A method for transient stability assessment of misjudgment-oriented graph convolutional network

Abstract

A kind of transient stability evaluation method for misjudgment graph convolution network, comprising the following steps: step 1: different operating scenarios of power system are simulated to construct sample set;Step 2: obtain improved adaptive graph convolution network GCN;Step 3: based on the improved adaptive graph convolution network GCN, an evaluation model is established, and offline training is carried out;Step 4: in the process of online application, the stability class and stability level obtained by the model are coordinated by constructing an auxiliary decision module, to judge the real operating state of power system.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of power system transient stability analysis, and particularly relates to a graph convolution network transient stability evaluation method for misjudgment. BACKGROUND

[0002] Transient stability generally refers to the ability of a power system to maintain synchronization when subjected to severe disturbances, such as short-circuit faults on transmission lines. With the continuous expansion of the scale of renewable energy access based on low-inertia converters, the stability margin of the power system is reduced, and the operating conditions are becoming increasingly tense. In this case, developing a transient stability assessment (TSA) method with low misjudgment rate to provide early warning of risks and timely start emergency control becomes a problem worth attention.

[0003] Conventionally, a reliable TSA method is the time-domain simulation method, which solves the high-dimensional nonlinear differential-algebraic equations of the power system through numerical integration algorithms, but the increasing computational burden brought about by the expansion of the system scale limits the online application of TSA. The emergence of artificial intelligence methods enables online TSA to be widely applied. The patent document with application publication number CN114629111A discloses a power system transient stability assessment method based on multi-source information, which uses time-domain simulation to construct a massive data set, and uses the operating information and fault information as the input features of a convolutional neural network model for TSA. The patent document with application publication number CN112909939A discloses a power system transient stability assessment method based on pre-fault and post-fault power flow characteristics, which uses a support vector machine classifier to make judgments on the transient stability state of the power system under the concepts of safe domain and stable domain according to pre-fault and post-fault information. In the above-mentioned methods, the shallow network model relies on the experience of experts to select input features, which greatly limits its generalization ability and TSA accuracy in practical applications, and the system dynamic characteristics from the start of the fault to the end of the fault or longer are used as inputs, which also limits the evaluation rate to some extent.

[0004] In summary, most of the current power system TSA methods do not maximize the online TSA and ensure high accuracy. In order to achieve this goal, the method proposed by the present application uses only the snapshot information within the first cycle after the fault occurs, integrates two stability judgment methods to improve accuracy, has higher computational efficiency, and also has higher robustness. SUMMARY

[0005] The present application proposes a graph convolution network transient stability evaluation method for misjudgment, aiming to solve the technical problems of low evaluation rate and high misjudgment rate existing in the prior art.

[0006] To solve the above technical problems, the technical scheme adopted by the present application is:

[0007] A method for transient stability evaluation of misjudgment graph convolutional network, comprising the following steps:

[0008] Step 1: Construct a sample set by simulating different operation scenarios of the power system;

[0009] Step 2: Obtain an improved adaptive graph convolutional network GCN;

[0010] Step 3: Establish an evaluation model based on the improved adaptive graph convolutional network GCN and perform offline training;

[0011] Step 4: In the online application process, coordinate the stability class and stability level obtained by the model through the construction of an auxiliary decision-making module to judge the real operation state of the power system.

[0012] In step 1, the construction of the sample set specifically includes the following steps:

[0013] Step 1-1: Simulate the operation scenario of the power system and select the input features of the model;

[0014] Step 1-2: Construct the stability evaluation index of the sample.

[0015] In step 1-1, the sampling data of bus voltage amplitude and phase angle, load active and reactive power, and generator active and reactive power within the first cycle after the N-1 three-phase short-circuit fault under normal fluctuation of the generator and load are taken as input to construct the feature matrix X m ∈R N×6 .

[0016] In step 1-2, each sample is assigned two stability labels, transient stability index TSI and stability level index SLI, which are defined as follows:

[0017]

[0018]

[0019] Where: Δδ max is the power angle difference between any two generators; when |Δδ| max >180°, i.e. TSI<0, the system is transiently unstable, marked as c=[1,0], otherwise, marked as c=[0,1]; ε is a stability threshold, and the value range of SLI is [-1,1].

[0020] In step 2, the improved neural network adaptive graph convolutional network GCN is specifically:

[0021] Input layer→first graph convolutional layer→first anti-residual module→second graph convolutional layer→second anti-residual module→third graph convolutional layer→third anti-residual module→output layer.

