Transient voltage stability evaluation method based on probabilistic neural network

Abstract

A transient voltage stability evaluation method based on a probabilistic neural network comprises the following steps: step 1, analyzing characteristics after wind power grid connection based on historical operation data of a power system to obtain system operation data samples, and classifying the samples according to critical fault clearing time margin values to form a corresponding initial sample set; 2, for the initial sample set, using a fusion feature t distribution-based random neighbor embedding algorithm to carry out dimension reduction processing on the feature quantity; 3, optimizing parameters of the probabilistic neural network by adopting a genetic ant colony hybrid algorithm; and 4, performing model training by using the training set as the input characteristic quantity of the optimized PNN to obtain an evaluation model, and finally performing transient voltage stability evaluation. According to the method, for security risk situation cognition, prospective early warning and formulation of prevention and emergency control measures, voltage instability and voltage collapse caused by the fact that large-scale new energy is merged into the power grid can be prevented.

Description

Technical Field

[0001] The invention belongs to the field of transient voltage stability assessment of power systems, and in particular relates to a transient voltage stability assessment method based on a probabilistic neural network. Background Art

[0002] With the continuous expansion of wind farms, the inherent random volatility and unpredictability of wind energy have become increasingly significant. After being integrated into the power system, it has posed new challenges to the safe and stable operation of the system. Rapid and reliable transient voltage stability assessment results can be prepared for a rainy day and effectively prevent the occurrence of transient voltage instability, thereby ensuring the reliable operation and continuous power supply capacity of the power system. Therefore, it is crucial to quickly and accurately assess the transient voltage stability after wind power is connected to the grid to ensure the safe and stable operation of the power system.

[0003] The patent document with application publication number CN116316603A discloses a new energy voltage stability analysis method based on high-voltage AC / DC power grid, establishes a static voltage stability index layer and a transient voltage stability index layer, creates corresponding evaluation indicators for the corresponding index layers, and assigns corresponding evaluation values; establishes a triangular fuzzy function, processes the evaluation value to obtain a numerical index, establishes a standard database for the obtained numerical index, makes an adaptive evaluation value for the obtained node voltage based on the index system, and numerically processes the evaluation value to obtain the corresponding evaluation index, compares the evaluation index with the numerical index, and judges the voltage stability state of the current AC / DC large power grid system by the comparison result. When the method needs to establish a triangular fuzzy function to obtain the node voltage during the transient voltage stability evaluation process, it may cause the evaluation process to become complicated, thereby causing the transient voltage stability evaluation result to be unreliable. The patent document with application publication number CN115187042A discloses a probabilistic voltage evaluation method, device and system. The probabilistic voltage evaluation device includes a data acquisition unit, a model construction unit and a result output unit. By dividing the wind and solar scenes, a probability input model and a voltage stability assessment model are constructed according to different wind and solar scenes based on a hybrid vine structure, and according to the inherited Latin hypercube sampling method and the probability input model, the voltage stability assessment model is used to calculate and output the probability voltage assessment result. However, this method requires dividing the wind and solar scenes, and when the scenes are complex, it is impossible to accurately perform transient voltage stability assessment, which may lead to a decrease in the accuracy of the transient voltage stability assessment results, or even make the assessment results unreliable. None of the above methods can provide a reasonable solution that can both simplify the transient voltage assessment process and more accurately perform transient voltage stability assessment on the system.

[0004] Therefore, the present invention proposes a transient voltage stability assessment method based on probabilistic neural network, which is of great significance for ensuring the safe and stable operation of the power system, improving the power supply quality, guiding system planning and transformation, and promoting technological innovation and development, and guarantees the transient voltage stable operation of the power system. Summary of the invention

[0005] The present invention proposes a transient voltage stability assessment method based on a probabilistic neural network, which can more accurately and efficiently assess the transient voltage stability after the connection of renewable energy to the grid.

[0006] In order to achieve the above-mentioned invention object, the present invention adopts the following technical method: a transient voltage stability assessment method based on probabilistic neural network, comprising the following steps:

[0007] Step 1: Based on the historical operation data of the power system and the characteristics of wind power after grid connection, obtain system operation data samples, classify them according to the size of the critical clearing time (CCT) margin value, and form a corresponding initial sample set;

[0008] Step 2: For the initial sample set, a t-distributed Stochastic Neighbor Embedding (t-SNE) algorithm based on fusion features is used to reduce the dimension of the feature quantity;

[0009] Step 3: Use the genetic ant colony hybrid algorithm to optimize the parameters of the probabilistic neural network (PNN);

[0010] Step 4: Use the training set as the input feature quantity of the optimized PNN to train the model and obtain the evaluation model. Finally, substitute the test set into the trained model, use the CCT margin value as the model output, and evaluate the transient voltage stability.

