Power system transient stability assessment method based on deep bayesian active learning

Abstract

The application relates to the technical field of power system automation, and discloses a power system transient stability evaluation method based on deep Bayesian active learning, which comprises the following steps: S1, in an offline stage, for a target power system, power grid power flow mode data sets are generated by randomly sampling the power grid power flow mode of the target power system; S2, power grid power flow mode sample sets containing target attributes are generated through transient stability simulation analysis; S3, a transient stability evaluation model based on a deep Bayesian neural network is constructed; S4, a minimum credibility-based active learning method is used to obtain samples to be labeled from the power grid power flow mode sample sets; and S5, in a real-time operation stage of the power system, the transient stability of the power system is predicted by using the transient stability evaluation model. The transient stability evaluation model based on the deep Bayesian neural network can give a prediction result to evaluate the transient stability of the power system.

Description

Technical Field

[0001] This invention relates to the field of power system automation technology, specifically to a power system transient stability assessment method based on deep Bayesian active learning. Background Technology

[0002] Online transient stability assessment can identify power flow patterns that do not meet the requirements for stable operation of the power grid in real time, which is an important prerequisite for preventing transient instability of the power system and avoiding large-scale power outages. Traditional transient stability analysis usually adopts a model-driven method with time-domain simulation and criteria, which is accurate and reliable. However, with the increasing scale of interconnected power grids and the integration of complex dynamic components such as new energy generation and DC transmission systems, model-driven methods face difficulties in terms of high computational complexity and long operation time, making it difficult to meet the requirements of real-time stability analysis. Machine learning algorithms, represented by deep neural networks, can establish a mapping relationship between input and output through a large number of training samples, and can be used to achieve rapid prediction of transient stability without relying on model simulation analysis.

[0003] Because transient instability samples in power systems are difficult to obtain, it is usually necessary to generate the samples needed for training deep neural network models through stochastic power system simulation. Although sample generation can be completed offline, there are still problems such as heavy computational load for sample simulation analysis and the potential existence of a large number of redundant samples. Summary of the Invention

[0004] To address the aforementioned issues, this invention proposes a power system transient stability assessment method based on deep Bayesian active learning. This method utilizes a transient stability assessment model based on a deep Bayesian neural network to provide predictions for evaluating the transient stability of the power system. The active learning framework is combined to improve the efficiency of sample selection, reduce the consumption of simulation resources for redundant samples, and enhance the efficiency of offline sample simulation and transient stability assessment model training.

[0005] This invention is achieved through the following technical solution:

[0006] A power system transient stability assessment method based on deep Bayesian active learning includes:

[0007] S1. In the offline phase, for the target power system, the power flow mode of the target power system is randomly sampled to generate a power flow mode dataset;

[0008] S2. The k-medoids algorithm is used to cluster and extract representative power flow mode samples from the power flow mode dataset. Transient stability simulation analysis is then used to generate a power flow mode sample set containing target attributes.

[0009] S3. Construct a transient stability assessment model based on a deep Bayesian neural network, and train the parameters of the transient stability assessment model using sample data from a power grid power flow mode sample set containing target attributes.

[0010] S4. Using an active learning method based on minimum confidence, obtain samples to be labeled from the power grid power flow mode sample set, update the power grid power flow mode sample set, and update the parameters of the trained transient stability evaluation model.

[0011] S5. During the real-time operation phase of the power system, relevant operating parameters of the power system are collected through the SCADA system, and the relevant operating parameters are input into the updated transient stability assessment model. The transient stability assessment model is then used to predict the transient stability of the power system.

[0012] As an optimization, the specific method for generating the power flow mode dataset in step S1 is as follows:

[0013] S1.1 For the target power system, determine the fluctuation range of the load level of the target power system according to the daily load curve. Based on the initial power flow mode of the target power system, adjust the active power output of the generator and the active and reactive power of the load synchronously according to the following formula to generate the power flow mode of the grid under different load levels.