[0022] The mathematical expression of the graph convolutional layer is as follows:

[0023] H (r+1) = f(H r ,A) (3)

[0024]

[0025] In the formula, H (r+1) and H r are the r+1th and rth node features respectively; f(·,·) is a function operation; A is a correlation matrix; h(·) is a nonlinear operation; is a normalized correlation matrix; W r is a learning operation matrix;

[0026] When the graph convolutional layer is used, first, the first layer of graph convolution and the anti-residual module expand the channel to achieve the effect of dimensionality increase; then, the second layer of graph convolution and the anti-residual module independently perform two-dimensional convolution operation on each feature channel, and the channel number remains unchanged and the feature map size becomes smaller; finally, the third layer of graph convolution and the anti-residual module are used for dimensionality reduction to output the feature map; after the dimensionality expansion, deep convolution and dimensionality reduction of the above anti-residual layer, the representation ability of the feature map is improved.

[0027] In step 3, the following steps are specifically included:

[0028] Step 3-1: constructing an evaluation model based on the improved adaptive graph convolutional network GCN;

[0029] Step 3-2: adding a stable class evaluation module SCAM and a stable level evaluation module SLAM to reduce the incidence of misjudgment;

[0030] Step 3-3: establishing a loss function of the joint stable class evaluation module SCAM and the stable level evaluation module SLAM to improve the evaluation accuracy of the model.

[0031] In step 3-1, the adaptive graph convolutional network GCN takes the feature information selected by the anti-residual module as input, which includes an initial node feature matrix X out ∈R N×d and an initial adjacency matrix A∈R N×N , where N is the number of samples and d is the feature dimension, the adjacency matrix A is set to be parameter adaptive update, and the weight of the neighborhood node is adjusted adaptively, which can better fit the topology structure change in the actual power system, and the calculation formula of the graph convolutional layer is as follows:

[0032]

[0033] where X out is the node feature; is the neighborhood aggregation information; " / / " is the concatenation along the feature dimension; W is the weight matrix; σ(·) is the nonlinear activation function; Z1 is the node feature matrix, and the input layer of the GCN network has N nodes, corresponding to the number of nodes in the power system.

[0034] In step 3-2, SCAM and SLAM composed of auto-encoder are connected at the end of adaptive graph convolutional network GCN, SCAM judges the stability class, and SLAM estimates the stability level. The expressions of the two modules are as follows:

[0035]

[0036]

[0037] where: is the confidence of the ith sample; is the output of the SCAM module; z SLAM is the output of the SLAM; is the transient stability level.

[0038] In step 3-3, the loss function contains SCAM, SLAM and a regularization term, as follows:

[0039] L = L SCAM + L SLAM + βL1 (8)

[0040]

[0041]

[0042] where β is the regularization coefficient; L1 is the regularization; c ni is the class of the label; is the confidence of the nth sample label; In order to reduce the interference of sample imbalance on evaluation, the loss function coefficient λ n of stable samples is set to 1, and the loss function coefficient λ n of unstable samples is set to λ = number of stable samples / number of unstable samples; N is the total number of samples; is the predicted SLI of the nth sample; SLI n is the true SLI of the nth sample. In the training process, the loss function is optimized by Adam.

[0043] In the online application stage, when the real power system data is input into the adaptive GCN evaluation model, the stable label is obtained from the trained evaluation model and the stable level For SCAM, the system is stable when , otherwise unstable; for SLAM, the system remains stable when , the system loses stability when , and if , the system state is uncertain; the auxiliary decision module is constructed to coordinate SCAM and SLAM, and when the instability judgment is made for both modules, the instability judgment is output; when both modules give stable judgment, the stable is output; when the outputs of the two modules conflict, the SLAM that can output the stable level is used as the reference; when the SLAM output is uncertain, the output of the SCAM module is corrected to obtain the stable result.

