Transient voltage stability evaluation method based on unbalanced residual information learning

Abstract

The invention discloses a transient voltage stability evaluation method based on unbalanced residual information learning. The method comprises the following steps: S1, building a target power system simulation model corresponding to a target power system; s2, performing transient voltage time-domain simulation on the target power system simulation model, and constructing a transient time-domain simulation sample set; s3, carrying out sample category weight calculation on the transient time domain simulation sample set, and constructing a Focal-Loss loss function used for processing sample category imbalance; s4, constructing a residual space-time diagram convolutional neural network model; s5, training the residual space-time diagram convolutional neural network model until a preset performance evaluation index threshold of space-time feature learning is met, and obtaining a trained residual space-time diagram convolutional neural network model; and S6, power system transient voltage stability evaluation is carried out through the trained residual space-time diagram convolutional neural network model.

Description

Technical Field

[0001] The present invention relates to the technical field of power systems, and in particular to a transient voltage stability assessment method based on unbalanced residual information learning. Background Art

[0002] With the large-scale connection of new energy to the grid via direct current, its power backbone network will form a complex hybrid pattern of multiple conventional / flexible direct current circuits. Taking the above factors into consideration, these trends will make the network structure and operating characteristics of traditional power systems increasingly complex. When the system is subjected to large transient disturbances, the rapid response characteristics and complex dynamic interactions of DC converter stations and a large number of dynamic loads can easily cause the problem of insufficient transient voltage support. In large-scale power grids with high-proportion penetration of new energy, large-scale complex hybrid of AC and DC, and massive heterogeneous loads, if the transient voltage stability cannot be reliably sensed in real time and control measures cannot be taken in time, it will easily cause voltage instability or even collapse accidents, posing a serious threat to the safe absorption of new energy and the stable supply of electricity.

[0003] Due to the high-dimensional, time-varying and strong nonlinear characteristics of the transient voltage stability problem in actual power systems, the research on its evaluation methods has always been a hot topic. In recent years, many domestic and foreign scholars have proposed transient voltage stability evaluation methods for power grids from the perspectives of model-driven or data-driven. Among them, the model-driven method has the complexity of model construction and large amount of calculation, which makes it difficult to adapt to the real-time evaluation needs of complex power grid structures and variable operating conditions. Although the existing data-driven methods can process large-scale data, most of them only rely on single time series data or isolated time series information of different monitoring nodes, lacking the mining of spatial and temporal coupling correlations in the time series evolution process of each monitoring node in the power grid after the fault occurs, making it difficult to fully capture the intrinsic spatiotemporal characteristics of transient voltage instability in the power grid. In addition, the probability of instability in actual power systems is very small, resulting in the number of stable samples being greater than the number of unstable samples. If this type of severely unbalanced data is put into the data-driven method for learning, it will inevitably lead to insufficient attention to the unstable samples in the evaluation, thereby affecting the transient voltage stability evaluation performance of the data-driven method. Summary of the invention

[0004] In view of this, the present invention provides a transient voltage stability assessment method based on unbalanced residual information learning, which is used to at least solve the problem of limitations of transient voltage stability category imbalance and spatiotemporal feature mining technology in the prior art when evaluating transient stability of power systems.

[0005] In order to achieve the above object, the present invention adopts the following technical solution: A transient voltage stability assessment method based on unbalanced residual information learning comprises the following steps: S1. Build a target power system simulation model corresponding to the target power system; S2. Perform transient voltage time domain simulation on the target power system simulation model to obtain M The time-domain simulation results are obtained, and the time-space graph model corresponding to the target power system simulation model under each simulation condition is constructed according to the time-space graph construction method. G , while targeting M The time domain simulation results are respectively used to analyze the system state S Perform voltage stability or instability calibration based on M Calibration results and spatiotemporal graph models G Construct a transient time domain simulation sample set, denoted as ; S3. Construct the Focal-Loss loss function by calculating the weight factors of the stable and unstable categories of the transient time domain simulation sample set and set the adjustment factor for sample classification , we get the Focal-Loss loss function for training the residual spatiotemporal graph convolutional neural network model to learn the imbalanced sample categories; S4. Construct a residual spatiotemporal graph convolutional neural network model, and use the residual spatiotemporal graph convolutional neural network model to learn the spatiotemporal features of the transient time domain simulation sample set to obtain the prediction result of the transient voltage state; S5. Determine whether the prediction result is correct according to the actual transient voltage state result of the target power system. If it is not correct, perform the next round of training on the residual spatiotemporal graph convolutional neural network model until the preset performance evaluation index threshold of spatiotemporal feature learning is met, and obtain the trained residual spatiotemporal graph convolutional neural network model; S6. The transient voltage stability of the power system is evaluated by using the trained residual spatiotemporal graph convolutional neural network model.

