A power system transient voltage stability evaluation method based on intelligent feature selection

Abstract

The application discloses a kind of power system transient voltage stability evaluation method based on intelligent feature selection belonging to the technical field of power system transient voltage stability evaluation.The method comprises: based on SMRMR, the feature is initially screened;Convert mutual information value into SU value, obtain the redundancy between any two features and the correlation between each feature and category based on SU value, then use incremental search to select the approximate optimal feature defined by MRMR;Based on QIBSO, the approximate optimal feature selected in step 1 is searched and selected;Based on the composite screening framework, the transient voltage stability is evaluated;From the time sequence dimension and the space dimension, the original feature is constructed, and the feature subset selected in step 2 is used as the input of machine learning model to evaluate the power system transient voltage stability online.The method can greatly reduce the dimension of feature subset, improve the evaluation accuracy and reduce the calculation time of online evaluation.

Description

Technical Field

[0001] This invention relates to the field of power system transient voltage stability assessment technology, and in particular to a power system transient voltage stability assessment method based on intelligent feature selection. Background Technology

[0002] The continuous growth in the grid connection scale of new energy power sources such as wind and solar power, the "hollowing out" of load centers caused by the withdrawal of traditional generator sets, and the high proportion of power electronic equipment in the new power grid operation characteristics will make the transient voltage stability problem more prominent, and the characteristic quantities reflecting system stability will also gradually increase.

[0003] Feature selection is a fundamental task in machine learning and statistics, effectively selecting a subset of low-dimensional key features from a set of original high-dimensional features to improve classifier evaluation accuracy. Basic methods include filtering, embedding, and wrapping. Filtering uses feature correlation scores created based on statistical measures to select features, making it relatively easy to compute. However, filtering ranks features based only on a single association between features and class labels, thus tending to ignore correlations between features and is generally used in the early stages of the feature selection process. Embedding automatically selects important features based on feature weights during model training, offering high accuracy but suffering from overfitting and high computational resource consumption. Wrapping integrates a classification algorithm into the feature selection process, selecting a feature subset based on classification performance. Because it directly optimizes the target classification algorithm, wrapping typically achieves better classification performance and generalization ability than the other two methods.

[0004] In fact, if the original feature set has a dimension of N, finding the optimal subset requires 2... N Searching on such scales is computationally extremely difficult. Therefore, some studies have begun to employ metaheuristic-based optimization methods to find optimal feature subsets within a reasonable timeframe based on evaluation functions and stopping criteria, and have attempted to apply these methods to the selection of feature subsets for transient stability in power systems. While metaheuristic algorithms perform well in solving feature selection problems, single metaheuristic methods still fall short when facing ultra-large-scale feature selection problems.

[0005] In recent years, the rapid development of artificial intelligence has provided a new perspective for the field of intelligent optimization. Hybrid metaheuristic methods combining intelligent optimization components are expected to better address the aforementioned challenges. How to extract spatiotemporal key response features suitable for transient voltage stability assessment of novel power systems from ultra-high-dimensional features, and how to improve metaheuristic methods to find a more suitable subset of key response features for transient voltage stability assessment, are pressing issues that need to be addressed. Summary of the Invention

[0006] The purpose of this invention is to propose a power system transient voltage stability assessment method based on intelligent feature selection, characterized in that the method includes the following steps:

[0007] Step 1: Perform initial screening of features based on SMRMR; convert mutual information values ​​into SU values, obtain the redundancy between any two features and the correlation between each feature and the category based on the SU values, and then use incremental search to select the approximate optimal features defined by MRMR;

[0008] Step 2: Based on QIBSO, refine the near-optimal features selected in Step 1;

[0009] Step 3: Evaluate transient voltage stability based on a composite screening framework; construct original features from the temporal and spatial dimensions, and use the feature subset selected in Step 2 as input to the machine learning model to evaluate the transient voltage stability of the power system online.