[0044] Compared with the prior art, the present application has the following technical effects:

[0045] 1) In the construction of the sample set, the time efficiency of TSA is fully considered, the system feature information within the first cycle after the fault occurs is used as the input, and good online TSA accuracy can be realized;

[0046] 2) A transient stability judgment standard based on TSI and SLI two indexes is proposed. The auxiliary decision module integrates the stability class and stability level judgment method, effectively reducing the model misjudgment rate;

[0047] 3) A transient stability evaluation method for misjudgment of graph convolution network is proposed, which fully considers the actual operation state of the system, uses an adaptive graph convolution network GCN evaluation model and adds a residual module, improves the feature expression and evaluation rate, and meets the requirements of online evaluation of power system transient stability;

[0048] 4) The purpose of the present application is to expand the TSA method of the power system to adapt to the power system with the addition of new energy, while ensuring the online TSA accuracy of the power system. In the model input aspect, the system feature information within the first cycle after the fault occurs is used as the input, which can realize good online TSA accuracy; in the system model construction aspect, the anti-residual module is used to improve the input information expression, and the adaptive graph convolution network GCN evaluation model is used to improve the robustness of the model, and in order to improve the evaluation accuracy, two transient stability judgment methods are integrated. Thus a fast TSA method for misjudgment of graph convolution network is provided. BRIEF DESCRIPTION OF DRAWINGS

[0049] The present application will be further described below in conjunction with the drawings and examples:

[0050] Figure 1 is a system flowchart of the present application;

[0051] Figure 2 is an adaptive GCN network structure diagram of the present application;

[0052] Figure 3 is a working principle diagram of the auxiliary decision module in the example of the present application;

[0053] Figure 4 is an IEEE 39-node system schematic diagram in the example of the present application;

[0054] Figure 5 is a TSI judgment example diagram in the example of the present application;

[0055] Figure 6 is an SLI judgment example diagram in the example of the present application. DETAILED DESCRIPTION

[0056] A method for transient stability assessment of a misjudged graph convolutional network, as shown in Figure 1 , the method comprises the following steps:

[0057] Step 1: Simulate different operating scenarios of a power system to construct a sample set;

[0058] Step 2: Obtain an improved adaptive graph convolutional network GCN;

[0059] Step 3: Establish an evaluation model based on the improved adaptive graph convolutional network GCN, and perform offline training;

[0060] Step 4: In the online application process, the stability class and stability level obtained by the model are coordinated by constructing an auxiliary decision module to judge the real operating state of the power system.

[0061] In step 1: Specifically, since the TSA of the post-fault power system is to quickly take emergency control measures when the power system is unstable, it requires the evaluation model to achieve fast and accurate TSA. Therefore, the construction of the sample set is particularly important. In terms of electrical feature selection, it not only needs to contain sufficient information for model training, but also needs to meet the requirement of rapidness, and the time span is as small as possible; in terms of stability label construction, it reflects the stability level of the power system under this operating condition. Therefore, the construction of the sample set is very important for the establishment of the whole model.

[0062] The construction of the sample set comprises the following steps:

[0063] 1) Simulate the N-1 three-phase short-circuit fault of the power system under normal fluctuations of generators and loads, take the sampling data of bus voltage amplitude and phase angle, load active and reactive power and generator active and reactive power within the first cycle after the fault occurs as input, and construct a feature matrix X m ∈R N×6 .

[0064] 2) Each sample is given two stability labels, TSI and SLI, defined as follows:

[0065]

[0066]

[0067] In the formula: Δδ max is the power angle difference between any two generators; when |Δδ| max >180°, that is, TSI<0, the system is transiently unstable, marked as c=[1,0], otherwise, marked as c=[0,1]; ε is a stability threshold, and the value range of SLI is [-1,1].

[0068] In step 2: Specifically, because the graph convolution network has many parameters, and the features are easily lost during feature extraction, a reverse residual module is added in the graph convolution module, as shown in Figure 2 , which increases the learning mode of the network to improve the expression ability of the feature map. In order to ensure the integrity of the electrical feature information, a three-layer network is constructed to extract features from multiple dimensions and multiple scales to enhance the generalization ability of the model and help the model accurately detect small feature objects in complex backgrounds.

[0069] The overall structure of the improved adaptive GCN network is: input layer→first graph convolution layer→first reverse residual module→second graph convolution layer→second reverse residual module→third graph convolution layer→third reverse residual module→output layer.

[0070] Firstly, the first layer of graph convolution and reverse residual module expands the channel to achieve dimensionality effect; then, the second layer of graph convolution and reverse residual module performs two-dimensional convolution operation on each feature channel independently, and the channel number remains unchanged and the feature map size becomes smaller; finally, the third layer of graph convolution and reverse residual module restores the dimension to output the feature map; after the dimension expansion, deep convolution and dimension reduction of the above reverse residual layer, the representation ability of the feature map is improved.