[0014] S1.2 Obtain the partition information of the target power system and the total number of partitions, Narea. Randomly select two partitions, namely the first partition and the second partition. For the first partition, calculate the initial value of the total active power generation of the first partition. Randomly sample within the range of (0, 20%) times the initial value to obtain the growth amount of the total active power generation of the first partition. Distribute the growth amount to each first generator according to the initial active power output ratio of each first generator in the first partition. For the second partition, achieve the active power balance of the entire power system by reducing the active power generation by an equal amount, that is, distribute the reduction amount to each second generator according to the initial active power output ratio of each second generator in the second partition.

[0015] S1.3. Summarize the data obtained in steps S1.1 and S1.2 to obtain the power grid flow mode dataset.

[0016] As an optimization, in S2, the specific method for generating a power flow mode sample set containing the target attributes is as follows:

[0017] Step S2.1: Obtain the transmission sections of each partition of the target power system. For all power flow mode samples in the power flow mode data obtained in S1, calculate the active power transmission volume of each transmission section.

[0018] Step S2.2: Using the active power of all transmission sections as the feature attribute of the power flow mode sample, the k-medoids algorithm is used to perform cluster analysis on the power flow mode to obtain representative power flow mode samples, and representative power flow mode samples of the cluster center are extracted.

[0019] Step S2.3: For representative power flow pattern samples, set a set of anticipated power grid faults, perform transient stability analysis based on time-domain simulation, and label the representative power flow pattern samples with the transient stability of the target power system as the target attribute to generate a transient stability sample set. Meanwhile, the unlabeled representative power flow mode samples constitute the unlabeled sample set. .

[0020] As an optimization, the specific implementation method of S3 is as follows:

[0021] S3.1. Using the active power output of each generator in the target power system as the input feature and the transient stability of the target power system as the output target, construct a transient stability evaluation model based on a deep Bayesian neural network.

[0022] S3.2, Transiently stable sample set generated based on S2 The parameters of the transient stability evaluation model based on deep Bayesian neural network were trained.

[0023] As an optimization, the structure of the deep Bayesian neural network is a multilayer perceptron model.

[0024] As an optimization, the specific method for S4 is as follows:

[0025] S4.1. Use the transient stability assessment model to predict the stability of all unlabeled samples;

[0026] S4.2 Calculate the uncertainty index for all unlabeled samples using the following formula;

[0027] ;

[0028] Step S4.3: Sort all unlabeled samples according to the uncertainty index, and take the top 10% of unlabeled samples to form an unlabeled sample pool;

[0029] Step S4.4: Use the k-medoids algorithm to cluster the samples in the unlabeled sample pool, and combine the clustered samples together to form a sample set to be labeled.

[0030] Step S4.5: For the samples to be labeled in the sample set to be labeled, set a set of anticipated faults of the target power system, perform transient stability analysis based on time-domain simulation, and label the samples to be labeled with the transient stability of the target power system as the target attribute. Add the newly labeled samples to the transient stability sample set. The sample set was never labeled. Delete;

[0031] Step S4.6: Utilize the updated transiently stable sample set The transient stability assessment model is updated and learned by model parameter updating.

[0032] As an optimization

[0033] ;

[0034] .

[0035] As an optimization, the specific method for S5 is as follows:

[0036] S5.1 During the real-time operation phase of the power system, the real-time values ​​of relevant operating parameters of the target power system are obtained through the state estimation module of the SCADA system of the target power system.

[0037] S5.2 Input the relevant operating parameters into the transient stability assessment model, use the transient stability assessment model to quickly predict the transient stability of the current target power system's operating condition, and return the prediction results to the power grid operation and dispatch personnel. If the current target power system's operating condition is judged to be unstable, an alarm signal should also be returned.