[0010] Step 2 specifically includes:

[0011] Step 21: Generate an initial solution randomly or heuristically, generate a set of reference solutions from the initial solution as a search region, and add the initial solution to the tabu list so that it will not be called again;

[0012] Step 22: Assign the reference solution to the bee for local search, during which time the visited solutions are evaluated and local optima are sought;

[0013] Step 23: After the local search is completed, each bee returns the optimal solution it found and adds all the optimal solutions to the dance table to improve the global optimal solution or generate a random solution;

[0014] Step 24: When the global optimal solution is improved, the next iteration selects the global optimal solution as the initial solution for development; if all dance solutions are in the tabu list, a random solution is generated as the initial solution, and the dance solution with the greatest difference from all solutions in the tabu list is selected for exploration; stop when the maximum number of iterations is reached.

[0015] Step 22 specifically includes:

[0016] Step 221: Each bee inputs a reference solution from the search domain and determines the next possible solution to visit based on the reference solution; the bees are considered as agents, constituting the state s of the environment. t It is a subset of the features of all possible solutions near the bee, action a t To flip, r t For a pair (s) t ,a t Related rewards;

[0017] Step 222: Calculate the Q-value for each accessible solution; the environment consists of a set of states S = {s1, s2, ..., s...} n As described in}, A = {a1, a2, ..., a} n} represents each state s of the agent in S. i A set of selectable actions, each time a certain state s is reached. t Perform an action a t The intelligent agent will then receive a corresponding reward. t The state-value function of strategy π The advantages and disadvantages of MDP strategies using infinite horizon and discounted value can be expressed as follows:

[0018]

[0019] In the formula, γ represents the expected total reward obtained by choosing an action according to policy π starting in state s; 0≤γ≤1 is the discount parameter, which is modified from the original fixed discount parameter to a variable that decreases with the number of iterations, and its variation is shown in the following formula:

[0020]

[0021] In the formula, i is the iteration number, I max The maximum number of iterations;

[0022] Define the following equation as the optimal Q-function for all state pairs.

[0023]

[0024] The optimal value function is then...

[0025]

[0026] The MDP problem simplifies to finding the optimal value of the Q-function for all action-state pairs, and is solved iteratively using the temporal difference method; the Q-function is updated according to the following rules:

[0027]

[0028] In the formula, the learning rate α is used to determine the impact of new information on the existing Q value, and a' is the next possible action;

[0029] Step 223: Store the Q-value of each calculated solution in the bee's Q-table; where each box in the first dimension of the optimized Q-table is a vector representing the number of features the solution has, and its size is the total number of features n retained after SMRMR dimensionality reduction; the second dimension is a list summarizing the combinations of the same number of features in the solution, which will be gradually constructed and contains information characterizing the solution, namely the Q-value and the action that led to the state;

[0030] Step 224: All bees share the Q-table and decide which action to take to maximize long-term gains under the guidance of the Q-table; the Q-table of each bee is shared with other bees so that bees can search based on the globally optimal Q-value during the search process;

[0031] Step 225: Introduce an exploration factor ε to give each action an exploration opportunity, with the probability generated randomly; when the probability is less than ε, the bee does not refer to the Q table and randomly selects an action to perform; when the probability is greater than ε, the bee executes the optimal search strategy.

[0032] In step 22, when evaluating the solution for the access, the F-beta score is used for missed detection penalty, i.e.

[0033]

[0034] In the formula, Err is the classification error obtained by a classifier using the feature subset for evaluation; β is the weight; P is the subset evaluation index precision; and R is the subset evaluation index recall.

[0035] The spatial dimensions of step 3 include the output and power angle of thermal power units, the output of new energy sources, the power flow and load power of the 500kV main grid transmission line, the voltage, phase and frequency changes of the 500kV bus, the trigger delay angle of the positive and negative poles at the sending and receiving ends, DC current, DC voltage, arc extinction angle, commutation angle, and DC power.

[0036] The beneficial effects of this invention are as follows:

[0037] 1. The method of the present invention can greatly reduce the dimensionality of feature subsets, improve the evaluation accuracy, and reduce the online evaluation calculation time.

[0038] 2. The method of this invention improves upon the original optimization algorithm, which suffers from high initial fitness, slow convergence speed, and getting trapped in local optima, effectively enhancing optimization efficiency and fitness value. Attached Figure Description

[0039] Figure 1 This is a flowchart of the power system transient voltage stability assessment method based on intelligent feature selection according to the present invention;

[0040] Figure 2The Q-table used in this invention is a common Q-table.