[0071] The mathematical expression of the graph convolution layer is as follows:

[0072] H (r+1) =f(H r ,A) (3)

[0073]

[0074] wherein H (r+1) , H r are the r+1th and rth node features respectively; f(·,·) is a function operation; A is a correlation matrix; h(·) is a nonlinear operation; is a normalized correlation matrix; W r is a learning operation matrix.

[0075] In step 3: Specifically, based on the improved adaptive GCN network, an evaluation model is constructed, and the collected electrical feature information is taken as input for model training:

[0076] 1) The adaptive GCN network contains an initial node feature matrix X out ∈R N×d and an initial adjacency matrix A∈R N×N , wherein N is the sample number, d is the feature dimension, the adjacency matrix A is set to be parameter adaptive update, and the weight of the neighborhood node is adjusted adaptively, which can better fit the topology structure change in the actual power system. The calculation formula of the graph convolution layer is as follows:

[0077]

[0078] wherein X out is a node feature; is neighborhood aggregation information; " / / " is splicing along the feature dimension; W is a weight matrix; σ(·) is a nonlinear activation function; Z1 is a node feature matrix. The input layer of the GCN network has N nodes, corresponding to the number of nodes in the power system.

[0079] 2) Add a stable class discrimination module to reduce the incidence of misjudgment. Under the adaptive GCN network, SCAM and SLAM composed of self-encoders are connected at the lower end. SCAM judges the stability class, and SLAM estimates the stability level. The expressions of the two modules are as follows:

[0080]

[0081]

[0082] wherein: is the confidence of the ith sample; is the output of the SCAM module; z SLAM is the output of the SLAM; is the transient stability level.

[0083] 3) Establish a loss function combining SCAM and SLAM to improve the evaluation accuracy of the model. The loss function contains SCAM, SLAM and a regularization term, as follows:

[0084] L = L SCAM + L SLAM + βL1 (8)

[0085]

[0086]

[0087] wherein: β is a regularization coefficient; L1 is regularization; c ni is the class of the label; is the confidence of the n-th sample label; in order to reduce the interference of sample imbalance on evaluation, the loss function coefficient λ n of unstable samples is set to 1, and the loss function coefficient λ n of stable samples is set to 1; N is the total number of samples; and is the predicted SLI of the n-th sample; SLI n is the true SLI of the n-th sample. In the training process, the loss function is optimized by Adam.

[0088] In step 4: specifically, in the online application stage, when the real power system data is input into the model, the stable label and the stability level are obtained from the trained evaluation model. When the SCAM is stable, the system is stable. Conversely, it is unstable; for the SLAM, when the system remains stable, when the system is unstable, and if the system state is uncertain. The auxiliary decision module is constructed to coordinate the SCAM and the SLAM. When both modules make an unstable judgment, the unstable judgment is output; when both modules give a stable judgment, the stable is output; when the outputs of the two modules conflict, the SLAM which can output the stability level is used as the criterion; when the SLAM output is uncertain, the output of the SCAM module is corrected to obtain the final stable result.

[0089] Embodiment:

[0090] The method proposed in the application is simulated on an IEEE 39-node system as shown in Figure 4 The following modifications are made: it is assumed that two wind machines replace synchronous generators G3 and G8, and the power generation capacity is the same as that of the original generators.

[0091] The active power of the generator, the fan and the load fluctuates between 80% to 120%, the fault is set at 10% to 100% with a step of 10% on the transmission line, the fault type is N-1 three-phase short-circuit fault, the fault duration is set to 0.1s, and the simulation time is 4s. The simulation generates 8179 samples. After the sample is generated, each sample has two labels according to the TSI index and the SLI index, and the samples are randomly divided into a training set and a test set in a ratio of 4:1, the training set is used to train the model, and the test set is used to test the performance of the model. As shown in Figure 5 、 Figure 6 The curve of the generator power angle changes under certain fault conditions, and the stability condition is judged by the curve change.

[0092] Compared with the prior art, the application provides a transient stability assessment method for misjudgment of a graph convolution network. The application first uses extremely short time snapshot feature information as model input, expands the control time after evaluation, and improves the evaluation speed. Then, an adaptive GCN evaluation model integrating a stability class and stability level judgment module is used, two judgment methods are mutually corrected, the accuracy of TSA is improved, and the occurrence of misjudgment is reduced. Through the application, the TSA method of the power system can be more rapid and accurate, and the transient stability of the power system is ensured.

