Power system transient stability assessment method based on information entropy and spectrum graph network

Abstract

The present invention provides a method for transient stability assessment of power systems based on information entropy and spectrum graph networks, including an offline training part and an online assessment part. In the offline training part, the historical operation data of the target power system to be analyzed or the data generated by the simulation model are used to form tensor data sets of time windows of different lengths to train four transient stability assessment sub-agents based on spectrum graph convolution respectively, and the trained sub-agents are used to iteratively update the sub-agents used for online assessment. In the online assessment stage, the measured parameters collected during the operation of the system are sent to a startup criterion based on information entropy for judgment. When the obtained information entropy exceeds the set threshold, a startup instruction is issued. The input data tensor generation module is respectively made into multiple groups of data tensors according to the time sequence of the data input, and is successively distributed to the sub-agents for transient stability assessment. After the evaluation results are collected successively, the final transient stability judgment result is output according to the design decision logic.

Description

Technical Field

[0001] The present invention relates to the technical field of power system stability analysis, and in particular to a method for evaluating transient stability of a power system based on information entropy and a spectrum graph network. Background Art

[0002] With the gradual advancement of the dual carbon goals and the construction of smart grids, the proportion of new energy units and power electronic equipment connected to the grid has increased significantly, resulting in reduced system inertia and increased transient safety risks. The random volatility and uncontrollability of new energy output have led to increasingly complex and changeable operating conditions of the power system, which has also increased the difficulty of system safety analysis and regulation, posing severe challenges to the accuracy and reliability of power system safety and stability analysis.

[0003] There are four main types of transient stability analysis methods for power systems, namely time domain simulation methods, energy function methods, extended equal area rule methods and artificial intelligence methods. Among them, with the rapid development of artificial intelligence technology algorithms and hardware computing power in recent years, more and more research focuses on using artificial intelligence technology to enable the analysis, dispatching, operation and control of power systems. At present, there are similar transient stability analysis implementation schemes based on artificial intelligence technology, such as common deep learning models such as convolutional neural network (CNN), long short-term memory neural network (LSTM) and deep belief network (DBN) for stability judgment, and the use of generative adversarial network (GAN) to make up for the insufficient amount of fault data and correct data imbalance problems; using stacked denoising autoencoder (SAE) to extract multi-layer features of input data for stability judgment and using cascaded deep learning model evaluation methods. The above analysis methods based on deep learning models do not take into account the topological information of the power system when extracting features, resulting in limited accuracy of the analysis results. The graph convolutional neural network (GCN) and graph attention network (GAT) models used subsequently fully consider the spatial correlation between system state information, introduce the topological structure of the power system, and conduct a more comprehensive assessment of system stability. However, the input data of the above-mentioned graph neural network algorithm requires the measurement data of all nodes in the system topology. In actual power systems, only some nodes are equipped with synchronized phasor measurement units (PMUs), which makes it difficult to provide all the measurement data required for transient stability analysis, and therefore face difficulties in practical application. In addition, existing research focuses less on the framework design of the online operation of the transient stability assessment agent, and in actual applications, it does not design when and what criteria to use to start the transient stability assessment agent.

[0004] China's invention patent disclosure CN110609477A discloses a power system transient stability judgment system and method based on deep learning, wherein a generative adversarial network (GAN) is used to make up for the insufficient amount of fault data and correct the problem of data imbalance, and a convolutional neural network (CNN) is used to perform transient stability assessment. China's invention patent disclosure CN109086913A discloses a power system transient stability assessment method and system based on deep learning, which uses a stacked denoising autoencoder (SAE) to extract multi-layer features from input data, and then uses a support vector machine (SVM) algorithm to perform transient stability assessment. China's invention patent disclosure CN109783997A discloses a power system transient stability assessment method based on a deep neural network, which uses the electrical measurement parameters of important buses and generators obtained by a wide area measurement system (WAMS) as input, uses a long short-term memory neural network (LSTM) and an improved convolutional neural network to extract data features, and then imports a lightweight gradient boosting machine to perform transient stability assessment. China's invention patent disclosure CN110336277A discloses a method for evaluating transient stability of power systems based on deep belief networks, and proposes a method for evaluating transient stability of power systems based on deep belief networks (DBNs). The post-fault feature quantity is extracted according to the maximum phase angle difference of the power system bus phase angle during the swing process, and the feature is input into the deep belief network to achieve rapid evaluation of transient stability of the power system. China's invention patent disclosure CN107846012A discloses a method for evaluating transient stability of power systems based on deep learning technology, which uses a stacked autoencoder (SAE) to extract features from feature variable vectors layer by layer, and then uses a convolutional neural network (CNN) to build a stability classification model. China's invention patent disclosure CN110889255A discloses a method for evaluating transient stability of power systems based on cascaded deep forests, and uses a cascaded deep forest algorithm to build a transient stability evaluation model, which has the advantage of strong small sample learning ability.