[0041] Figure 3 This is a schematic diagram of the QIBSO algorithm structure;

[0042] Figure 4 Wiring diagram for a hybrid AC / DC power grid with a high proportion of renewable energy sources;

[0043] Figure 5 The fitness curves for each optimization algorithm are shown. Detailed Implementation

[0044] This invention proposes a method for evaluating the transient voltage stability of power systems based on intelligent feature selection. The invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0045] Figure 1 This is a flowchart of the power system transient voltage stability assessment method based on intelligent feature selection according to the present invention; specifically including:

[0046] 1. Feature screening based on SMRMR

[0047] 1) Mutual information describes the amount of information shared between two random variables, that is, the correlation between two random variables X and Y. It is defined as:

[0048]

[0049] In the formula, the random variable X = {x1, x2, ..., x} N}, x i Let X represent the possible values ​​of X, and p(x) be the probability density of X. The random variable Y is defined similarly to X; p(x,y) is the joint probability density of X and Y.

[0050] 2) Symmetric uncertainty (SU) feature evaluation methods measure feature quality by determining the uncertainty of a feature relative to other feature sets. Therefore, mutual information values ​​are converted to SU values ​​to compensate for the bias of mutual information methods towards features with more values. The formula for calculating SU is:

[0051]

[0052]

[0053] Here, H(X) and H(Y) are the marginal entropies of X and Y, respectively.

[0054] 3) After obtaining the redundancy between any two features and the correlation between each feature and the category based on the SU value, an incremental search method is used to select the approximate optimal feature defined by MRMR. These two calculation steps are collectively referred to as the SMRMR method, as shown in equations (4)-(8). The importance ranking of the feature subsets based on SMRMR can be obtained by sorting them in descending order of Φ(D,R).

[0055]

[0056]

[0057] maxΦ(D,R),Φ=DR (6)

[0058]

[0059] In the formula, F is the original feature set, S is the selected feature subset, and F m-1 A subset containing m-1 features; f i f j Let S represent the i-th and j-th features in S, D(S,C) represent the correlation between the feature set and the category, R(S) represent the redundancy between features, and Φ(D,R) represent the simultaneous optimization of correlation and redundancy; |*| is the set dimension calculated.

[0060] 2. Feature Refinement Based on QIBSO

[0061] To further optimize the selected subset, an improved metaheuristic bee colony optimization algorithm (QIBSO) based on reinforcement learning is proposed for feature selection. The search process consists of four steps:

[0062] 1) Generate an initial solution randomly or heuristically, determine a search region from it, and add the initial solution to the tabu list so that it is no longer called.

[0063] The solution is represented as a Boolean vector of length n after SMRMR selection. If a feature is selected, the corresponding position is set to 1; otherwise, it is set to 0.

[0064] The search region is a set of reference solutions generated from the initial solution. Reference solutions are generated by flipping n / flip features in the initial solution. Flipping a boolean feature from 0 to 1 adds a feature and deleting it, while flipping represents an empirical parameter. The number of reference solutions equals the number of bees.

[0065] For the initial iteration, the initial solution is generated in a purely random manner. A Pareto-based initial solution selection scheme is then adopted instead. The new method selects a fixed number of features as the initial solution to reduce the possible combinations of features. Based on the Pareto principle-determined initial solution, the retained feature combinations are guided by the SMRMR ranking score Φ, prioritizing the retention of feature combinations with higher Φ values ​​as the initial solution. Similarly, when generating the reference solution, features with higher Φ values ​​are preferentially flipped.

[0066] The remaining iterations are generated based on the principles of exploration (searching for a global solution) or development (enhancing a local solution).

[0067] 2) A reference solution is assigned to the bee. The bee performs a local search, during which it evaluates the solutions visited and seeks a locally optimal solution.