[0038] As an optimization, in S5.2, the method for determining whether the current target power system is unstable relies on the deep Bayesian neural network trained and constructed in S4. The specific steps are as follows:

[0039] S5.2.1 Setting the total number of Monte Carlo feedforward operations N MCFF =100, and set the input feature x to the real-time value of the relevant operating parameters of the target power system;

[0040] S5.2.2, Set the loop variable s=1;

[0041] S5.2.3, For random variables According to the standard normal distribution Perform random sampling, and according to Calculate the weight parameters w of the deep Bayesian neural network;

[0042] S5.2.4. Based on the weight parameters w generated by sampling, perform feedforward operation on the deep Bayesian neural network to obtain the prediction result y of transient stable classification;

[0043] S5.2.5 Determine if the loop variable s is greater than N. MCFF If yes, jump to S5.2.6; otherwise, set the loop variable s = s + 1 and return to S5.2.3.

[0044] S5.2.6, Based on N in the aforementioned loop MCFF The result of the Monte Carlo feedforward operation, let N be... MCFF In the Monte Carlo feedforward operation, there are a total of N... stable The next judgment is that the system's operating mode is "transiently stable," and there are N... unstable The system is initially determined to be in a "transient instability" state. Further calculations are made to determine the probabilities of transient stability and transient instability, respectively: N... stable / N MCFF and N unstable / N MCFF ;

[0045] S5.2.7, If N stable / N MCFF >N unstable / N MCFF If N is stable, then the operating condition of the target power system is determined to be "transiently stable"; otherwise, if N is not stable, then the operating condition of the target power system is determined to be "transiently stable". stable / N MCFF <N unstable / N MCFF If so, the operating condition of the target power system is determined to be "transient instability".

[0046] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0047] This invention proposes a power system transient stability assessment method based on deep Bayesian active learning. The transient stability assessment model based on deep Bayesian neural network can provide its prediction results to assess the transient stability of the power system. The active learning framework is combined to improve the efficiency of sample selection, reduce the simulation resource consumption of redundant samples, and improve the efficiency of offline sample simulation and transient stability assessment model training. Attached Figure Description

[0048] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings:

[0049] Figure 1 This is a flowchart of a power system transient stability assessment method based on deep Bayesian active learning. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0051] Example

[0052] This invention is achieved through the following technical solution:

[0053] like Figure 1 As shown, a power system transient stability assessment method based on deep Bayesian active learning includes:

[0054] S1. In the offline phase, for the target power system, the power flow mode of the target power system is randomly sampled to generate a power flow mode dataset.

[0055] In this embodiment, the specific method for generating the power grid flow mode dataset in step S1 is as follows:

[0056] S1.1 For the target power system, determine the fluctuation range of the load level of the target power system according to the daily load curve. Based on the initial power flow mode of the target power system, adjust the active power output of the generator and the active and reactive power of the load synchronously according to the following formula to generate the power flow mode of the grid under different load levels.

[0057] ;

[0058] ;

[0059] ;

[0060] In the formula, , and These represent the initial values ​​of generator active power output and load active and reactive power for the initial power flow mode of the target power system. Represents the set of generators in the target power system. Represents the load set. The proportionality coefficient for the load level. , and This refers to the adjusted generator active power output and the load active and reactive power.

[0061] S1.2. Based on the zoning information of the target power system, two zoning areas are randomly selected, namely, Zone 1 and Zone 2. For Zone 1, the initial value of the total active power generation in Zone 1 is calculated. Random sampling is performed within the range of (0, 20%) times the initial value to obtain the total active power generation growth of Zone 1. The total active power generation growth is then allocated to each of the first generators in Zone 1 according to the proportion of their initial active power output (the initial active power output of the generator is the aforementioned). Determine which first generators are connected within the first zone of the target power system, and then calculate the [data / efficiencies] of these first generators. (The proportion); for the second zone, the active power balance of the entire power system is achieved by reducing an equal amount of active power generation, that is, the reduction is distributed to each second generator according to the initial active power output ratio of each second generator in the second zone. Similarly, Then calculate these second generators The proportion.

[0062] Power grid zoning refers to dividing a power grid into several sub-regional power grids that are "closely interconnected within each region and weakly interconnected between regions" based on the power grid's structural characteristics. Power grid zoning itself is a research topic, but this invention does not involve research on methods for power grid zoning.

[0063] For a target power grid (target power system), its grid divisions (i.e., the total number of divisions and which buses each division consists of) are generally known in advance by the power grid dispatch and operation personnel.