[0011] In step 1, based on the historical operation data of the power system and the characteristics of wind power after grid connection, the system operation data samples are obtained and classified according to the size of the CCT margin value to form the corresponding initial sample set. The calculation formula of the CCT margin value is as follows:

[0012] T s =T cr -T cct (1)

[0013] Where: T s is the CCT margin value; T cr is the critical fault clearing time; Tcct is the fault clearing time.

[0014] In power system fault analysis, if the margin value of CCT is positive, that is, the fault is successfully removed before approaching the critical edge of system stability, such samples are classified as transient stability. On the contrary, when the margin value turns negative, it means that the fault removal operation occurs after the critical edge of system stability, and such samples are considered as transient instability.

[0015] In step 2, the t-SNE algorithm is used to reduce the dimension of the feature quantity for the initial sample set.

[0016] Among them, the specific steps of the t-SNE algorithm are as follows:

[0017] Step 2-1: Assume the original data is A = {a1, a2, ..., a n}, the data after dimensionality reduction is B = {b1, b2, ..., b n};

[0018] Step 2-2: Initialize the number of iterations, learning rate and momentum, and calculate the number of a in high-dimensional space i As the center, a j is a neighboring point, with a probability distribution of Gaussian distribution P j|i and with a j Centered on i is the neighboring point, according to the probability density P of Gaussian distribution i|j , as follows:

[0019]

[0020] Where: a k ∈{a1,a2,…,a n}; σ i and σ j They are the corresponding data points a i and a j The variance of the Gaussian distribution of

[0021] Step 2-3: Calculate the joint probability distribution P ij And normalize it:

[0022]

[0023] Step 2-4: Initialize the reduced-dimensional data B (0) , and calculate b in the low-dimensional space of the t distribution i With b j The probability density Q ij , as follows:

[0024] B (0)={b1,b2,…,b n} (5)

[0025]

[0026] Where: b k ∈{b1,b2,…,b n};

[0027] Step 2-5: H is the loss function in the dimensionality reduction process obtained by Kullback-Leibler divergence KL. The low-dimensional distribution joint probability density Q ij Joint probability density P with high-dimensional distribution ij Fitting, divergence KL is the relative entropy that measures the degree of fit of the joint probability density of two distributions, as follows:

[0028]

[0029] Get the gradient By decreasing the gradient of H, we can learn the appropriate b i , calculate the optimized gradient

[0030]

[0031] Step 2-6: Introduce the learning rate γ and momentum factor α to control the learning accuracy, iterate t times to perform gradient descent, and obtain the low-dimensional data B (t) , the gradient update formula is as follows:

[0032]

[0033] Step 2-7: The learning rate γ controls the step size of the gradient descent, and the momentum factor α indicates the influence of the direction and size of the previous step of gradient descent on the direction and size of the current gradient descent. The values ​​of the two can be adjusted according to the characteristics of the data set and the complexity of the training model. If the current gradient direction is consistent with the previous step gradient direction, increase the weight of this step; if not, reduce the weight.

[0034] In step 3, the genetic ant colony hybrid algorithm is used to optimize the parameters of the PNN. The genetic ant colony hybrid algorithm is an adaptive hybrid algorithm, and its principle is to combine the genetic algorithm and the ant colony algorithm.

[0035] The specific steps are as follows:

[0036] Step 3-1: Encode the weights, thresholds or other key parameters of the PNN to form the initial population of the genetic algorithm. These codes can be binary codes, real number codes or other suitable coding methods. At the same time, initialize the relevant parameters of the ant colony algorithm, such as the initial pheromone concentration, the number of ants, the pheromone evaporation rate, etc.