[0006] Preferably, the specific contents of the space-time graph construction method in S2 include: According to the construction method of space-time graph, construct the k The observation time window length of a time domain simulation sample is The graphical model corresponding to the physical topology of the target power system G , and obtain the final space-time graph model ; The first The monitoring nodes are represented as vertices in the graph model. , recorded as X and , N is the number of monitoring nodes, The monitoring node and The connection relationship between the monitoring nodes is used as the edge in the graph model , collect all It is represented as a topological adjacency matrix, denoted as A and , among which k The length of the sample observation time window is The graphical model corresponding to the physical topology of the target power system ; At any monitoring node Get the node time series data on the top, if there is a connection, record it as , otherwise it is recorded as ; Finally, the final spatiotemporal graph model is constructed To characterize the dynamic characteristics of the transient response data of the target power system in the time and space dimensions.

[0007] Preferably, the specific content of S2 includes: The transient voltage time domain simulation is performed on the target power system simulation model, and the voltage stability of the power grid under various typical operating conditions is simulated by combining the system historical operation mode and fault settings. The total number is The time domain simulation results of the times; After each time domain simulation, record the physical adjacency matrix of the target power system And the length of the observation time window on the corresponding monitoring node after the fault occurs is The temporal evolution characteristics of ,include The voltage amplitude under the observation time window length , Active Power and reactive power , and the system status is analyzed based on the transient voltage stability engineering criterion used in the actual power system. Calibration, where Indicates that the power system state result of this time domain transient simulation is voltage instability. It indicates that the result of the simulated power system state is voltage stability; based on M Calibration results and spatiotemporal graph models G Construct the input sample set as ,in, .

[0008] Preferably, M It is the product of the typical operating level number of generators, new energy output and load consumption, the proportion of transient operating dynamic load and the typical fault disturbance number.

[0009] Preferably, the Focal-Loss loss function in S3 specifically includes: calculating the weight factors of the stable and unstable categories of the transient time domain simulation sample set and set the adjustment factor for sample classification ; Category weight factor Set according to the ratio of positive samples to negative samples: The unstable samples of the transient time-domain simulation sample set are taken as positive samples, and the stable samples of the transient time-domain simulation sample set are taken as negative samples; Assume that the number of positive samples is c , the number of negative samples is d The total number of samples is M = c + d , then the weight ratio of unstable samples is , for the weight ratio of stable samples ; Adjustment Factor Used to control the degree of loss weight reduction for sample classification, while enhancing the residual spatiotemporal graph convolutional network model's attention to difficult-to-evaluate samples: ; Based on category weight factor and adjustment factors The Focal-Loss loss function is: ; in, is the predicted probability of the residual spatiotemporal graph convolutional network model for the sample. For unstable samples , for a stable sample .

[0010] Preferably, the residual spatiotemporal graph convolutional neural network model in S4 specifically includes: a residual spatiotemporal graph convolution module and a fully connected classification module; The residual spatiotemporal graph convolution module includes a spatial graph convolution layer, a temporal convolution layer, and a residual convolution layer; The spatial graph convolution layer is used to obtain any spatiotemporal graph input sample in the input sample set and perform topological adjacency matrix analysis on the power system. Perform spatial feature mining and realize the feature aggregation of neighbor node time series information to the central node on the power system topology map, and output spatial features ; Temporal convolution layer, used to further Perform convolution operation to extract the spatiotemporal features of transient voltage response data and output the timing features ; Residual convolution layer is used to Perform convolution operation and output residual features And keep the output features of the spatiotemporal graph Dimensionality is consistent; Fully connected classification module for residual features and spatiotemporal characteristics The fusion is performed and the dimensionality of the fused features is reduced through a linear layer to further extract high-dimensional features for classifying transient voltage stability and instability. The high-dimensional features are converted into probability distributions through a softmax activation function to complete the mapping of the short-term transient voltage stability state space and realize the transient voltage stability or instability assessment of the system.