[0068] Improved fitness value through missed detection penalty strategy: The optimization algorithm is evaluated based on achieving higher evaluation accuracy and lower feature subset dimensionality. The fitness value (Fit) is typically used as the metric, and its objective function is defined as follows:

[0069]

[0070] In the formula, F is the set of candidate features retained by SMRMR, S is the selected feature subset, and 0 < λ1 and λ2 ≤ 1 are weight coefficients.

[0071] Err represents the classification error obtained by a classifier when evaluating a subset of features. The transient voltage stability assessment problem in power systems is particularly concerned with missed detections, as these can lead to serious consequences such as system collapse. Positive classes represent stable samples, and negative classes represent unstable samples. Missed detections (instability misclassified as stability) are denoted as false positives (FP), incorrect detections (stability misclassified as instability) are denoted as false negatives (FN), and correct stability and instability judgments are denoted as true positives (TP) and true negatives (TN), respectively. Based on the above concepts, the subset evaluation metrics Precision (P) and Recall (R) are defined as follows:

[0072]

[0073]

[0074] To effectively reduce the false negative rate, we give greater attention to functional programming (FP) when setting the Fit objective function, and use the F-beta score for false negative penalty, as follows:

[0075]

[0076] In this embodiment, β is set to 0.2 to give greater weight to accuracy.

[0077] Q-learning replaces blind local search: A Q-learning reinforcement learning algorithm is used to replace the blind local search of bees in the original bee colony optimization to speed up the optimization process. It includes the following steps:

[0078] i) Each bee inputs a reference solution from the search domain and determines the next possible solution to visit based on the reference solution. Depending on the specific problem, define the elements and environment: treat the bee as an agent, and define the state s of the environment. t It is a subset of the features of all possible solutions near the bee, action a t To flip, with a pair (s t ,a t Related rewards r t Calculated based on equation (8).

[0079] ii) Calculate the Q-value for each accessible solution. Suppose the environment consists of a set of states S = {s1, s2, ..., s...} n As described in}, A = {a1, a2, ..., a} n} represents each state s of the agent in S. i A set of actions that can be selected, whenever a certain state s is reached. t Perform an action a t The intelligent agent will then receive a corresponding reward. t The state-value function of strategy π The merits and demerits of the MDP strategy are quantified by the concept of infinite horizon and discounting, and can be expressed as follows:

[0080]

[0081] In the formula, E π This represents the expected total reward obtained by choosing an action according to policy π starting in state s; 0≤γ≤1 is a discount parameter. If γ is close to 0, the agent tends to choose immediate rewards; otherwise, it tends to consider long-term rewards. Consider using a larger value in the early stages of iteration to consider a larger proportion of future rewards, and a smaller value in the later stages to focus on immediate development rewards. The original fixed parameter value is modified to a variable that decreases with the number of iterations, and its variation law is shown in the following formula:

[0082]

[0083] In the formula, i is the iteration number, I max This represents the maximum number of iterations.

[0084] If equation (14) is the optimal Q-function for all state-action pairs, then the agent's role is to learn an optimal policy π*(S→A) that maximizes the Bellman equation, i.e., the optimal value function can be expressed as equation (15):

[0085]

[0086]

[0087] The MDP problem simplifies to finding the optimal value of the Q-function for all action-state pairs. Since solving the Bellman equation is computationally complex, the Temporal Difference (TD) method is used iteratively. The Q-function is updated according to the following rules:

[0088]

[0089] In the formula, the learning rate α is used to determine the impact of new information on the existing Q value, and a' is the next possible action.

[0090] iii) Store the Q value of each calculated solution in the bee's Q table. Figure 2 The original Q-table, constructed as an m×n matrix, wastes storage resources and access time. In the optimized Q-table, each box in the first dimension is a vector representing the number of features the solution possesses; therefore, its size is the total number of features n retained after SMRMR dimensionality reduction. The second dimension is a list summarizing combinations of the same number of features in the solution. This list is gradually constructed and contains information characterizing the solution: the Q-value and the action that led to that state.

[0091] iv) All bees share a Q-table, which guides their decisions on which actions to maximize long-term gains. Each bee's Q-table is shared with other bees, allowing them to search based on the globally optimal Q-value during the search process.