[0064] S1.3. Summarize the data obtained in steps S1.1 and S1.2 to obtain the power grid flow mode dataset.

[0065] S2. The k-medoids algorithm is used to cluster and extract representative power flow mode samples from the power flow mode dataset. Transient stability simulation analysis is then used to generate a power flow mode sample set containing target attributes.

[0066] In this embodiment, the specific method for generating the power flow mode sample set containing the target attributes in S2 is as follows:

[0067] Step S2.1: Obtain the transmission sections of each partition of the target power system. For all power flow mode samples in the power flow mode data obtained in S1, calculate the active power transmission volume of each transmission section.

[0068] Step S2.2: Using the active power of all transmission sections as the feature attribute of the power flow mode sample, the k-medoids algorithm is used to perform cluster analysis on the power flow mode to obtain representative power flow mode samples, and representative power flow mode samples of the cluster center are extracted.

[0069] Based on the k-medoids algorithm, the power flow mode samples (power flow mode dataset) generated by S1 are classified into several classes. The power flow mode samples within a class have similarity. After cluster analysis, this invention uses the power flow mode samples corresponding to the cluster centers as the "representative power flow mode samples" of each cluster.

[0070] Step S2.3: For representative power flow pattern samples, set a set of anticipated power grid faults, perform transient stability analysis based on time-domain simulation, and label the representative power flow pattern samples with the transient stability of the target power system as the target attribute to generate a transient stability sample set. Meanwhile, the unlabeled representative power flow mode samples constitute the unlabeled sample set. .

[0071] Time-domain simulation is an existing and commonly used method, so it will not be elaborated on here.

[0072] Annotation method:

[0073] For a representative power grid flow pattern sample (1 sample), a set of anticipated power grid faults is set (e.g., a three-phase short-circuit fault occurs at the head end of a 500kV transmission line; assuming the target power system has N faults). C (One anticipated fault). For each anticipated fault, time-domain simulation calculations are performed using the Power System Integrated Stability Program (PSASP). The PSASP program outputs generator power angle trajectory data of the target power system after the anticipated fault occurs under a representative grid power flow mode. The Transient Stability Index (TSI) is calculated using the following formula:

[0074] ;

[0075] ;

[0076] In the formula, The maximum unit power angle difference during the time-domain simulation is given. t represents the time variable; it is assumed that the anticipated fault occurs at the initial time (0 seconds), and the final time calculated in the time-domain simulation is T, which is taken as T = 20 seconds. Let be the set of generators in the target power system, where δ represents the generator power angle, and and These refer to the power angles of the i-th and j-th generators at time t, respectively.

[0077] When TSI > 0, the target power system can maintain transient stability under a given anticipated fault; conversely, when TSI < 0, the target power system will experience transient instability under a given anticipated fault.

[0078] Traverse N C For each anticipated fault, the transient stability is assessed using the aforementioned TSI index. When the target power system experiences all N... C When the target power system remains transiently stable under all anticipated faults, the target attribute of the corresponding representative power flow mode sample is "transient stability" and it is recorded. Conversely, when the target power system experiences transient instability under any anticipated fault, the target attribute of the corresponding representative power flow mode sample is "transient instability" and it is recorded.

[0079] S3. Construct a transient stability assessment model based on a deep Bayesian neural network, and train the parameters of the transient stability assessment model using sample data from a power flow mode sample set containing target attributes.

[0080] In this embodiment, the specific implementation method of S3 is as follows:

[0081] S3.1. Using the active power output of each generator in the target power system as the input feature and the transient stability of the target power system as the output target, construct a transient stability assessment model based on a deep Bayesian neural network; here, the transient stability assessment model refers to the aforementioned "deep Bayesian neural network", and it is emphasized that this deep Bayesian neural network is used for transient stability assessment.

[0082] S3.2, Transiently stable sample set generated based on S2 The parameters of the transient stability evaluation model based on deep Bayesian neural network were trained.

[0083] In this embodiment, the Deep Bayesian Neural Network (DBNN) is structured as a Multi-layer Perception (MLP) model.

