Artificial intelligence-based method for assessing power system transient stability considering sample imbalance

Abstract

The application discloses an artificial intelligence-based power system transient stability evaluation method considering sample imbalance, and comprises the following steps: step 1, generating an original data set containing system operation information based on a time domain simulation technology; step 2, proposing a synthetic minority over-sampling technique (SMOTE) method with adaptive eligible synthesizer selection to process the original data set, and establishing a sample balanced and rich data set; step 3, constructing a feature extraction model based on a parallel convolution algorithm; step 4, training and establishing a transient stability evaluation model based on a rich random forest based on an offline data set; step 5, quickly establishing an updated model; and step 6, completing online transient stability evaluation based on data collected by a phasor measurement unit. SM The application has the advantages that an ASN-SMOTE method is proposed, the sample imbalance problem is effectively solved, a feature extraction framework is constructed based on a parallel convolution algorithm, highly relevant key operation variables T SM are extracted, and the calculation efficiency of subsequent model evaluation is greatly improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of power system transient stability assessment, and particularly relates to a power system transient stability assessment method based on artificial intelligence and considering sample imbalance. BACKGROUND

[0002] Transient stability assessment (TSA) is of great significance in power system planning and operation. With the construction and development of smart grids, the traditional power system is undergoing major changes. Nowadays, artificial intelligence is increasingly widely used in power systems, and the advent of artificial intelligence makes related calculations of power systems more rapid and accurate. Therefore, how to combine appropriate artificial intelligence algorithms, timely grasp the system operating state, reduce economic losses caused by unforeseen power accidents, and build an online, rapid and accurate TSA method have become an urgent need of modern power systems.

[0003] The patent document with the authorized announcement number CN113937762A discloses a micro-grid transient stability assessment method based on a long short-term memory network, which adds a transient stability assessment network and network weight to the control unit of a virtual synchronous inverter for information interaction. When the micro-grid is normally operated, the transient stability assessment network does not perform assessment, and when the micro-grid detects that a fault occurs, the transient stability assessment network performs online assessment. The patent document with the authorized announcement number CN107171315B discloses a power system transient stability assessment method based on RPTSVM, which performs feature compression on power system indicators and projected energy function indicators through a maximum correlation minimum redundancy feature selection method, finds a power system feature subset with high sensitivity to transient changes of the power grid, and then performs power system transient stability assessment after the RPTSVM adds a regular term to the optimization objective function to reconstruct the classifier. However, the above methods do not fully consider the problem that when applied to large-scale power systems, they may consume a large amount of calculation time due to the difficulty and complexity of parameter tuning, and also do not consider the sample imbalance problem caused by the fact that with the continuous improvement of the stability of modern power systems, more stable samples are used for TSA research than unstable samples.

[0004] In summary, the current transient stability assessment method is difficult to meet the demand of modern power systems for high calculation efficiency and high precision of the transient stability assessment method, and lacks certain research on the sample imbalance problem. SUMMARY

[0005] The application is to solve the problem that the sample imbalance problem brings unpredictable influence to the data-driven transient stability evaluation method, so that the stable operation of the system faces great risks, and at the same time meets the demand of high calculation efficiency and high precision of the transient stability evaluation method of modern power system, and proposes an artificial intelligence-based power system transient stability evaluation method considering sample imbalance.

[0006] In order to achieve the above-mentioned application purposes, the application adopts the following technical methods:

[0007] The artificial intelligence-based power system transient stability evaluation method considering sample imbalance comprises the following steps:

[0008] Step 1: based on time domain simulation technology, obtain power system operation variables, construct transient stability margin index T SM , generate original data set containing a large amount of system operation state information;

[0009] Step 2: propose a synthetic minority oversampling method with adaptive qualified synthesizer selection (ASN-SMOTE) to process the original data set, and through adaptive synthesis of unstable samples, establish a sample balanced and rich data set;

[0010] Step 3: based on parallel convolution algorithm (PCA), construct feature extraction model, extract key operation variables highly related to T SM , and establish offline data set according to the extracted key operation variables and corresponding class labels;

[0011] Step 4: based on the offline data set, train the transient stability evaluation model based on the enriched random forest (ERF), extract the mapping relationship between the key operation variables of the power system and the corresponding class labels, and establish the transient stability evaluation model based on the ERF;

[0012] Step 5: considering that the system may appear new topology structure, timely check the evaluation accuracy of the ERF transient stability evaluation old model, and quickly establish the updated model;

[0013] Step 6: based on the phasor measurement unit (PMU), collect the required real-time measurement data, and use the trained ERF model to evaluate the transient stability state, and complete the online transient stability evaluation.