[0092] v) Introduce a trade-off strategy between development and exploration: By introducing an exploration factor ε, each action is given a certain exploration opportunity: the probability is generated randomly. When the probability is less than ε, the bee does not refer to the Q table and randomly selects an action to perform; when the probability is greater than ε, the bee executes the optimal search strategy.

[0093] 3) After the search is completed, each bee returns the optimal solution it found and adds it to the dance table.

[0094] 4) When the global optimal solution is improved, the next iteration selects this optimal reference solution as the initial solution for development; otherwise, exploration continues. During exploration, the dance table solution that differs most significantly from all solutions in the tabu list is selected. If all dance solutions are in the tabu list, a random solution is generated as the initial solution. The algorithm stops when the maximum number of iterations is reached.

[0095] The overall flow and optimization process of the proposed QIBSO algorithm are as follows: Figure 3 As shown.

[0096] 3. Application of composite screening framework in transient voltage stability assessment

[0097] The application of the composite screening framework in transient voltage stability assessment mainly involves three steps: original feature construction, composite feature selection, and online assessment.

[0098] 1) Construction of original features: Construct the original feature set, which is constructed in two dimensions: temporal and spatial.

[0099] The temporal dimension is constructed using the sliding window sample generation method, resulting in a total of 4 dimensions.

[0100] In the spatial dimension, a comprehensive range of electrical quantities are collected, and the impact of renewable energy and DC access on the system's transient voltage stability is fully considered. These include: thermal power unit output and power angle, renewable energy output, 500kV main grid transmission line power flow and load power; 500kV bus voltage, phase and frequency variations; and DC parameters, including the positive and negative trigger delay angles of the sending and receiving ends, DC current, DC voltage, arc extinction angle, commutation angle, and DC power. The total spatial feature dimension is 292 dimensions.

[0101] By combining temporal and spatial features, each sample's original feature set contains 1168-dimensional features.

[0102] 2) Selection of composite features:

[0103] The SMRMR filtering algorithm is used to initially screen the original spatiotemporal feature data to obtain an optimized subset.

[0104] Based on the obtained optimized subset, the QIBSO algorithm is used and combined with the classification effect on the machine learning model for optimization, and a new feature subset is determined.

[0105] 3) Online evaluation: Use the new feature subset obtained in step 2) as input for classification by the machine learning model.

[0106] A kNN classifier is used to train the classification model, and the classifier is validated using 10-fold cross-validation with averaging. Based on the Missing Alarm Rate (MAR) and False Alarm Rate (FAR), the composite screening framework model evaluation metrics, Missing Alarm Rate (MAR) and False Alarm Rate (FAR), and the Accuracy (ACC), are defined as follows:

[0107]

[0108]

[0109]

[0110] The results are output using a sliding window majority voting method based on the majority voting concept. Each cross-section within the sliding window provides its evaluation result for that moment. When the majority results are consistent, a judgment result is output to determine whether to issue an instability alarm. Taking all factors into consideration, the sliding window length is set to 6 time-series cross-sections, and the voting threshold is 5 or 6 votes, to balance the requirements of total evaluation time and evaluation accuracy.

[0111] Example Analysis:

[0112] Figure 4 The diagram shows the wiring diagram of a hybrid AC / DC power grid with a high proportion of renewable energy integration, consisting of 100 nodes, including 32 nodes in the 500kV main grid. The installed capacity ratio is renewable energy (2.4 million kW): conventional power (6.3 million kW) = 1:2.62. The basic operation mode is 1.8 million kW of renewable energy generation (1.2 million kW of wind power and 0.6 million kW of photovoltaic power) and 3.86 million kW of conventional power generation. The system has one DC circuit with a DC power input of 800,000 kW. A simulation example was built based on the Electric Power Research Institute's BPA simulation platform, and data processing and model training were performed jointly in MATLAB 2020 and Python 3.9.

[0113] Evaluation of the effectiveness of feature selection in composite frameworks:

[0114] Table 1 shows that after two stages of fine screening, only 15.02 features with high importance weights were retained from the original feature set, and over 98% of the features were considered to have no significant impact on the learning process. Regarding evaluation accuracy, the evaluation model can deliver the lowest false negative rate with the minimum feature subset dimension, and improves the overall evaluation accuracy. Due to the significantly reduced feature subset dimension, the stability determination result can be obtained in just 0.041 seconds during online evaluation, meeting the requirements of online real-time applications. This verifies the effectiveness of the proposed feature selection method in transient voltage stability evaluation.