[0084] The parameters of a model based on a deep Bayesian neural network follow a certain probability distribution model and are not unique values ​​like those of a conventional neural network. Therefore, unlike the feedforward computation of a neural network, the mathematical expression of Bayesian neural network prediction is as shown in the following formula, and the prediction result needs to be obtained by Monte Carlo sampling and solving for the output expectation.

[0085] ;

[0086] Since the posterior distribution of deep Bayesian neural network parameters is difficult to obtain directly, it is necessary to combine variational inference and the Bayes-by-BackProp algorithm. This involves minimizing the Kullback-Leibler divergence between the posterior and variational distributions to obtain the variational distribution of the network parameters, and then using the variational distribution to approximate the posterior distribution. The mathematical expression for minimizing the Kullback-Leibler divergence is:

[0087] ;

[0088] The variational distribution is generally a normal distribution, where θ represents the variational distribution parameters to be determined (i.e., the mean and variance of the normal distribution). Reparameterization is used to train and optimize the variational distribution parameters. The variational posterior distribution of the neural network weight parameters is assumed to be an independent Gaussian distribution, and each weight parameter can be obtained through random sampling based on the mean μ and standard deviation σ of this variational posterior Gaussian distribution. To ensure that the standard deviation is non-negative, the standard deviation is further parameterized, setting...

[0089] ;

[0090] Therefore, the parameters of the variational posterior distribution are... The sampling of network weight parameters can be obtained by the following formula:

[0091] ;

[0092] In the formula, random variable It follows a standard normal distribution, that is .

[0093] By applying variational inference and reparameterization techniques, the algorithm flow for parameter training at each step based on the Bayes by Backprop algorithm is as follows:

[0094] (1) Random variables according to the standard normal distribution Perform sampling;

[0095] (2) Calculate the network weight parameter w using the following formula;

[0096] ;

[0097] (3) Estimate the prediction loss for the current batch of training data using the following formula;

[0098] ;

[0099] (4) Calculate the gradient of the mean using the following formula:

[0100] ;

[0101] (5) Calculate the gradient of the standard deviation parameter using the following formula:

[0102] ;

[0103] (6) Update the parameters of the variational posterior distribution using the following formula:

[0104] ;

[0105] Steps 4-6 of the above process can be completed by calling optimizers such as SGD and ADAM from existing machine learning libraries such as PyTorch and TensorFlow. Repeat the above process until all training data in each batch has been traversed, completing parameter learning for the current epoch; then repeat each epoch until the pre-set total number of iterations for model training is reached, ultimately obtaining a transient stability assessment model based on a deep Bayesian neural network for power system transient stability evaluation.

[0106] w represents the weight parameters of the transient stability evaluation model (i.e., the deep Bayesian neural network model), and the weight parameters w are given by the formula... The calculation yielded that, and The variational parameters of the deep Bayesian neural network model need to be determined through the aforementioned "parameter training based on the Bayes by Backprop algorithm". are random variables and . For a given sample set The predicted loss value, For the set of variational parameters, i.e. . and express For variational parameters and The gradient. This is the learning rate for optimizers such as SGD and ADAM, and can be the default setting for machine learning libraries such as PyTorch and TensorFlow.

[0107] S4. Using an active learning method based on minimum confidence, obtain samples to be labeled from the power grid power flow mode sample set, update the power grid power flow mode sample set, and update the parameters of the trained transient stability evaluation model.

[0108] In this embodiment, the specific method of S4 is as follows:

[0109] S4.1. Use the transient stability assessment model to predict the stability of all unlabeled samples;

[0110] S4.2 Calculate the uncertainty index for all unlabeled samples using the following formula;

[0111] ;

[0112] In the formula, Represents the transiently stable sample set. y represents the input features of the unlabeled sample, and y represents the target attribute of the unlabeled sample (i.e., the binary classification target of "transient stability" or "transient instability"). These represent the weight parameters of a deep Bayesian neural network.

[0113] ;

[0114] Let the given transiently stable training sample set be Given that the input sample to be analyzed is x, the predicted result y is the posterior probability of category c.