[0014] In step 1, an actual power system model is built using simulation technology, and different fault times, fault locations and load levels are set for batch time domain simulation to obtain power system operation variables, build transient stability margin indicators, and generate an original data set containing a large amount of system operation state information.

[0015] The transient stability margin can represent the transient stability level of the power system and is usually used to describe the operation state of the system. The critical removal time and the actual removal time of the fault are used to build the transient stability margin indicator T SM , and the specific calculation is as follows:

[0016]

[0017] In the formula: t CC is the maximum removal time of the system under stable operation under the fault; t AC is the actual removal time of the system under the fault; T SM is a continuous value in the range of-1 to 1. According to the transient stability evaluation standard of the actual power system, a threshold α can be set, and T SM The operation state under the current system fault is classified as stable state and unstable state, and is labeled, that is:

[0018]

[0019] In step 2, considering that the unbalanced sample problem of the modern power system may bring unpredictable influence to the data-driven transient stability evaluation method, and the stable operation of the system faces great risks. An ASN-SMOTE method is proposed to process the original data set, and through adaptive synthesis of unstable samples, the number of unstable samples and stable samples is basically balanced, and a sample balanced and rich data set is established.

[0020] The principle of the ASN-SMOTE method is as follows:

[0021] Step 2-1: Based on the K nearest neighbor algorithm, the nearest sample of each minority class sample is judged, the minority class samples around the noise and decision boundary are accurately identified, and are filtered;

[0022] 1) The majority class set, the minority class set and the complete data set are respectively denoted as M, H and S. M h and M m are the number of minority class samples and the number of majority class samples. For each sample h i in H, the Euclidean distance to each sample in S is calculated, and the nearest sample D(h i ) is found:

[0023]

[0024] where ||h i -s||2 is the Euclidean distance between h i and s.

[0025] 2) If D(h i )∈M, h i is considered as a noise or a minority class sample near the decision boundary, denoted as a non-restricted sample Z u = Z u ∪h i , Z u is the non-restricted minority class sample set of h i . Otherwise, Z q = Z q ∪h i , Z q is the qualified minority class sample set of h i .

[0026] Step 2-2: Based on the adaptive neighbor selection strategy, select the appropriate nearest neighbor that meets the condition of the minority class sample;

[0027] 1) For each sample in Z q , denoted as q j , j = 1, 2, …, M q , find the k nearest neighbors of q j in S by Euclidean distance calculation, denoted as

[0028] 2) If all the k nearest neighbors, i.e. are minority class samples, they are all added to the set Q j , where Q j represents the qualified neighbor set of the qualified minority class sample Q j . Otherwise, perform step 3);

[0029] 3) If at least one of the k nearest neighbors is a majority sample, find the nearest majority sample among the k nearest neighbors of q j , denoted as

[0030] 4) Add the neighbor in the minority class whose distance to q j is less than to the set q j .

[0031] Step 2-3: Based on the linear interpolation method, synthesize the minority class samples;

[0032] 1) For the qualified minority class samples, determine the number of synthesized samples (Ml ) as follows:

[0033]

[0034] 2) randomly select one eligible neighbor from Q i , denoted as q' i , and synthesize a sample by linear interpolation as follows:

[0035] l = q i + ω × (q' i - q i ) (5)

[0036] where ω ∈ (0, 1) is the weight. This random selection and linear interpolation are performed n times for each eligible minority class sample.

[0037] In step 3, considering that obtaining sufficient input features from real-time power systems is crucial for subsequent accurate evaluation results, which leads to an increase in input features as the system size increases. The present application adopts PCA to solve the problem of a large number of input features of the above deep learning model, constructs a feature extraction model, extracts key operating variables highly related to T SM , and establishes an offline data set according to the extracted key operating variables and corresponding class labels.