[0115] Table 1 Comparison of Feature Selection Evaluation Indicators

[0116]

[0117] The convergence results of four commonly used heuristic optimization algorithms in the packaging method are compared with those of the proposed method in this invention, including Whale Algorithm (WOA), Particle Swarm Optimization (PSO), Genetic Algorithm (GA), and Grey Wolf Algorithm (GWO). Furthermore, to verify the effectiveness of the proposed improvement strategy for the original BSO algorithm, the convergence curve is compared with that of QIBSO. The average convergence curve after 20 tests is shown below. Figure 5 As shown.

[0118] The QIBSO hybrid metaheuristic algorithm, as an improvement upon BSO, firstly improves the initial and reference solution generation process, enabling the algorithm to search around local or global optima from the early stages, resulting in the lowest initial fitness value. Secondly, it replaces the blind local search process with a unique Q-learning mechanism, allowing each bee to benefit from its own and other bees' Q-tables to perform optimal search actions, minimizing useless search processes and resulting in a significant decrease in the fitness curve in the early stages, enabling a faster transition to the fine-grained development phase. Thirdly, the development-exploration tradeoff strategy helps the agent tend to explore unknown solutions in the early stages to escape local optima. Ultimately, it achieves a better solution than BSO, with the lowest final fitness value.

[0119] In addition, the above methods are compared in terms of feature subset dimension, accuracy indicators of the three models, and optimization time, as shown in Table 2.

[0120] Table 2 Comparison of Evaluation Indicators for Various Optimization Methods

[0121]

[0122] By comparing the advantages and disadvantages of different methods in terms of evaluation indicators, the following conclusions can be drawn:

[0123] 1) BSO has a significant advantage over other optimization algorithms in terms of subset dimension, false negative rate and accuracy, which are of great interest. However, it takes a long time to optimize and has a high false negative rate, so there is room for improvement.

[0124] 2) QIBSO has made significant improvements over BSO in its weaknesses, with optimization time reduced by about 20% and false positive rate effectively reduced. Its overall ranking is among the top of all methods.

[0125] This invention proposes a two-stage composite feature selection framework and applies it to transient voltage stability assessment in novel power systems. The first stage employs an improved MRMR method with SU value enhancement for coarse feature selection. A reinforcement learning-based metaheuristic optimization algorithm is then applied to and optimized for the second stage of feature refinement. This selects the optimal feature subset for transient voltage stability assessment, significantly reducing the feature subset dimensionality, improving assessment accuracy, and decreasing online assessment computation time. Furthermore, this invention improves upon the original optimization algorithm's issues of high initial fitness, slow convergence, and getting trapped in local optima, effectively enhancing optimization efficiency and fitness values. It demonstrates significant advantages over common metaheuristic optimization algorithms used for transient voltage stability assessment.