[0011] Preferably, the spatial graph convolutional layer outputs spatial features The specific contents include: The adjacency matrix of the target power system topology Degree Matrix Conversion, the conversion method is: ; In the formula, is the degree matrix Line Elements of a column; Based on the identity matrix Sum degree matrix , and further use the Laplace graph transformation formula to transform the power system topology matrix Perform spatial feature mining, where the Laplace graph transformation formula is: ; in, is the identity matrix; degree matrix It is used to represent the degree of each vertex in the graph model. The degree is the number of connections between each node and other nodes on the power grid adjacency matrix. Perform spatial feature mining based on the solution formula Get the target power system topology adjacency matrix The orthonormal eigenvectors of and the spatial eigenvalue matrix , the graph node information aggregation formula is used to realize the feature aggregation of the neighbor node time series information on the power system topology graph to the central node, where the graph node information aggregation formula is: ; In the formula, is the graph convolution kernel, is the length of the time window for observing nodes in the graph matrix The temporal evolution characteristics of .

[0012] Preferably, the temporal convolutional layer outputs temporal features The specific contents include: ; Among them, ReLU is a nonlinear activation function. is the convolution kernel parameter, It is the convolution operation.

[0013] Preferably, the fully connected classification module reduces the dimension of the fused features through a linear layer, further extracts high-dimensional features for classifying transient voltage stability and instability, and converts the high-dimensional features into probability distribution through the fully connected classification formula. , the mapping of the short-term transient voltage stability state space is completed, and the transient voltage stability state evaluation of the system is realized. The full connection classification formula is: .

[0014] Preferably, the specific content of S5 includes: The accuracy and misclassification rate are selected as the performance evaluation indicators of the residual spatiotemporal graph convolutional neural network model for spatiotemporal feature learning. The accuracy rate Acc is used to reflect the overall performance of the residual spatiotemporal graph convolutional neural network model, and the misclassification rate MAR reflects the probability that an unstable sample is misclassified as stable by the residual spatiotemporal graph convolutional neural network model. The higher the accuracy rate and the lower the misclassification rate, the better the evaluation performance of the residual spatiotemporal graph convolutional neural network model. The total number of correct transient time-domain simulation samples evaluated by the residual spatiotemporal graph convolutional neural network model in this round of training is , the total number of unstable samples is evaluated as unstable And the total number of unstable samples evaluated as stable , and calculate the transient stability assessment accuracy and misjudgment rate of the residual spatiotemporal graph convolutional neural network model. The calculation method is: ; If the accuracy and misjudgment rate of all transient time-domain simulation sample sets using the residual space-time graph convolutional network model reach the preset performance evaluation index threshold, a trained residual space-time graph convolutional neural network model is obtained; if not, the residual space-time graph convolutional neural network model is re-trained for space-time feature mining.