Claims

1. A method for misjudgment of graph convolution network transient stability evaluation, characterized in that, The method comprises the following steps: Step 1: constructing a sample set by simulating different operation scenarios of the power system; Step 2: obtaining an improved adaptive graph convolution network GCN; Step 3: establishing an evaluation model based on the improved adaptive graph convolution network GCN, and performing offline training; Step 4: in the online application process, coordinating the stable categories and stable levels obtained by the model through the construction of an auxiliary decision-making module to judge the real operation state of the power system; In step 1, the construction of the sample set specifically comprises the following steps: Step 1-1: simulating the operation scenario of the power system and selecting the input features of the model; Step 1-2: constructing the stability evaluation indexes of the sample; In step 1-2, each sample is assigned two stability labels, namely the transient stability index TSI and the stability level index SLI, which are defined as follows: （1）； （2）； In the formula: is the absolute value of the power angle difference between any two generators; when > 180°, i.e., TSI < 0, the system is transiently unstable, marked as c = [1, 0], otherwise, marked as c = [0, 1]; is a stability threshold, and the value range of SLI is [-1, 1].

2. The method of claim 1, wherein: In step 1-1, the simulation power system occurs N -1 three-phase short-circuit fault, the sampling data of bus voltage amplitude and phase angle, load active and reactive power and generator active and reactive power within the first cycle after the fault occurs are taken as input, and a feature matrix is constructed .

3. The method of claim 1, wherein: In step 2, the improved neural network adaptive graph convolution network GCN is specifically as follows: Input layer→first graph convolution layer→first reverse residual module→second graph convolution layer→second reverse residual module→third graph convolution layer→third reverse residual module→output layer.

4. The method of claim 3, wherein, The mathematical expression of the graph convolution layer is as follows: （3）； （4）； In the formula: , are the characteristics of the first r +1 and the rth node respectively; is a function operation; is a correlation matrix; is a nonlinear operation; is a normalized correlation matrix; is a learning operation matrix; When the graph convolution layer is used, first, the first graph convolution layer and the first reverse residual module expand the channel to achieve the effect of dimensionality reduction; then, the second graph convolution layer and the second reverse residual module independently perform two-dimensional convolution operation on each feature channel, and the channel number remains unchanged and the feature map size becomes smaller; finally, the third graph convolution layer and the third reverse residual module are used for dimensionality reduction, and the feature map is output; after the dimension expansion, deep convolution and dimension reduction of the reverse residual module, the representation ability of the feature map is improved.

5. The method of claim 1, wherein: In step 3, it specifically comprises the following steps: Step 3-1: constructing an evaluation model based on the improved adaptive graph convolution network GCN; Step 3-2: adding a stable category evaluation module SCAM and a stable level evaluation module SLAM to reduce the occurrence rate of misjudgment; Step 3-3: establishing a loss function of the joint stable category evaluation module SCAM and the stable level evaluation module SLAM to improve the evaluation accuracy of the model.

6. The method of claim 5, wherein: In step 3-1, the adaptive graph convolutional network GCN takes the feature information selected by the anti-residual module as input, which contains the initial node feature matrix and the initial adjacency matrix , where is the number of samples, is the feature dimension, and the adjacency matrix is set to be adaptively updated. By adaptively adjusting the weights of the neighborhood nodes, the topology structure change in the actual power system can be better fitted. The calculation formula of the graph convolutional layer is as follows: （5）； In the formula: is the node feature; is the neighborhood aggregation information; is spliced along the feature dimension; is the weight matrix; is a nonlinear activation function; is the node feature matrix, and the input layer of the GCN network has N nodes, corresponding to the number of nodes in the power system.

7. The method of claim 5, wherein: In step 3-2, the SCAM and the SLAM composed of the autoencoder are connected at the lower end of the adaptive graph convolution network GCN, the SCAM judges the stability category, the SLAM estimates the stability level, and the expressions of the two modules are as follows: （6）； （7）； wherein: is the confidence of the i th sample; is the output of the SCAM module; is the output of the SLAM; is the level of transient stability.

8. The method of claim 5, wherein: In step 3-3, the loss function includes the SCAM, the SLAM and a regularization term, as follows: （8）； （9）； （10）； In the formula: The regularization coefficient is used. For regularization; The category of the label; For the first n The confidence level of each sample label; to reduce the interference of sample imbalance on the evaluation, the loss function coefficients of the stable samples are calculated. Set to 1, unstable sample setting =Number of stable samples Number of unstable samples; This represents the total number of samples; For the first n Predicted SLI for each sample; For the first n For each sample, the true SLI is obtained, and the loss function is optimized using Adam during training.

Images