Claims

1. A method for evaluating transient voltage stability of a power system based on intelligent feature selection, characterized in that, The method includes the following steps: Step 1: Perform initial screening of features based on SMRMR; convert mutual information values ​​into SU values, obtain the redundancy between any two features and the correlation between each feature and the category based on the SU values, and then use incremental search to select the approximate optimal features defined by MRMR; Step 2: Based on QIBSO, refine the near-optimal features selected in Step 1; Step 21: Generate an initial solution randomly or heuristically, generate a set of reference solutions from the initial solution as a search region, and add the initial solution to the tabu list so that it will not be called again; The solution is represented as follows: the feature subset selected by SMRMR is represented by a Boolean vector of length n; if the corresponding feature is selected, the position is set to 1, otherwise it is 0. The search region is a set of reference solutions generated from the initial solution; the reference solutions are generated by flipping n / flip features in the initial solution. Flipping a feature boolean from 0 to 1 adds a feature and deleting it, while flipping is an empirical parameter; the number of reference solutions is equal to the number of bees. For the first iteration, a certain number of features are selected as the initial solution. Based on the Pareto principle to determine the initial solution, the feature combinations to be retained are guided by the SMRMR ranking score, and the feature combinations with high ranking scores are retained as the initial solution. When generating the reference solution, the features with high ranking scores are flipped. Step 22: Assign the reference solution to the bee for local search, during which time the visited solutions are evaluated and local optima are sought; Step 22 specifically includes: Step 221: Each bee inputs a reference solution from the search domain and determines the next possible solution to visit based on the reference solution; the bee is considered an agent, constituting the state of the environment. It is a subset of all possible solution features near the bee, actions To flip, For a pair Related rewards; Step 222: Calculate the Q-value for each accessible solution; the environment consists of a set of states. As described, It is the intelligent agent in Each state A set of selectable actions, each time a certain state is reached. Perform an action The intelligent agent will then receive a corresponding reward. The state-value function of strategy π The advantages and disadvantages of MDP strategies using infinite horizon and discounted value can be expressed as follows: ； In the formula, Indicates the state The expected total reward obtained by choosing an action based on strategy π at the beginning; The discount parameter is modified from the original fixed discount parameter to a variable that decreases with the number of iterations, and its variation is shown in the following formula: ； In the formula, Let be the iteration number. The maximum number of iterations; Define the following equation as the optimal Q-function for all state pairs. ； The optimal value function is then... ； The MDP problem simplifies to finding the optimal value of the Q-function for all action-state pairs, and is solved iteratively using the temporal difference method; the Q-function is updated according to the following rules: ； In the formula, the learning rate Used to determine the impact of new information on existing Q values. This is a possible next step; Step 223: Store the Q-value of each calculated solution in the bee's Q-table; where each box in the first dimension of the optimized Q-table is a vector representing the number of features the solution has, and its size is the total number of features n retained after SMRMR dimensionality reduction; the second dimension is a list summarizing the combinations of the same number of features in the solution, which will be gradually constructed and contains information characterizing the solution, namely the Q-value and the action that led to the state; Step 224: All bees share the Q-table and decide which action to take to maximize long-term gains under the guidance of the Q-table; the Q-table of each bee is shared with other bees so that bees can search based on the globally optimal Q-value during the search process; Step 225: Introduce Exploration Factors Each action is given an opportunity to explore, with probabilities generated randomly; when the probability is less than... At that time, the bee does not refer to the Q-table and randomly selects an action to perform; when the probability is greater than At that time, the bee executes the optimal search strategy; Step 3: Evaluate transient voltage stability based on a composite screening framework; construct original features from the temporal and spatial dimensions, and use the feature subset selected in Step 2 as input to the machine learning model to evaluate the transient voltage stability of the power system online.

2. The power system transient voltage stability assessment method based on intelligent feature selection according to claim 1, characterized in that, Step 2 specifically also includes: Step 23: After the local search is completed, each bee returns the optimal solution it found and adds all the optimal solutions to the dance table to improve the global optimal solution or generate a random solution; Step 24: When the global optimal solution is improved, the next iteration selects the global optimal solution as the initial solution for development; if all dance solutions are in the tabu list, a random solution is generated as the initial solution, and the dance solution with the greatest difference from all solutions in the tabu list is selected for exploration; stop when the maximum number of iterations is reached.

3. The power system transient voltage stability assessment method based on intelligent feature selection according to claim 2, characterized in that, In step 22, when evaluating the solution for the access, the F-beta score is used for missed detection penalty, i.e. ； In the formula, The classification error obtained by evaluating a classifier using this subset of features; is the weight; P is the precision of the subset evaluation index, and R is the recall of the subset evaluation index.

4. The power system transient voltage stability assessment method based on intelligent feature selection according to claim 1, characterized in that, The spatial dimensions of step 3 include the output and power angle of thermal power units, the output of new energy sources, the power flow and load power of the 500kV main grid transmission line, the voltage, phase and frequency changes of the 500kV bus, the trigger delay angle of the positive and negative poles at the sending and receiving ends, DC current, DC voltage, arc extinction angle, commutation angle, and DC power.

Images