[0015] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a transient voltage stability assessment method based on unbalanced residual information learning, which has the following beneficial effects: The spatial topological matrix of the physical information of the power system is combined with the time series characteristics of the transient voltage response of each monitoring node after the fault, and the dynamic spatiotemporal evolution law of the system transient voltage instability is feature mined from the two perspectives of data and grid structure modeling itself. Different from the traditional analysis method based on single time series data or power system topology, the present invention combines the physical topology of the power system and electrical measurement data to effectively take into account the spatiotemporal characteristics of the transient voltage after the fault. In view of the problem of smoothness of spatiotemporal feature extraction caused by the increase in the number of layers of the spatiotemporal graph convolution network, the present invention introduces residual convolution to retain the original features, better realizes the mapping relationship between spatiotemporal feature extraction and the transient voltage stable state, and improves the accurate prediction of transient voltage instability. At the same time, the existing methods are insufficient in dealing with sample imbalance problems. The present invention introduces the Focal-Loss loss function to dynamically adjust the weights of positive and negative samples, so that the model pays more attention to minority samples (transient voltage instability samples), thereby improving the evaluation accuracy of transient voltage instability. The transient voltage stability assessment method based on unbalanced residual information learning of the present invention can not only improve the prediction accuracy of transient voltage stability assessment, but also has strong adaptability and can provide accurate transient voltage stability warning under complex power grid structures and multiple fault scenarios. In addition, with the popularization of power grid monitoring equipment, the model can also monitor the power grid status in real time, provide accurate basis for power grid dispatching and control decision-making, and provide strong support for the safe operation of the power grid and optimization of control strategies. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0017] Figure 1 A flowchart of a transient voltage stability assessment method based on unbalanced residual information learning provided by the present invention; Figure 2 A structural diagram of the residual spatiotemporal graph convolutional neural network model provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0018] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0019] The present invention provides a transient voltage stability assessment method based on unbalanced residual information learning, such as Figure 1 As shown, the following steps are included: S1. Build a target power system simulation model corresponding to the target power system; S2. Perform transient voltage time domain simulation on the target power system simulation model to obtain M The time-domain simulation results are obtained, and the time-space graph model corresponding to the target power system simulation model under each simulation condition is constructed according to the time-space graph construction method. G , while targeting M The time domain simulation results are respectively used to analyze the system state S Perform voltage stability or instability calibration based on M Calibration results and spatiotemporal graph models G Construct a transient time domain simulation sample set, denoted as ; S3. Construct the Focal-Loss loss function by calculating the weight factors of the stable and unstable categories of the transient time domain simulation sample set and set the adjustment factor for sample classification , we get the Focal-Loss loss function for training the residual spatiotemporal graph convolutional neural network model to learn the imbalanced sample categories; S4. Construct a residual spatiotemporal graph convolutional neural network model, and use the residual spatiotemporal graph convolutional neural network model to learn the spatiotemporal features of the transient time domain simulation sample set to obtain the prediction result of the transient voltage state; S5. Determine whether the prediction result is correct according to the actual transient voltage state result of the target power system. If it is not correct, perform the next round of training on the residual spatiotemporal graph convolutional neural network model until the preset performance evaluation index threshold of spatiotemporal feature learning is met, and obtain the trained residual spatiotemporal graph convolutional neural network model; S6. The transient voltage stability of the power system is evaluated by using the trained residual spatiotemporal graph convolutional neural network model.

[0020] It should be noted that: DSP-Studio electromechanical transient time-domain simulation software is used to build a model of the target power system to fully capture the temporal and spatial characteristics of the transient voltage instability evolution of each monitoring node in the system; The residual spatiotemporal graph convolutional neural network model uses the python deep neural network model pytorch framework.

[0021] In order to further implement the above technical solution, the specific contents of the spatiotemporal graph construction method in S2 include: According to the construction method of space-time graph, construct the k The observation time window length of a time domain simulation sample is The graphical model corresponding to the physical topology of the target power system G , and obtain the final space-time graph model ; Each monitoring node in the target power system under the observation time window is represented as a vertex in the graph model, denoted as X and , N The number of monitoring nodes is represented by the connection relationship between each monitoring node as a topological adjacency matrix, which is recorded as A and , then k The length of the sample observation time window is The graphical model corresponding to the physical topology of the target power system ; At any monitoring node Get node time series data on the edge Nodes used to reflect the target power system and nodes The physical connection relationship between them, if there is a connection, is recorded as , otherwise it is recorded as ; Finally, the final spatiotemporal graph model is constructed To characterize the dynamic characteristics of the transient response data of the target power system in the time and space dimensions.

[0022] It should be noted that: The monitoring nodes in the system include: generators, substations and loads, and the node time series data include: voltage amplitude, active power and reactive power.

[0023] In order to further implement the above technical solution, the monitoring nodes in the system include: generators, substations and loads, and the node time series data include: voltage amplitude, active power and reactive power.