[0037] Step 3-2: Define a fitness function to evaluate the performance of PNN on a given data set. This function will be used for individual selection in the genetic algorithm and path selection in the ant colony algorithm. Then, according to the fitness function, select individuals with higher fitness from the current population as the parents of the next generation, and perform crossover and mutation operations;

[0038] Step 3-3: Generate new individuals through crossover operation and introduce new genetic information through mutation operation to increase the diversity of the population. In the ant colony algorithm, the pheromone concentration is updated according to the fitness of the path (i.e., different parameter combinations of PNN). A better path will accumulate more pheromones, thereby attracting more ants to choose this path;

[0039] Step 3-4: Use the excellent individuals generated by the genetic algorithm (i.e., the PNN parameter combination with higher fitness) as the initial solution or seed solution of the ant colony algorithm to enhance the search ability of the ant colony algorithm. In the crossover and mutation process of the genetic algorithm, the pheromone heuristic rules of the ant colony algorithm can be introduced to guide the selection of crossover points and the direction of mutation, thereby improving the efficiency of the genetic algorithm;

[0040] Step 3-5: Repeat the above steps to continuously generate new PNN parameter combinations and evaluate them through the fitness function. As the iteration proceeds, the performance of the PNN will gradually improve. When the preset number of iterations is reached, the fitness reaches the preset threshold, or the quality of the solution is no longer significantly improved, the algorithm terminates and outputs the optimal PNN parameter combination.

[0041] In step 4, the training set is used as the input feature quantity of the optimized PNN for model training to obtain the evaluation model. Finally, the test set is substituted into the trained model, the CCT margin value is used as the model output, and the transient voltage stability is evaluated. Among them, the transient voltage stability evaluation accuracy c is used for evaluation, as follows:

[0042]

[0043] Where: N is the total number of training samples; w is the number of samples evaluated correctly.

[0044] By adopting the above technical solution, the present invention has the following technical effects:

[0045] 1) Since the power system is a complex nonlinear system, the access of new energy to the power grid further increases the complexity of the system. The use of traditional linear methods for data processing may lead to low computational efficiency. Therefore, the present invention considers using t-SNE to reduce the dimension of the input features. The uniqueness of the t-SNE algorithm lies in its powerful ability to capture nonlinear relationships, which is particularly suitable when the composite features are derived from multiple data sources and involve complex nonlinear interactions. Through this algorithm, the original high-dimensional composite feature space can be effectively reduced to a low-dimensional space that is easier to understand and analyze, which helps to improve the efficiency of data-driven decision support and model optimization;

[0046] 2) Traditional genetic algorithms and ant colony algorithms are prone to local convergence, which leads to a decrease in search efficiency in the later stage. Therefore, the present invention adopts a genetic ant colony hybrid algorithm to improve search efficiency. Two different population algorithms make up for each other's shortcomings until the results of the two algorithms are consistent, avoiding the algorithm from falling into local convergence, and can effectively optimize the parameters and structure of PNN, improving its classification performance and generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The present invention will be further described below in conjunction with the accompanying drawings and embodiments:

[0048] Figure 1 is a flow chart of the method of the present invention;

[0049] Figure 2 It is the flow chart of genetic ant colony hybrid algorithm;

[0050] Figure 3 This is the basic structure diagram of PNN;

[0051] Figure 4 It is a schematic diagram of the improved IEEE 39-node system in an example of the present invention. DETAILED DESCRIPTION

[0052] The present invention will be further described below in conjunction with the accompanying drawings.

[0053] The present invention proposes a transient voltage stability assessment method based on probabilistic neural network. Figure 1 As shown, the method comprises the following steps:

[0054] Step 1: Based on the historical operation data of the power system and the analysis of the characteristics of wind power after grid connection, obtain system operation data samples, classify them according to the size of the CCT margin value, and form the corresponding initial sample set. The calculation formula of the CCT margin value is as follows:

[0055] T s =T cr -T cct (1)

[0056] Where: T s is the CCT margin value; T cr is the critical fault clearing time; T cct is the fault clearing time.

[0057] In power system fault analysis, if the margin value of CCT is positive, that is, the fault is successfully removed before approaching the critical edge of system stability, such samples are classified as transient stability. On the contrary, when the margin value turns negative, it means that the fault removal operation occurs after the critical edge of system stability, and such samples are considered as transient instability.