[0038] In step 4, based on the offline data set, the traditional random forest is enhanced by weighted random sampling to train the ERF-based transient stability evaluation model, extract the mapping relationship between the key operating variables of the power system and the corresponding class labels, and establish the ERF-based transient stability evaluation model.

[0039] For the traditional random forest (RF), the selection of eligible feature subsets at each node is performed by simple random sampling, so almost all of these subsets may contain a large number of non-information features. In order to overcome this limitation of the traditional RF, the present application proposes a new random forest method, namely ERF. This method uses weighted random sampling to assign lower weights to features with less information, and improves the performance of the traditional random forest method by reducing the contribution of trees with less node information. Compared with the traditional random forest method, this method obtains more and better base learners, thereby forming a better fitting.

[0040] In step 5, for the new topology, the evaluation accuracy of the ERF transient stability evaluation old model should be timely checked from the perspective of system stability, and if the evaluation accuracy cannot meet the requirements, the model must be quickly updated to improve the online evaluation performance of the model.

[0041] The ERF transient stability evaluation old model is verified whether the evaluation accuracy meets the requirements, and the specific process is as follows:

[0042] Step 5-1: search the possible topological structure of the system in the future stage through the operator, such as system expansion, large photovoltaic power station or wind turbine grid connection, electrical equipment removal caused by system maintenance, etc.;

[0043] Step 5-2: obtain the operation data of the new scene through time domain simulation, and build a small capacity test set D new ;

[0044] Step 5-3: based on the test set D new , the prediction accuracy P new of the ERF transient stability evaluation model on the new topology is calculated, when P new is less than the minimum expected A set1 of the operator, the transient stability evaluation model accuracy is insufficient, and needs to be updated in time;

[0045] Step 5-4: repeat the above steps until the evaluation accuracy of all ERF transient stability evaluation old models meets the requirements.

[0046] In step 6, with the rapid deployment and development of PMU in the power system, the data acquisition speed of the power system is getting faster and faster. Based on the real-time measurement data required by PMU acquisition, the ERF model is trained to evaluate the transient stability state of the power system, and the online transient stability evaluation is completed.

[0047] The above technical scheme can bring the following technical effects:

[0048] (1) An ASN-SMOTE method is proposed, which effectively solves the sample imbalance problem, and a feature extraction framework is constructed based on a parallel convolution algorithm, which extracts key operating variables highly related to T SM , greatly improving the calculation efficiency of subsequent model evaluation.

[0049] (2) A transient stability evaluation method based on ERF model is proposed, which can realize efficient and accurate real-time evaluation of the transient stability state of modern complex power system, and considering that new topological structure may appear in the system, a fast updating model is established. Compared with general data-driven models, the real-time transient stability evaluation model based on ERF has higher operation efficiency and classification accuracy, and stronger generalization ability. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 is the flow chart of the method of the present application;

[0051] Figure 2 is a schematic diagram of the hierarchical updating strategy in the fast updating model of the present application;

[0052] Figure 3 is a schematic diagram of a New England 10-machine 39-bus system in examples of the present application. DETAILED DESCRIPTION

[0053] The method for considering sample imbalance in power system transient stability assessment based on artificial intelligence, as shown in the formula (1), includes the following steps: Figure 1

[0054] Step 1: Based on time domain simulation technology, obtain power system operation variables, construct transient stability margin index T SM , generate an original data set containing a large amount of system operation state information;

[0055] Specifically, in the simulation technology, an actual power system model is built, 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, load active power and reactive power, line loss, etc., to construct a transient stability margin index, and generate an original data set containing a large amount of system operation state information.

[0056] The transient stability margin can represent the transient stability level of the power system, and is usually used to describe the operation state of the system. The present application uses the critical clearing time and the actual clearing time of the fault to construct the transient stability margin index T SM , and the specific calculation is as follows:

[0057]

[0058] In the formula, t CC is the maximum clearing time of the system under stable operation under the fault; t AC is the actual clearing time of the system under the fault; T SM is a continuous value in the range of -1 to 1. According to the transient stability judgment standard of the actual power system, a threshold α can be set, and T SM is used to classify the operation state under the current system fault into stable state and unstable state, and label it, i.e.