[0024] In order to further implement the above technical solution, the specific contents of S2 include: The transient voltage time domain simulation is performed on the target power system simulation model, and the voltage stability of the power grid under various typical operating conditions is simulated by combining the system historical operation mode and fault settings. The total number is The time domain simulation results of the times; After each time domain simulation, record the physical adjacency matrix of the target power system And the length of the observation time window on the corresponding monitoring node after the fault occurs is The temporal evolution characteristics of ,include The voltage amplitude under the observation time window length , Active Power and reactive power , and the system status is analyzed based on the transient voltage stability engineering criterion used in the actual power system. Calibration, where Indicates that the power system state result of this time domain transient simulation is voltage instability. It indicates that the result of the simulated power system state is voltage stability; based on M Calibration results and spatiotemporal graph models G Construct the input sample set as ,in, .

[0025] In order to further implement the above technical solution, M It is the product of the typical operating level number of generators, new energy output and load consumption, the proportion of transient operating dynamic load and the typical fault disturbance number.

[0026] In order to further implement the above technical solution, the Focal-Loss loss function in S3 specifically includes: calculating the weight factors of the stable and unstable categories of the transient time domain simulation sample set and set the adjustment factor for sample classification ; Category weight factor Set according to the ratio of positive samples to negative samples: The unstable samples of the transient time-domain simulation sample set are taken as positive samples, and the stable samples of the transient time-domain simulation sample set are taken as negative samples; Assume that the number of positive samples is c , the number of negative samples is d The total number of samples is , then the weight ratio of unstable samples is , for the weight ratio of stable samples ; Adjustment Factor Used to control the degree of loss weight reduction for sample classification, while enhancing the residual spatiotemporal graph convolutional network model's attention to difficult-to-evaluate samples: ; Based on category weight factor and adjustment factors The Focal-Loss loss function is: ; in, is the predicted probability of the residual spatiotemporal graph convolutional network model for the sample. For unstable samples , for a stable sample .

[0027] In order to further implement the above technical solution, the residual spatiotemporal graph convolutional neural network model in S4 specifically includes: a residual spatiotemporal graph convolution module and a fully connected classification module; The residual spatiotemporal graph convolution module includes a spatial graph convolution layer, a temporal convolution layer, and a residual convolution layer; The spatial graph convolution layer is used to obtain any spatiotemporal graph input sample in the input sample set and perform topological adjacency matrix analysis on the power system. Perform spatial feature mining and realize the feature aggregation of neighbor node time series information to the central node on the power system topology map, and output spatial features ; Temporal convolution layer, used to further Perform convolution operation to extract the spatiotemporal features of transient voltage response data and output the timing features ; Residual convolution layer is used to Perform convolution operation and output residual features And keep the output features of the spatiotemporal graph Dimensionality is consistent; Fully connected classification module for residual features and spatiotemporal characteristics The fusion is performed and the dimensionality of the fused features is reduced through a linear layer to further extract high-dimensional features for classifying transient voltage stability and instability. The high-dimensional features are converted into probability distributions through a softmax activation function to complete the mapping of the short-term transient voltage stability state space and realize the transient voltage stability or instability assessment of the system.

[0028] In order to further implement the above technical solution, the spatial graph convolution layer outputs spatial features The specific contents include: The adjacency matrix of the target power system topology Degree Matrix Conversion, the conversion method is: ; In the formula, is the degree matrix Line Elements of a column; Based on the identity matrix Sum degree matrix , and further use the Laplace graph transformation formula to transform the power system topology matrix Perform spatial feature mining, where the Laplace graph transformation formula is: ; Among them, the identity matrix is a special square matrix where all elements on the main diagonal are 1 and all other elements are 0. It is also a special square matrix used to represent the degree of each vertex in the graph. This degree is the number of connections between each node and other nodes in the power grid adjacency matrix. Perform spatial feature mining based on the solution formula Get the target power system topology adjacency matrix The orthonormal eigenvectors of and the spatial eigenvalue matrix , the graph node information aggregation formula is used to realize the feature aggregation of the neighbor node time series information on the power system topology graph to the central node, where the graph node information aggregation formula is: ; In the formula, is the graph convolution kernel, is the length of the time window for observing nodes in the graph matrix The temporal evolution characteristics of .