[0058] Step 2: For the initial sample set, use the t-SNE algorithm to reduce the dimension of the feature quantity. The specific steps of the t-SNE algorithm are as follows:

[0059] Step 2-1: Assume the original data is A = {a1, a2, ..., a n}, the data after dimensionality reduction is B = {b1, b2, ..., b n};

[0060] Step 2-2: Initialize the number of iterations, learning rate and momentum, and calculate the number of a in high-dimensional space i As the center, a j is a neighboring point, with a probability distribution of Gaussian distribution P j|i and with a j Centered on i is the neighboring point, according to the probability density P of Gaussian distribution i|j , as follows:

[0061]

[0062] Where: a k ∈{a1,a2,…,a n}; σ i and σ j They are the corresponding data points a i and a j The variance of the Gaussian distribution of

[0063] Step 2-3: Calculate the joint probability distribution P ij And normalize it:

[0064]

[0065] Step 2-4: Initialize the reduced-dimensional data B (0) , and calculate b in the low-dimensional space of the t distribution i With b j The probability density Q ij , as follows:

[0066] B (0) ={b1,b2,…,b n} (5)

[0067]

[0068] Where: b k ∈{b1,b2,…,b n};

[0069] Step 2-5: H is the loss function in the dimensionality reduction process obtained by Kullback-Leibler divergence KL. The low-dimensional distribution joint probability density Q ij Joint probability density P with high-dimensional distribution ij Fitting, divergence KL is the relative entropy that measures the degree of fit of the joint probability density of two distributions, as follows:

[0070]

[0071] Get the gradient By decreasing the gradient of H, we can learn the appropriate b i , calculate the optimized gradient

[0072]

[0073] Step 2-6: Introduce the learning rate γ and momentum factor α to control the learning accuracy, iterate t times to perform gradient descent, and obtain the low-dimensional data B (t) , the gradient update formula is as follows:

[0074]

[0075] Step 2-7: The learning rate γ controls the step size of the gradient descent, and the momentum factor α indicates the influence of the direction and size of the previous step of gradient descent on the direction and size of the current gradient descent. The values ​​of the two can be adjusted according to the characteristics of the data set and the complexity of the training model. If the current gradient direction is consistent with the previous step gradient direction, increase the weight of this step; if not, reduce the weight.

[0076] Step 3: Use the genetic ant colony hybrid algorithm to optimize the parameters of PNN. PNN consists of four layers of nodes, and the specific structure is as follows:

[0077] (1) Input layer: responsible for inputting feature vectors into the network. The number of neurons in the input layer is equal to the number of sample features.

[0078] (2) Pattern layer: This layer contains one neuron for each training set. The activation function used to activate PNN, namely the radial basis function (RBF), is embedded in this layer.

[0079] Subtract the input weights (IW1,1) from the input vector p and sum the squares of the differences to get the Euclidean distance, which represents how close the input is to the training set vector. These elements are multiplied one by one and the bias x is sent to the radial basis transfer function using the dot product function. The output y is:

[0080] y=radbas(||IW 1.1 -p||x) (11)

[0081] Where: radbas is the radial basis activation function, and its general form is:

[0082]

[0083] Where: m is the number of test data, and the training algorithm used to train RBF is the orthogonal least squares method, which provides a systematic method for selecting the RBF center.

[0084] (3) Summation layer: It is responsible for connecting the pattern layer units of each class, calculating the sum of the matching degrees of each class of samples, and taking the average to obtain the initial probability sum of the input sample belonging to each class. The number of neurons in the summation layer is equal to the number of sample categories.

[0085] (4) Output layer: responsible for outputting the category with the highest score in the summation layer.

[0086] The genetic ant colony hybrid algorithm is an adaptive hybrid algorithm that combines the genetic algorithm with the ant colony algorithm. The specific steps are as follows:

[0087] Step 3-1: Encode the weights, thresholds or other key parameters of the PNN to form the initial population of the genetic algorithm. These codes can be binary codes, real number codes or other suitable coding methods. At the same time, initialize the relevant parameters of the ant colony algorithm, such as the initial pheromone concentration, the number of ants, the pheromone evaporation rate, etc.