[0059]

[0060] Step 2: Process the original data set based on an ASN-SMOTE method, and establish a sample balanced and rich data set by adaptively synthesizing unstable samples;

[0061] ​Specifically, considering the continuous improvement in the stability of modern power systems, the number of stable samples used for transient stability research far exceeds the number of unstable samples. This resulting sample imbalance problem may have unpredictable impacts on data-driven transient stability assessment methods, posing a significant risk to the stable operation of the system. Therefore, addressing the sample imbalance problem is crucial. This invention proposes the ASN-SMOTE method to process the original dataset and adaptively synthesizes unstable samples, thereby achieving a near-balance between the number of unstable and stable samples and establishing a balanced and rich dataset.

[0062] The principle of the ASN-SMOTE method is as follows:

[0063] Step 2-1: Based on the K-nearest neighbor algorithm, determine the nearest sample of each minority class sample, accurately identify the minority class samples around the noise and decision boundary, and filter them;

[0064] 1) Represent the majority class set, minority class set, and complete dataset as M, H, and S, respectively. M h and M m These represent the number of minority class samples and the number of majority class samples, respectively. For each sample h in H... i Calculate the Euclidean distance to each sample in S, and find the nearest sample D(h). i ):

[0065]

[0066] In the formula: ||h i -s||2 is for h i The Euclidean distance between s and s.

[0067] 2) If D(h) i If ()∈M, then hi is considered a minority class sample near the noise or decision boundary, denoted as an unbounded sample Z. u =Z u ∪h i Z u It is h i The unrestricted minority class sample set. Otherwise, Z q =Z q ∪h i Z q for h i The set of minority class samples that meet the criteria.

[0068] Step 2-2: Based on the adaptive neighbor selection strategy, select the appropriate nearest neighbor that meets the minority class instance condition;

[0069] 1) For Z q Each sample in the dataset is denoted as q. j j = 1, 2, ..., Mq The k nearest neighbors of q in S are found by the Euclidean distance calculation, denoted as j

[0070] 2) If all the k nearest neighbors, i.e. are minority class samples, they are all added to the set Q j , where Q j represents the qualified neighbor set of the qualified minority class samples Q j . Otherwise, step 3) is performed;

[0071] 3) If at least one of the k nearest neighbors is a majority sample, find the nearest majority sample in the k nearest neighbors of q j , denoted as

[0072] 4) Add the neighbor in the minority class whose distance to q j is less than to the set q j .

[0073] Step 2-3: Synthesize the minority class samples based on the linear interpolation method;

[0074] 1) For the qualified minority class samples, determine the number of synthesized samples (M l ) for each qualified minority class sample as follows:

[0075]

[0076] 2) Randomly select a qualified neighbor from Q i , denoted as q' i , and synthesize a sample by linear interpolation as follows:

[0077] l = q i + ω × (q' i - q i ) (5)

[0078] where ω ∈ (0, 1) is the weight value. This random selection and linear interpolation is performed n times for each qualified minority class sample.

[0079] Step 3: Based on PCA, construct a feature extraction model to extract key operating variables highly related to T SM , and establish an offline dataset according to the extracted key operating variables and corresponding class labels;

[0080] ​Specifically, considering that obtaining sufficient input features from the real-time power system is crucial for subsequent accurate evaluation results, which leads to an increase in input features as the system size increases. In addition, the dynamic information of system variables is a large amount of spatio-temporal data, which can be stored in the form of a multi-dimensional array, and has a similar data structure to an image. Inspired by this feature, the application uses PCA to solve the problem of a large number of input features of the above deep learning model, constructs a feature extraction model, and extracts key operating variables highly related to T SM The key operating variables highly related to T

[0081] The specific working principle of the convolution algorithm is as follows:

[0082] The convolution algorithm (CA) is actually a feature extraction method. Through CA, new features that can better represent potential problems are searched. For multi-channel input data, normalization is needed to unify the order of magnitude of different channels. The specific method is as follows:

[0083]

[0084] In the formula: p, q represent the channel and the feature respectively.

[0085] After normalization, the multi-channel data is assumed to be a high-dimensional data plane, and the learned weight kernel is slid onto the data plane. The same weight is shared in the same spatial position (i, j). An abstract feature is generated each time the kernel slides. The mathematical expression of the convolution operation is as follows:

[0086]

[0087] In the formula: W is the learned weight kernel, X is the multi-dimensional data, represents convolution calculation, and b is the bias. The window size and number of the convolution kernel are hyperparameters.