[0115] ;

[0116] Represents the weight parameters given a transiently stable sample set. The posterior probability, That is, based on the weight parameters posterior probability Calculated expected value, This is an indicator of uncertainty.

[0117] Step S4.3: Sort all unlabeled samples according to the uncertainty index, and take the top 10% of unlabeled samples (from largest to smallest) to form the unlabeled sample pool;

[0118] Step S4.4: Use the k-medoids algorithm to cluster the samples in the unlabeled sample pool, and combine the clustered samples together to form a sample set to be labeled.

[0119] Step S4.5: For the samples to be labeled in the sample set, set a set of anticipated faults for the target power system (e.g., a three-phase short-circuit fault occurring at the head end of a 500kV transmission line), perform transient stability analysis based on time-domain simulation, and label the samples to be labeled using the transient stability of the target power system as the target attribute. Add the newly labeled samples to the transient stability sample set. The sample set was never labeled. Delete;

[0120] Step S4.6: Utilize the updated transiently stable sample set The transient stability assessment model is updated and learned by model parameter updating.

[0121] S5. During the real-time operation phase of the power system, relevant operating parameters of the power system are collected through the SCADA system, and the relevant operating parameters are input into the updated transient stability assessment model. The transient stability assessment model is then used to predict the transient stability of the power system.

[0122] In this embodiment, the specific method of S5 is as follows:

[0123] S5.1 During the real-time operation phase of the power system, the real-time values ​​of relevant operating parameters of the target power system are obtained through the state estimation module of the SCADA system of the target power system. The relevant operating parameters here include generator active power output, active power flow and reactive power flow of transmission lines, node voltage amplitude and phase.

[0124] S5.2 Input the relevant operating parameters into the transient stability assessment model, use the transient stability assessment model to quickly predict the transient stability of the current target power system's operating condition, and return the prediction results (the prediction results are the binary prediction results of the target power system's "transient stability" or "transient instability") to the power grid operation dispatch personnel. If the current target power system's operating condition is judged to be unstable, an alarm signal should also be returned.

[0125] In this invention, the method for determining whether the current operating condition of the target power system is unstable relies on a deep Bayesian neural network trained and constructed using S4. The specific steps are as follows:

[0126] S5.2.1 Setting the total number of Monte Carlo feedforward operations N MCFF =100, and set the input feature x to the real-time value of the relevant operating parameters of the target power system;

[0127] S5.2.2, Set the loop variable s=1;

[0128] S5.2.3, For random variables According to the standard normal distribution Perform random sampling, and according to Calculate the weight parameters w of the deep Bayesian neural network;

[0129] S5.2.4. Based on the weight parameters w generated by sampling, perform feedforward operation on the deep Bayesian neural network to obtain the prediction result y of transient stable classification;

[0130] S5.2.5 Determine if the loop variable s is greater than N. MCFF If yes, jump to S5.2.6; otherwise, set the loop variable s = s + 1 and return to S5.2.3.

[0131] S5.2.6, Based on N in the aforementioned loopMCFF The result of the Monte Carlo feedforward operation, let N be... MCFF In the Monte Carlo feedforward operation, there are a total of N... stable The next judgment is that the system's operating mode is "transiently stable," and there are N... unstable The system is initially determined to be in a "transient instability" state. Further calculations are made to determine the probabilities of transient stability and transient instability, respectively: N... stable / N MCFF and N unstable / N MCFF ;

[0132] S5.2.7, If N stable / N MCFF >N unstable / N MCFF If N is stable, then the operating condition of the target power system is determined to be "transiently stable"; otherwise, if N is not stable, then the operating condition of the target power system is determined to be "transiently stable". stable / N MCFF <N unstable / N MCFF If so, the operating condition of the target power system is determined to be "transient instability".