[0088] Step 3-2: Define a fitness function to evaluate the performance of PNN on a given data set. This function will be used for individual selection in the genetic algorithm and path selection in the ant colony algorithm. Then, according to the fitness function, select individuals with higher fitness from the current population as the parents of the next generation, and perform crossover and mutation operations;

[0089] Step 3-3: Generate new individuals through crossover operation and introduce new genetic information through mutation operation to increase the diversity of the population. In the ant colony algorithm, the pheromone concentration is updated according to the fitness of the path (i.e., different parameter combinations of PNN). A better path will accumulate more pheromones, thereby attracting more ants to choose this path;

[0090] Step 3-4: Use the excellent individuals generated by the genetic algorithm (i.e., the PNN parameter combination with higher fitness) as the initial solution or seed solution of the ant colony algorithm to enhance the search ability of the ant colony algorithm. In the crossover and mutation process of the genetic algorithm, the pheromone heuristic rules of the ant colony algorithm can be introduced to guide the selection of crossover points and the direction of mutation, thereby improving the efficiency of the genetic algorithm;

[0091] Step 3-5: Repeat the above steps to continuously generate new PNN parameter combinations and evaluate them through the fitness function. As the iteration proceeds, the performance of the PNN will gradually improve. When the preset number of iterations is reached, the fitness reaches the preset threshold, or the quality of the solution is no longer significantly improved, the algorithm terminates and outputs the optimal PNN parameter combination.

[0092] Step 4: Use the training set as the input feature quantity of the optimized PNN to train the model and obtain the evaluation model. Finally, substitute the test set into the trained model, use the CCT margin value as the model output, and evaluate the transient voltage stability. The transient voltage stability evaluation accuracy c is used for evaluation, as follows:

[0093]

[0094] Where: N is the total number of training samples; w is the number of samples evaluated correctly.

[0095] Example:

[0096] The embodiment used in the present invention is based on the IEEE 39-node power system as a simulation model and Psat as a simulation platform. This system uses 100MVA as a power benchmark, 345kV as a voltage benchmark, and 10 generators all use a 4th-order model. 30 1.5MW wind turbines are equivalent to 1 wind turbine connected to the power grid. The voltage level of the wind turbine is 0.69kV, and it is connected to the system after being stepped up by a step-up transformer. The system is transiently analyzed by power flow calculation and time domain simulation.

[0097] Through simulation analysis, a total of 2080 samples were collected, including 1380 stable samples and 700 unstable samples. 90% of the samples were randomly selected as training sets, and the remaining 10% were used as test sets. The CCT margin value was used to classify the samples according to the method described above, and the reduced-dimensional training set was used as the input of the PNN model, and the CCT margin value was used as the final output of the model. After the network training was completed, the test set was input into the trained test PNN network, and the model was evaluated using the transient voltage stability assessment accuracy c.

[0098] Finally, the experiment showed that the accuracy of transient voltage stability assessment of wind power grid-connected system based on PNN was as high as 99.26%, indicating that the method proposed in the present invention has strong assessment ability.

[0099] 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. Ordinary technicians in the relevant field can still modify or make equivalent substitutions to the specific implementation methods of the present invention with reference to the above embodiments. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention are within the scope of protection of the claims of the present invention to be approved.

Claims

1. A transient voltage stability assessment method based on probabilistic neural network, characterized in that: The following steps are involved: Step 1: Based on the historical operation data of the power system and the characteristics of wind power after grid connection, obtain system operation data samples, classify them according to the size of the critical fault removal time margin value, and form a corresponding initial sample set; Step 2: For the initial sample set, a t-distributed random neighbor embedding algorithm based on fusion features is used to reduce the dimension of the feature quantity; Step 3: Use the genetic ant colony hybrid algorithm to optimize the parameters of the probabilistic neural network; Step 4: Use the training set as the input feature quantity of the optimized PNN to train the model and obtain the evaluation model. Finally, substitute the test set into the trained model, use the CCT margin value as the model output, and evaluate the transient voltage stability.

2. The transient voltage stability assessment method based on probabilistic neural network according to claim 1 is characterized in that: In step 1, based on the historical operation data of the power system and the characteristics of wind power after grid connection, the system operation data samples are obtained and classified according to the size of the CCT margin value to form a corresponding initial sample set; wherein, the calculation formula of the CCT margin value is as follows: T s =T cr -T cct (1) Where: T s is the CCT margin value; T cr is the critical fault clearing time; T cct is the fault clearing time; In power system fault analysis, if the margin value of CCT is positive, that is, the fault is successfully removed before approaching the critical edge of system stability, such samples are classified as transient stability categories; on the contrary, when the margin value turns negative, it means that the fault removal operation occurs after the critical edge of system stability, and such samples are regarded as transient instability categories.