[0088] Step 4: Based on the offline data set, train the ERF-based transient stability evaluation model to extract the mapping relationship between the key operating variables of the power system and the corresponding class labels, and establish the ERF-based transient stability evaluation model;

[0089] Specifically, based on the offline data set, the traditional random forest is enhanced by weighted random sampling to train the ERF-based transient stability evaluation model, extract the mapping relationship between the key operating variables of the power system and the corresponding class labels, and establish the ERF-based transient stability evaluation model.

[0090] Traditional Random Forest (RF) methods select a subset of features that meet certain criteria at each node using simple random sampling. Therefore, almost all of these subsets may contain a large number of non-informative features. To overcome this limitation of traditional RF, this invention proposes a novel Random Forest method, ERF. This method uses weighted random sampling, assigning lower weights to features with less information, thereby improving the performance of traditional Random Forest methods by reducing the contribution of trees with less information at their nodes. Compared to traditional Random Forest methods, this method obtains more and better base learners, resulting in a better fit.

[0091] This invention introduces ERF (Effective Dynamics Response) into transient stability assessment to quickly and accurately evaluate the stable operating state of a power system. By constructing a mapping relationship between key operating variables and corresponding class labels in a trained ERF classifier, the current stable operating state of the power system can be immediately assessed upon obtaining real-time PMU (Power Management Unit) data.

[0092] Step 5: Considering the possibility of new topologies in the system, promptly verify the accuracy of the old ERF transient stability assessment model and quickly establish an updated model;

[0093] Specifically, despite significant efforts made in the offline phase to make the training dataset comprehensive, given the almost unlimited operating conditions of power systems and the inevitable emergence of new network topologies, for new topologies, from the perspective of system stability, the evaluation accuracy of the old ERF transient stability assessment model should be checked promptly. If the evaluation accuracy is insufficient, the model must be updated quickly to improve its online evaluation performance.

[0094] The specific process for verifying whether the evaluation accuracy of the old model for ERF transient stability assessment meets the requirements is as follows:

[0095] Step 5-1: Search for possible topologies of the system in future stages using operators, such as system expansion, grid connection of large photovoltaic power plants or wind turbines, and removal of electrical equipment due to system maintenance.

[0096] Step 5-2: Obtain runtime data for the new scenario through time-domain simulation and construct a small-capacity test set D. new ;

[0097] Step 5-3: Based on test set D new Calculate the prediction accuracy P of the ERF transient stability assessment model on the new topology. new When P new The minimum expectation A of the operator is less than set1 At times, the accuracy of the transient stability assessment model is insufficient, and it needs to be updated promptly;

[0098] A specific hierarchical update strategy is shown in Figure 2 set1 set2 set3 is a parameter manually set by the operator. According to the prediction accuracy P new The update strategy is divided into three levels, and the number of forests updated in different levels is different. The updated forest is a strong classifier, and the un-updated forest is a weak classifier. Larger voting weights are applied to strong classifiers, and the new update process prioritizes weak classifiers.

[0099] Step 5-4: Repeat the above steps until the evaluation accuracy of all ERF transient stability evaluation old models meets the requirements.

[0100] Step 6: Based on the real-time measurement data required by the PMU, the transient stability state is evaluated using the trained ERF model, and the online transient stability evaluation is completed.

[0101] Specifically, with the rapid deployment and development of PMU in the power system, the data acquisition speed of the power system is getting faster and faster. Based on the real-time measurement data required by the PMU, the transient stability state of the power system is evaluated using the trained ERF model, and the online transient stability evaluation is completed. If the evaluation result is transient instability, a risk warning signal is sent to the system operator and emergency control measures are taken immediately to maintain system stability; when the evaluation result is transient stability, the ERF-based transient stability evaluation model continues to monitor the stability of the system.

[0102] Embodiment:

[0103] The embodiment used by the present application is based on a New England 10-machine 39-node system, which consists of 10 generators, 39 buses and 19 load nodes, and the network topology is as shown in Figure 3 The test includes all steps described in the method of the present application, and is carried out on a notebook computer with an Intel Core i7 processor and 16GB of memory.