[0133] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for evaluating the transient stability of a power system based on deep Bayesian active learning, characterized in that, include: S1. In the offline phase, for the target power system, the power flow mode of the target power system is randomly sampled to generate a power flow mode dataset; S2. The k-medoids algorithm is used to cluster and extract representative power flow mode samples from the power flow mode dataset. Transient stability simulation analysis is then used to generate a power flow mode sample set containing target attributes. S3. Construct a transient stability assessment model based on a deep Bayesian neural network, and train the parameters of the transient stability assessment model using sample data from a power grid power flow mode sample set containing target attributes. S4. Using an active learning method based on minimum confidence, obtain samples to be labeled from the power grid power flow mode sample set, update the power grid power flow mode sample set, and update the parameters of the trained transient stability evaluation model. The specific method for S4 is as follows: S4.

1. Use the transient stability assessment model to predict the stability of all unlabeled samples; S4.2 Calculate the uncertainty index for all unlabeled samples using the following formula; ； In the formula, Represents the transiently stable sample set. represents the input features of the unlabeled samples, and y represents the target attribute of the unlabeled samples. These represent the weight parameters of a deep Bayesian neural network. ； Let the given transiently stable training sample set be Given that the input sample to be analyzed is x, the predicted result y is the posterior probability of category c. ； Represents the weight parameters given a transiently stable sample set. The posterior probability, That is, based on the weight parameters posterior probability Calculated expected value, This refers to an uncertainty indicator; Represents the transiently stable sample set. represents the input features of the unlabeled samples, and y represents the target attribute of the unlabeled samples. This represents the weight parameters of the deep Bayesian neural network; uppercase C represents the set of categories for the classification problem, and lowercase c represents the c-th category. Let the given transiently stable training sample set be Given that the input sample to be analyzed is x, the predicted result y is the posterior probability of category c. Represents a given transiently stable sample set Below, weight parameters The posterior probability, Indicates considering probability distribution Expected value; Step S4.3: Sort all unlabeled samples according to the uncertainty index, and take the top 10% of unlabeled samples to form an unlabeled sample pool; Step S4.4: Use the k-medoids algorithm to cluster the samples in the unlabeled sample pool, and combine the clustered samples together to form a sample set to be labeled. Step S4.5: For the samples to be labeled in the sample set to be labeled, set a set of anticipated faults of the target power system, perform transient stability analysis based on time-domain simulation, and label the samples to be labeled with the transient stability of the target power system as the target attribute. Add the newly labeled samples to the transient stability sample set. The sample set was never labeled. Delete; Step S4.6: Utilize the updated transiently stable sample set The transient stability assessment model is updated by learning the model parameters. S5. During the real-time operation phase of the power system, relevant operating parameters of the power system are collected through the SCADA system, and the relevant operating parameters are input into the updated transient stability assessment model. The transient stability assessment model is then used to predict the transient stability of the power system.

2. The power system transient stability assessment method based on deep Bayesian active learning according to claim 1, characterized in that, In step S1, the specific method for generating the power grid flow mode dataset is as follows: S1.1 For the target power system, determine the fluctuation range of the load level of the target power system according to the daily load curve. Based on the initial power flow mode of the target power system, adjust the active power output of the generator and the active and reactive power of the load synchronously according to the following formula to generate the power flow mode of the grid under different load levels. ； ； ； In the formula, , and These represent the initial values ​​of generator active power output and load active and reactive power for the initial power flow mode of the target power system. Represents the set of generators in the target power system. Represents the load set. The proportionality coefficient for the load level. , and This refers to the adjusted generator active power output and the load active and reactive power. S1.

2. Based on the zoning information of the target power system, two zoning areas are randomly selected, namely the first zoning area and the second zoning area. For the first zoning area, the initial value of the total active power generation of the first zoning area is calculated. Random sampling is performed within the range of (0, 20%) times the initial value to obtain the growth amount of the total active power generation of the first zoning area. The growth amount is allocated to each first generator according to the initial active power output ratio of each first generator in the first zoning area. For the second zoning area, the active power balance of the entire power system is achieved by reducing an equal amount of active power generation, that is, the reduction amount is allocated to each second generator according to the initial active power output ratio of each second generator in the second zoning area. S1.