3. The transient voltage stability assessment method based on probabilistic neural network according to claim 1, characterized in that: In step 2, the t-SNE algorithm is used to reduce the dimension of the feature quantity for the initial sample set. The specific steps of the t-SNE algorithm are as follows: Step 2-1: Assume the original data is A = {a1, a2, ..., a n }, the data after dimensionality reduction is B = {b1, b2, ..., b n }; Step 2-2: Initialize the number of iterations, learning rate and momentum, and calculate the number of a in high-dimensional space i As the center, a j is a neighboring point, with a probability distribution of Gaussian distribution P j|i and with a j Centered on i is the neighboring point, according to the probability density P of Gaussian distribution i|j , as follows: Where: a k ∈{a1,a2,…,a n }; σi and σ j They correspond to data points ai and a respectively j Gaussian distribution variance; Step 2-3: Calculate the joint probability distribution P ij And normalize it: Step 2-4: Initialize the reduced-dimensional data B (0) , and calculate b in the low-dimensional space of the t distribution i With b j The probability density Q ij , as follows: B (0) ={b1,b2,…,b n } (5) Where: b k ∈{b1,b2,…,b n }; Step 2-5: H is the loss function in the dimensionality reduction process obtained by Kullback-Leibler divergence KL; the low-dimensional distribution joint probability density Q ij Joint probability density P with high-dimensional distribution ij Fitting, divergence KL is the relative entropy that measures the degree of fit of the joint probability density of two distributions, as follows: Get the gradient By decreasing the gradient of H, we can learn the appropriate b i , calculate the optimized gradient Step 2-6: Introduce the learning rate γ and momentum factor α to control the learning accuracy, iterate t times to perform gradient descent, and obtain the low-dimensional data B (t) , the gradient update formula is as follows: Step 2-7: The learning rate γ controls the step size of the gradient descent. The momentum factor α represents the influence of the direction and size of the previous step's gradient descent on the direction and size of the current gradient descent. The values ​​of the two can be adjusted according to the characteristics of the data set and the complexity of the training model. If the current gradient direction is consistent with the previous step's gradient direction, the weight of this step is increased; if not, the weight is reduced.

4. The transient voltage stability assessment method based on probabilistic neural network according to claim 1, characterized in that: In step 3, the genetic ant colony hybrid algorithm is used to optimize the parameters of the PNN. In order to make full use of the global search capability of the genetic algorithm and the local search and pheromone mechanism of the ant colony algorithm to improve the initialization, parameter adjustment and performance optimization of the PNN, the specific steps are as follows: Step 3-1: Encode the weight, threshold or other key parameters of the PNN to form the initial population of the genetic algorithm; these codes can be binary codes, real number codes or other suitable coding methods; at the same time, initialize the relevant parameters of the ant colony algorithm, such as the initial pheromone concentration, the number of ants, and the pheromone evaporation rate; Step 3-2: Define a fitness function to evaluate the performance of PNN on a given data set. This function will be used for individual selection in the genetic algorithm and path selection in the ant colony algorithm. Then, according to the fitness function, select individuals with higher fitness from the current population as the parents of the next generation, and perform crossover and mutation operations. Step 3-3: Generate new individuals through crossover operation and introduce new genetic information through mutation operation to increase the diversity of the population; in the ant colony algorithm, update the pheromone concentration according to the fitness of the path; the better path will accumulate more pheromones, thereby attracting more ants to choose the path; Step 3-4: Use the excellent individuals generated by the genetic algorithm (i.e., the PNN parameter combination with higher fitness) as the initial solution or seed solution of the ant colony algorithm to enhance the search ability of the ant colony algorithm; in the crossover and mutation process of the genetic algorithm, the pheromone heuristic rule of the ant colony algorithm can be introduced to guide the selection of crossover points and the direction of mutation, thereby improving the efficiency of the genetic algorithm; Step 3-5: Repeat the above steps to continuously generate new PNN parameter combinations and evaluate them through the fitness function; as the iteration proceeds, the performance of the PNN will gradually improve. When the preset number of iterations is reached, the fitness reaches the preset threshold, or the quality of the solution is no longer significantly improved, the algorithm terminates and outputs the optimal PNN parameter combination.

5. The transient voltage stability assessment method based on probabilistic neural network according to claim 1, characterized in that: In step 4, the training set is used as the input feature quantity of the optimized PNN for model training to obtain an evaluation model. Finally, the test set is substituted into the trained model, the CCT margin value is used as the model output, and the transient voltage stability is evaluated; wherein, the accuracy rate is used to evaluate the transient voltage stability discrimination capability.

Images