[0104] ​​​In the process of transient stability simulation, for the 39-bus system, first, 11 different load levels are set in the range of 80% to 130% with a step of 5%, and the generator output is changed accordingly to maintain power balance; In addition, three-phase short-circuit faults are set at 20%, 50%, and 80% of the bus and transmission line, with fault duration set to 0.1s, 0.2s, and 0.3s respectively, and the total simulation period is 10s; Then the time-domain simulation generates 9642 original samples, of which stable samples and unstable samples are about 7:1. After the ASN-SMOTE method oversampling, the number of unstable samples and stable samples is balanced, and each sample contains 403 operating variables. Through the feature extraction framework, 30 operating variables are finally extracted as key operating variables; Finally, 70% of the samples are randomly selected for model training, and the remaining 30% of the samples are used for model verification.

[0105] The evaluation indexes Acc and F 1-score are selected to test the classification performance of the transient stability assessment model based on ERF. Among them, TP is the correct assessment of stable samples, FN is the incorrect assessment of stable samples, FP is the incorrect assessment of unstable samples, and TN is the correct assessment of unstable samples. The specific expressions are as follows:

[0106]

[0107]

[0108] In the formula: P r is the ratio of correctly assessed stable samples to all assessed stable samples; R e is the ratio of correctly assessed stable samples to all actual stable samples; F 1-score is defined as the harmonic mean of precision P r and recall R e ; Acc is defined as the overall accuracy. (The larger the Acc value, the better the classification performance of the TSA model; the value of F 1-score falls in (0, 1), and the closer F 1-score is to 1, the better the classification performance of the TSA model)

[0109] In order to demonstrate the significant classification performance of the ERF-based transient stability assessment model of the present application, it is tested and compared with other classifiers, and five other methods, namely, Logistic Regression (LR), Artificial Neural Network (ANN), Support Vector Machine (SVM), Decision Tree (DT) and Random Forest (RF), are tested using the same input samples. The overall accuracy Acc and F 1-score The comparison results of the aspects are shown in Table 1. As can be seen from the table, compared with other classifier models, the ERF-based TSA model has more significant classification performance, and the Acc and F 1-score are as high as 98.78% and 98.69%, respectively, and the classification performance meets the actual needs and meets the purposes to be achieved by the present application.

[0110] In order to verify the necessity and superiority of the ASN-SMOTE method of the present application, it is used to sample the original data set generated in step one, and the ROS, SMOTE and ADASYN methods are used to sample the original data set, and then the ERF model is used for evaluation, and the comparison test results are shown in Table 2. The results show that the classification performance of the TSA model trained by the data oversampling data set is obviously better than that of the TSA model trained by the original data, and the oversampling method used in the present application is more excellent than other oversampling methods in the subsequent model evaluation stage. It can be seen that the ASN-SMOTE method meets the actual needs and meets the purposes to be achieved by the present application.

[0111] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application and not to limit them, and the ordinary skilled in the art can still modify or equivalently replace the specific embodiments of the present application according to the above examples, and any modification or equivalent replacement of the present application without departing from the spirit and scope of the present application is within the protection scope of the claims of the present application.

[0112] Table 1

[0113] Model Acc F 1-score ]]> LR 96.31％ 96.08％ ANN 97.48％ 97.13％ SVM 97.69％ 97.42％ DT 95.25％ 94.94％ RF 97.81％ 97.66％ ERF 98.78％ 98.69％

[0114] Table 2

[0115] Sampling method Acc F 1-score ]]> No sampling 94.01％ 93.27％ ROS 96.45％ 95.87％ SMOTE 96.78％ 96.77％ ADASYN 97.65％ 96.58％ ASN-SMOTE 98.39％ 98.11％