3. Summarize the data obtained in steps S1.1 and S1.2 to obtain the power grid flow mode dataset.

3. The power system transient stability assessment method based on deep Bayesian active learning according to claim 1, characterized in that, In S2, the specific method for generating a power flow mode sample set containing target attributes is as follows: Step S2.1: Obtain the transmission sections of each partition of the target power system. For all power flow mode samples in the power flow mode data obtained in S1, calculate the active power transmission volume of each transmission section. Step S2.2: Using the active power of all transmission sections as the feature attribute of the power flow mode sample, the k-medoids algorithm is used to perform cluster analysis on the power flow mode to obtain representative power flow mode samples, and the representative power flow mode samples of the cluster center are extracted. Step S2.3: For representative power flow pattern samples, set a set of anticipated power grid faults, perform transient stability analysis based on time-domain simulation, and label the representative power flow pattern samples with the transient stability of the target power system as the target attribute to generate a transient stability sample set. Meanwhile, the unlabeled representative power flow mode samples constitute the unlabeled sample set. .

4. The power system transient stability assessment method based on deep Bayesian active learning according to claim 1, characterized in that, The specific implementation method of S3 is as follows: S3.

1. Using the active power output of each generator in the target power system as the input feature and the transient stability of the target power system as the output target, construct a transient stability evaluation model based on a deep Bayesian neural network. S3.2, Transiently stable sample set generated based on S2 The parameters of the transient stability evaluation model based on deep Bayesian neural network were trained.

5. A power system transient stability assessment method based on deep Bayesian active learning according to claim 1 or 4, characterized in that, The deep Bayesian neural network is structured as a multilayer perceptron model.

6. The power system transient stability assessment method based on deep Bayesian active learning according to claim 1, characterized in that, The specific method for S5 is as follows: S5.1 During the real-time operation phase of the power system, the real-time values ​​of relevant operating parameters of the target power system are obtained through the state estimation module of the SCADA system of the target power system. S5.2 Input the relevant operating parameters into the transient stability assessment model, use the transient stability assessment model to quickly predict the transient stability of the current target power system's operating condition, and return the prediction results to the power grid operation and dispatch personnel. If the current target power system's operating condition is judged to be unstable, an alarm signal should also be returned.

7. The power system transient stability assessment method based on deep Bayesian active learning according to claim 6, characterized in that, The relevant operating parameters of the target power system include generator active power output, active and reactive power flow of transmission lines, node voltage amplitude and phase.

8. The power system transient stability assessment method based on deep Bayesian active learning according to claim 7, characterized in that, In S5.2, the method for determining whether the current operating condition of the target power system is unstable relies on the deep Bayesian neural network trained and constructed in S4. The specific steps are as follows: S5.2.1 Setting the total number of Monte Carlo feedforward operations N MCFF =100, and set the input feature x to the real-time value of the relevant operating parameters of the target power system; S5.2.2, Set the loop variable s=1; S5.2.3, For random variables According to the standard normal distribution Perform random sampling, and according to Calculate the weight parameters w of the deep Bayesian neural network; S5.2.

4. Based on the weight parameters w generated by sampling, perform feedforward operation on the deep Bayesian neural network to obtain the prediction result y of transient stable classification; S5.2.5 Determine if the loop variable s is greater than N. MCFF If yes, jump to S5.2.6; otherwise, set the loop variable s = s + 1 and return to S5.2.

3. S5.2.6, Based on N in the aforementioned loop MCFF Let N be the result of the Monte Carlo feedforward operation. MCFF In the Monte Carlo feedforward operation, there are a total of N stable The next judgment is that the system's operating mode is "transiently stable," and there are N... unstable The system is initially determined to be in a "transient instability" state. Further calculations are made to determine the probabilities of transient stability and transient instability, respectively: N... stable / N MCFF and N unstable / N MCFF ; S5.2.7, If N stable / N MCFF >N unstable / N MCFF If N is true, then the operating condition of the target power system is determined to be "transiently stable"; otherwise, if N is false, the operating condition is determined to be "transiently stable". stable / N MCFF <N unstable / N MCFF If so, the operating condition of the target power system is determined to be "transient instability".

Images