[0029] In order to further implement the above technical solution, the temporal convolution layer outputs temporal features The specific contents include: ; Among them, ReLU is a nonlinear activation function. is the convolution kernel parameter, It is the convolution operation.

[0030] In order to further implement the above technical solution, the fully connected classification module reduces the dimension of the fused features through the linear layer, further extracts the high-dimensional features for classifying transient voltage stability and instability, and converts the high-dimensional features into probability distribution through the fully connected classification formula. , the mapping of the short-term transient voltage stability state space is completed, and the transient voltage stability state evaluation of the system is realized. The full connection classification formula is: .

[0031] It should be noted that: As the number of spatiotemporal graph network layers increases, the feature vectors of different nodes will become more and more similar. This phenomenon is called oversmoothing, which will significantly reduce the stable evaluation performance of the model. To enhance the ability to capture the spatiotemporal dependencies of transient processes, the residual convolution layer adds the output of the previous layer to the output of the current layer.

[0032] After being processed by the residual spatiotemporal graph convolutional network, the output information retains the temporal dependence of the transient voltage response of each node and integrates the spatial dependence of the power system topology, which can more comprehensively characterize the spatiotemporal characteristics of the transient response data.

[0033] The residual spatiotemporal graph convolutional network model constructed in this embodiment is as follows Figure 2 shown.

[0034] In order to further implement the above technical solution, the specific contents of S5 include: The accuracy and misclassification rate are selected as the performance evaluation indicators of the residual spatiotemporal graph convolutional neural network model for spatiotemporal feature learning. The accuracy rate Acc is used to reflect the overall performance of the residual spatiotemporal graph convolutional neural network model, and the misclassification rate MAR reflects the probability that an unstable sample is misclassified as stable by the residual spatiotemporal graph convolutional neural network model. The higher the accuracy rate and the lower the misclassification rate, the better the evaluation performance of the residual spatiotemporal graph convolutional neural network model. The total number of correct transient time-domain simulation samples evaluated by the residual spatiotemporal graph convolutional neural network model in this round of training is , the total number of unstable samples is evaluated as unstable And the total number of unstable samples evaluated as stable , and calculate the transient stability assessment accuracy and misjudgment rate of the residual spatiotemporal graph convolutional neural network model. The calculation method is: ; If the accuracy and misjudgment rate of all transient time-domain simulation sample sets using the residual spatiotemporal graph convolutional network model reach the preset performance evaluation index threshold (in this embodiment ), then the trained residual spatiotemporal graph convolutional neural network model is obtained; if it is not reached (in this embodiment ), the residual spatiotemporal graph convolutional neural network model is retrained for spatiotemporal feature mining.

[0035] It should be noted that: In this embodiment, the trained residual spatiotemporal graph convolutional neural network model is used to monitor the transient voltage state of the system in real time, extract the physical topology matrix of the power system after the fault and the observation time window length of each monitoring node on the matrix The time series response data is input into the trained residual spatiotemporal graph convolutional neural network model obtained by S5. Conduct transient voltage stability assessment of the corresponding system. , it means that the power system will be at risk of voltage instability in a short period of time after experiencing a transient fault, and may even cause voltage collapse. Voltage instability may lead to serious consequences, including damage to power equipment, large-scale power outages, or system collapse. Therefore, it is necessary to send an alarm to the power grid control center and prepare targeted control measures to restore the operation of the power system to stability. If the stable state is determined , indicating that the transient voltage of the power system is stable, there is no serious voltage fluctuation in the system, and the voltage will remain within a safe range for a certain period of time, but real-time monitoring is required.