Claims

1. A transient stability assessment method for power systems based on artificial intelligence and considering sample imbalance, characterized in that, Includes the following steps: Step 1: Based on time-domain simulation technology, obtain the operating variables of the power system and construct the transient stability margin index T. SM This generates a raw dataset containing a large amount of system operating status information. Step 2: A synthetic minority oversampling method, ASN-SMOTE, with adaptive qualified synthesizer selection is proposed to process the original dataset and establish a balanced and rich dataset by adaptively synthesizing unstable samples; the details are as follows: By employing the ASN-SMOTE method to process the original dataset and adaptively synthesizing unstable samples, the number of unstable and stable samples is roughly balanced, thus establishing a balanced and rich dataset. The principle of the ASN-SMOTE method is as follows: Step 2-1: Based on the K-nearest neighbor algorithm, determine the nearest sample of each minority class sample, accurately identify the minority class samples around the noise and decision boundary, and filter them; Step 2-2: Based on the adaptive neighbor selection strategy, select the appropriate nearest neighbor that meets the minority class sample conditions; Steps 2-3: Based on the linear interpolation method, synthesize the minority class samples; Step 3: Based on the parallel convolution algorithm PCA, construct a feature extraction model to extract features related to T. SM Highly relevant key runtime variables; an offline dataset is built based on the extracted key runtime variables and their corresponding class labels. Step 4: Based on the offline dataset, train a transient stability assessment model based on enriched random forest (ERF), extract the mapping relationship between key operating variables of the power system and their corresponding class labels, and establish an ERF-based transient stability assessment model. Step 5: Considering the possibility of new topologies in the system, promptly verify the accuracy of the old ERF transient stability assessment model and quickly establish an updated model; Step 6: Collect the required real-time measurement data based on the phasor measurement unit (PMU), and evaluate the transient stable state using the trained ERF model to complete the online transient stability assessment.

2. The transient stability assessment method for power systems based on artificial intelligence and considering sample imbalance, as described in claim 1, is characterized in that: In step 1, a model of the actual power system is built using simulation technology, and batch time-domain simulations are performed with different fault times, fault locations, and load levels to obtain power system operating variables, construct transient stability margin indices, and generate a raw dataset containing a large amount of system operating status information.

3. The transient stability assessment method for power systems based on artificial intelligence and considering sample imbalance, as described in claim 2, is characterized in that: The transient stability margin index T is constructed using the critical clearance time and actual clearance time of the fault. SM The specific calculations are as follows: In the formula: t CC The maximum clearing time for the system to maintain stable operation under this fault; t AC This refers to the actual time the system cleared the fault under this condition; T SM Given a continuous value within the range of -1 to 1, a threshold α can be set according to the transient stability evaluation criteria of actual power systems, combined with T. SM The current system's operating state under fault conditions is classified into stable and unstable states, and labeled accordingly:

4. The transient stability assessment method for power systems based on artificial intelligence and considering sample imbalance, as described in claim 1, is characterized in that: In step 3, PCA is used to address the issue of a large number of input features in the data-driven model described above, and a feature extraction model is constructed to extract features related to T. SM Highly relevant key runtime variables were identified, and an offline dataset was built based on the extracted key runtime variables and their corresponding class labels.

5. The transient stability assessment method for power systems based on artificial intelligence and considering sample imbalance according to claim 1, characterized in that: In step 4, based on the offline dataset, the traditional random forest is enhanced by weighted random sampling, an ERF-based transient stability assessment model is trained, the mapping relationship between key operating variables of the power system and their corresponding class labels is extracted, and an ERF-based transient stability assessment model is established.

6. The transient stability assessment method for power systems based on artificial intelligence and considering sample imbalance according to claim 1, characterized in that: In step 5, for the new topology, from the perspective of system stability, the evaluation accuracy of the old model for ERF transient stability assessment is checked in a timely manner. If the evaluation accuracy does not meet the requirements, the model must be updated quickly to improve the online evaluation performance of the model.

7. The transient stability assessment method for power systems based on artificial intelligence and considering sample imbalance, as described in claim 6, is characterized in that: The specific process for verifying whether the evaluation accuracy of the old model for ERF transient stability assessment meets the requirements is as follows: Step 5-1: Search for possible topologies of the system in future stages using operators; Step 5-2: Obtain runtime data for the new scenario through time-domain simulation and construct a small-capacity test set D. new ; Step 5-3: Based on test set D new Calculate the prediction accuracy P of the ERF transient stability assessment model on the new topology. new When P new The minimum expectation A of the operator is less than set1 At times, the accuracy of the transient stability assessment model is insufficient, and it needs to be updated promptly. Step 5-4: Repeat the above steps until the evaluation accuracy of all ERF transient stability assessment old models meets the requirements.

Images