used to optimize the subset of data features. Step 3: Convert a subset of the final power system data features into graph-structured data. Train the data using a spatiotemporal graph convolutional network (STGCN) learning framework. Use a STGCN encoder to capture the topological features of the power system graph data. Evaluate the transient stability of the power system based on a multi-export classification mechanism, and establish a transient stability assessment model. Step 4: Use the phasor measurement unit in the wide-area measurement system to collect the required real-time measurement data, input the actual measurement data into the trained transient stability assessment model, and finally output the transient stability assessment results of the power system.

2. The method for evaluating power system transient stability based on spatiotemporal graph convolutional networks according to claim 1, characterized in that: The specific method of step 1 is: A power system model was built in the power system simulation software, and batch time-domain simulations were performed with different fault times, fault locations, and load levels. A data set was obtained through the simulations, and faults were randomly introduced on the transmission lines with a sampling interval of 10 milliseconds and a total sampling duration of 10 seconds. Obtain power system operating variables. Sampled data includes the voltage amplitude and voltage phase angle of the generator bus, and the active power and reactive power of the generator bus. Select a continuous sequence of 20 time steps in the sampled data as a feature , by all sample features Composition of the dataset ; The transient stability state of the power system is defined as , 0 or 1, indicating that the system is transiently stable or transiently unstable, and the state of all samples Composing a label set ; The sample feature With transient stability label Correspondingly, construct the sample set ; The sample labels are determined based on the transient stability index, and an original data set containing a large amount of system operation status information is generated. The maximum power angle difference between any two generator sets after the system failure is selected as the basis for transient stability judgment, and the transient stability index of the power system is calculated. The calculation formula is: ； in, is the maximum power angle difference between the two generators at any time; when When , the system determines that it is a transient stable sample and determines the sample label to be 1; otherwise, the system determines that it is a transient unstable sample and determines the sample label to be 0.

3. The method for evaluating power system transient stability based on spatiotemporal graph convolutional networks according to claim 1, characterized in that: The specific method of step 2 is: Step 2-1: Define the sample data set in step 1 in the first stage The eigenvectors are normalized and the data range is transformed to , get the normalized sample set : ； Where, It is The first sample Features and is the maximum and minimum characteristic value of the sample; is the characteristic mean of the sample; In the first stage, the random forest algorithm is used to integrate multiple decision trees and use the contribution of the tree model to the feature when the node is split to evaluate the importance of the feature. The calculation formula for evaluating the importance of the feature is: ； Where, Indicates the Importance score of each feature; represents the number of decision trees in the forest; Represents a decision tree; Indicates the The classification accuracy of the decision tree under the original feature order; Indicates the Classification accuracy after randomly shuffling the order of features; Fix the order of other features and only randomly rearrange the first The sample value of each feature is used to destroy the true association between the feature and the label. If the accuracy drops significantly after the disruption, it means that the feature is important for model prediction. All features are sorted according to the feature importance score, and the features ranked in the top 80% of importance are selected to complete the initial screening; Step 2-2: Initialize the population in the second stage: Initialize random generation search agents, each representing a possible feature solution, with a dimension equal to the number of features in the original dataset , the vector expression of the agent is: ； in, Indicates selection Features, Indicates no choice; During each iteration, the position vector returned by each search agent The corresponding feature subset can be obtained Each feature subset will be used as the input of the learning model, and the K-nearest neighbor algorithm will be used as the subset classifier in the feature optimization process. Combined with the transient stability multidimensional time series data, the feature subset is obtained. The evaluation results guide the next iteration through the function score; Step 2-3: Calculate the fitness of each individual and select the individual with the best fitness as the dominant individual , update it to the global optimal solution ,at this time ; Using classification accuracy, Value calculation fitness index: Use the current feature subset on the training set for training and prediction, and calculate the corresponding fitness value. The calculation formula is: ； Where, and is the performance indicator of transient stability assessment when using the selected feature subset, 、 and They are accuracy, and the weight parameter corresponding to the importance of the number of selected feature subsets, is the feature dimension corresponding to feature subset selection, It is the dimension after the initial screening; Step 2-4: According to the random threshold With parameters Compare and update the feature subset: The new position of the remaining population in the search space is determined based on the position of the dominant individual. If the fitness function value of another individual exceeds that of the optimal individual, the optimal individual is updated, and the positions of other individuals are adjusted based on the newly established optimal individual. The corresponding update calculation formula is: ； ； ； Where, is the current iteration number; is the maximum number of iterations; is the fitness function value of the dominant individual; is the disturbance factor, simulating the effect of abundant food sources; is the coefficient simulating the reduction of food sources; is the coefficient that encourages the population to move to other abundant food sources available in the breeding area; , is a random integer between 1 and 4; When rand < ρ, a global search algorithm is used to update the position based on the current fitness. The calculation formula is: ； Where, Is a random value between 0 and 1; Is a dominant individual of Dimensional position; Define the initial perturbation probability as high, and traverse the remaining dimensional features of the dominant individuals whose scores increase after removing a certain dimensional feature; When rand ≥ ρ, select non-dominant individuals near the dominant individual for local search and fine-tune the position near the known good feature combination. The calculation formula is: ； Where, is a randomly selected non-dominant individual of Dimensional position; , make small adjustments; Traverse the adjacent dimension features whose scores decrease after removing a certain dimension feature from the dominant individuals; Step 2-5: Calculate the fitness of the feature subset after the updated position, adopt the update strategy of the exploration or development phase according to the fitness value, update the individual position again, and update the global optimal solution ; Step 2-6: Repeat steps 2-4 to 2-5 until the maximum number of iterations is reached , determine the total number of iterations of the proposed algorithm times, execute the loop from to , start the iterative process of transient stability key feature optimization algorithm, select new dominant individuals in the iterative process , and finally return to the global optimal solution , that is, the feature subset that removes redundant features.

4. The method for evaluating power system transient stability based on spatiotemporal graph convolutional networks according to claim 1, characterized in that: The specific method of step 3 is: The spatiotemporal graph convolutional network uses graph convolutional networks, gated linear units, and gated convolutional neural networks to capture the structural characteristics of the power grid space and the temporal variation characteristics of the transient process data volume, respectively, to extract transient stability features; Introducing multiple classification exit nodes at different depths in the network, adding a multi-branch early exit classification module, evaluating the prediction confidence of each exit node, and stopping the reasoning process if the confidence exceeds a predefined threshold, and outputting the current prediction; According to the transient stability feature set obtained in step 2 after the two-stage hybrid intelligent feature screening process, the transient stability feature set is converted into graph structure data, and a transient stability assessment model based on the spatiotemporal graph convolutional network is trained.

Images