[0036] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. A transient voltage stability assessment method based on unbalanced residual information learning, characterized in that: The following steps are involved: S1. Build a target power system simulation model corresponding to the target power system; S2. Perform transient voltage time domain simulation on the target power system simulation model to obtain M The time-domain simulation results are obtained, and the time-space graph model corresponding to the target power system simulation model under each simulation condition is constructed according to the time-space graph construction method. G , while targeting M The time domain simulation results are respectively used to analyze the system state S Perform voltage stability or instability calibration based on M Calibration results and spatiotemporal graph models G Construct a transient time domain simulation sample set, denoted as ; S3. Construct the Focal-Loss loss function by calculating the weight factors of the stable and unstable categories of the transient time domain simulation sample set and set the adjustment factor for sample classification Obtain the Focal-Loss loss function for training the residual spatiotemporal graph convolutional neural network model to learn the imbalanced sample categories; S4. Construct a residual spatiotemporal graph convolutional neural network model, and use the residual spatiotemporal graph convolutional neural network model to learn the spatiotemporal features of the transient time domain simulation sample set to obtain the prediction result of the transient voltage state; S5. Determine whether the prediction result is correct according to the actual transient voltage state result of the target power system. If it is not correct, perform the next round of training on the residual spatiotemporal graph convolutional neural network model until the preset performance evaluation index threshold of spatiotemporal feature learning is met, and obtain the trained residual spatiotemporal graph convolutional neural network model; S6. The transient voltage stability of the power system is evaluated by using the trained residual spatiotemporal graph convolutional neural network model.

2. The transient voltage stability assessment method based on unbalanced residual information learning according to claim 1 is characterized in that: The specific contents of the space-time graph construction method in S2 include: According to the construction method of space-time graph, construct the k The observation time window length of a time domain simulation sample is The graphical model corresponding to the physical topology of the target power system G , and obtain the final space-time graph model ; The first The monitoring nodes are represented as vertices in the graph model. , recorded as X and , N is the number of monitoring nodes, The monitoring node and The connection relationship between the monitoring nodes is used as the edge in the graph model , collect all It is represented as a topological adjacency matrix, denoted as A and , among which k The length of the sample observation time window is The graphical model corresponding to the physical topology of the target power system ; At any monitoring node Get the node time series data on the top, if there is a connection, record it as , otherwise it is recorded as ; Finally, the final spatiotemporal graph model is constructed To characterize the dynamic characteristics of the transient response data of the target power system in the time and space dimensions.

3. The transient voltage stability assessment method based on unbalanced residual information learning according to claim 1 is characterized in that: The specific contents of S2 include: The transient voltage time domain simulation is performed on the target power system simulation model, and the voltage stability of the power grid under various typical operating conditions is simulated by combining the system historical operation mode and fault settings. The total number is The time domain simulation results of the times; After each time domain simulation, record the physical adjacency matrix of the target power system And the length of the observation time window on the corresponding monitoring node after the fault occurs is The temporal evolution characteristics of ,include The voltage amplitude under the observation time window length , Active Power and reactive power , and the system status is analyzed based on the transient voltage stability engineering criterion used in the actual power system. Calibration, where Indicates that the power system state result of this time domain transient simulation is voltage instability. It indicates that the result of the simulated power system state is voltage stability; based on M Calibration results and spatiotemporal graph models G Construct the input sample set as ,in, 。 4. The transient voltage stability assessment method based on unbalanced residual information learning according to claim 1 is characterized in that: M It is the product of the typical operating level number of generators, new energy output and load consumption, the proportion of transient operating dynamic load and the typical fault disturbance number.

5. The transient voltage stability assessment method based on unbalanced residual information learning according to claim 1 is characterized in that: The Focal-Loss loss function in S3 specifically includes: Calculating the weight factors of stable and unstable categories of transient time domain simulation sample sets and set the adjustment factor for sample classification ; Category weight factor Set according to the ratio of positive samples to negative samples: The unstable samples of the transient time-domain simulation sample set are taken as positive samples, and the stable samples of the transient time-domain simulation sample set are taken as negative samples; Assume that the number of positive samples is c , the number of negative samples is d The total number of samples is M = c + d , then the weight ratio of unstable samples is , for the weight ratio of stable samples ; Adjustment Factor Used to control the degree of loss weight reduction for sample classification, while enhancing the residual spatiotemporal graph convolutional network model's attention to difficult-to-evaluate samples: ； Based on category weight factor and adjustment factors The Focal-Loss loss function is: ； in, is the predicted probability of the residual spatiotemporal graph convolutional network model for the sample. For unstable samples , for a stable sample .

6. The transient voltage stability assessment method based on unbalanced residual information learning according to claim 1 is characterized in that: The residual spatiotemporal graph convolutional neural network model in S4 specifically includes: a residual spatiotemporal graph convolution module and a fully connected classification module; The residual spatiotemporal graph convolution module includes a spatial graph convolution layer, a temporal convolution layer, and a residual convolution layer; The spatial graph convolution layer is used to obtain any spatiotemporal graph input sample in the input sample set and perform topological adjacency matrix analysis on the power system. Perform spatial feature mining and realize the feature aggregation of neighbor node time series information to the central node on the power system topology map, and output spatial features ; Temporal convolution layer, used to further Perform convolution operation to extract the spatiotemporal features of transient voltage response data and output the timing features ; Residual convolution layer is used to Perform convolution operation and output residual features And keep the output features of the spatiotemporal graph Dimensionality is consistent; Fully connected classification module for residual features and spatiotemporal characteristics The fusion is performed and the dimensionality of the fused features is reduced through a linear layer to further extract high-dimensional features for classifying transient voltage stability and instability. The high-dimensional features are converted into probability distributions through a softmax activation function to complete the mapping of the short-term transient voltage stability state space and realize the transient voltage stability or instability assessment of the system.

7. The transient voltage stability assessment method based on unbalanced residual information learning according to claim 6 is characterized in that: The spatial graph convolution layer outputs spatial features The specific contents include: The adjacency matrix of the target power system topology Degree Matrix Conversion, the conversion method is: ； In the formula, is the degree matrix Line Elements of a column; Based on the identity matrix Sum degree matrix , and further use the Laplace graph transformation formula to transform the power system topology matrix Perform spatial feature mining, where the Laplace graph transformation formula is: ； in, is the identity matrix; degree matrix It is used to represent the degree of each vertex in the graph model. The degree is the number of connections between each node and other nodes on the power grid adjacency matrix. Perform spatial feature mining based on the solution formula Get the target power system topology adjacency matrix The orthonormal eigenvectors of and the spatial eigenvalue matrix , the graph node information aggregation formula is used to realize the feature aggregation of the neighbor node time series information on the power system topology graph to the central node, where the graph node information aggregation formula is: ； In the formula, is the graph convolution kernel, is the length of the time window for observing nodes in the graph matrix The temporal evolution characteristics of .

8. The transient voltage stability assessment method based on unbalanced residual information learning according to claim 1, characterized in that: The temporal convolution layer outputs temporal features The specific contents include: ； Among them, ReLU is a nonlinear activation function. is the convolution kernel parameter, It is the convolution operation.

9. The transient voltage stability assessment method based on unbalanced residual information learning according to claim 6, characterized in that: The fully connected classification module reduces the dimension of the fused features through the linear layer, further extracts high-dimensional features for classifying transient voltage stability and instability, and converts the high-dimensional features into probability distribution through the fully connected classification formula. , the mapping of the short-term transient voltage stability state space is completed, and the transient voltage stability state evaluation of the system is realized. The full connection classification formula is: 。 10. The transient voltage stability assessment method based on unbalanced residual information learning according to claim 6, characterized in that: The specific contents of S5 include: The accuracy and misclassification rate are selected as the performance evaluation indicators of the residual spatiotemporal graph convolutional neural network model for spatiotemporal feature learning. The accuracy rate Acc is used to reflect the overall performance of the residual spatiotemporal graph convolutional neural network model, and the misclassification rate MAR reflects the probability that an unstable sample is misclassified as stable by the residual spatiotemporal graph convolutional neural network model. The higher the accuracy rate and the lower the misclassification rate, the better the evaluation performance of the residual spatiotemporal graph convolutional neural network model. The total number of correct transient time-domain simulation samples evaluated by the residual spatiotemporal graph convolutional neural network model in this round of training is , the total number of unstable samples is evaluated as unstable And the total number of unstable samples evaluated as stable , and calculate the transient stability assessment accuracy and misjudgment rate of the residual spatiotemporal graph convolutional neural network model. The calculation method is: ； If the accuracy and misjudgment rate of all transient time-domain simulation sample sets using the residual space-time graph convolutional network model reach the preset performance evaluation index threshold, a trained residual space-time graph convolutional neural network model is obtained; if not, the residual space-time graph convolutional neural network model is re-trained for space-time feature mining.

Images