[0005] Although the above six patent documents have applied deep learning algorithms to the field of transient stability assessment of power systems, none of them takes the topological information of the power system into consideration, resulting in limited accuracy of the analysis results. China Invention Patent Publication Specification CN111914484A discloses a recursive graph convolutional network system for transient stability assessment of power grids, which uses a spatial domain graph convolutional neural network (GCN) constructed based on the power system admittance matrix as a feature extractor, cascaded with a time series extractor containing a long short-term memory network (LSTM), and a time pooling layer is designed after the time series extractor to aggregate the prediction results of multiple time steps to form a recursive deep graph convolution model suitable for power grid stability assessment. China Invention Patent Publication Specification CN111914486A discloses a method for transient stability assessment of power systems based on a graph attention network, which uses an attention coefficient to replace the Laplacian matrix of a spatial domain graph convolutional neural network, and proposes a transient stability assessment method based on a graph attention network. Although the above two documents take into account the topological structure of the power system, the input data of the artificial intelligence algorithm requires the measurement data of all nodes in the system topology. However, in actual power systems, only some nodes are equipped with synchronized phasor measurement units (PMUs), which cannot provide all the measurement data required for transient stability analysis, and therefore face difficulties in practical applications.

[0006] In addition, existing research rarely focuses on the framework design of the online operation of the transient stability assessment agent, and in practical applications, no design is made on when and what criteria to use to start the transient stability assessment agent. Summary of the invention

[0007] An embodiment of the present invention provides a method for evaluating transient stability of a power system based on information entropy and a spectrum graph network, which is used to solve the problems existing in the prior art.

[0008] In order to achieve the above object, the present invention adopts the following technical scheme.

[0009] A method for transient stability assessment of power system based on information entropy and spectrum graph network, including an offline training process and an online assessment process;

[0010] The offline training process includes:

[0011] Constructing multiple sets of training data sets with different time window lengths based on historical operating data of the target power system and / or operating data generated by a simulation model;

[0012] According to the time sequence of the time window of the training data set, the training data set is input into a transient stability training sub-agent in sequence; the output result obtained by the transient stability training sub-agent is used to train and iteratively update the transient stability assessment sub-agent in the online assessment process;

[0013] The online assessment process includes:

[0014] Based on the measured data during the operation of the power system, a startup judgment is made through a startup criterion device based on information entropy. If the judgment result is a startup criterion, multiple groups of data tensors with different time window lengths are generated;

[0015] According to the time sequence of the time windows of the multiple groups of data tensors, the data tensors are sequentially input into a transient stability assessment sub-agent to obtain an assessment result;

[0016] Based on the evaluation results, a comprehensive judgment is made through decision logic to obtain the transient stability judgment result of the power system;

[0017] The transient stability training sub-agent and the transient stability assessment sub-agent are constructed based on spectral graph convolution and Chebyshev polynomial expansion methods.

[0018] Preferably, the transient stability training sub-agent and the transient stability assessment sub-agent include an input layer, a hidden layer, a pooling layer and an output layer connected in sequence; the hidden layer includes three single-layer graph neural networks connected in sequence, and the output of each single-layer graph neural network is activated by a Relu activation function; each single-layer graph neural network is constructed by the following process:

[0019] Define a graph G = (V, E, W) with n nodes and m edges, and get the calculation formula

[0020] L=ID -1 / 2 WD -1 / 2 (1)

[0021] In the formula, V and E are the node set and edge set of G respectively, W is the adjacency matrix, L is the normalized Laplace matrix, I is the identity matrix, and D is the diagonal matrix, whose diagonal elements are

[0022] D ii =∑ j W ij (2)

[0023] In the formula, i and j represent the node numbers, D ii Represents the sum of the weights of the edges connected to node i;

[0024] Define the eigenvector matrix of the normalized Laplace matrix L as U = [u1,u2,...,u n ], the eigenvalue matrix is ​​Λ=diag(λ1,λ2,...,λ n ), L=UΛU T , the Fourier transform of the graph is defined as

[0025]

[0026] In the formula, is the original signal x, after the orthogonal basis U T The expression projected onto the spectral domain, diag represents the diagonal matrix, u represents the eigenvector of the Laplace transform, and λ represents the corresponding eigenvalue;

[0027] Define the graph convolution operator as * G , the matrix dot product operator is ☉, and the spectrum graph convolution of two signals x and y is defined as the calculation formula

[0028] x* G y=U((U T x)☉(U T y)) (4);

[0029] For the original signal x, U T y=[θ0,θ1,...,θ n-1 ] T The overall definition is the convolution kernel in the spectral domain. The training parameters have n vectors θ, defining the diagonal matrix g θ =diag([θ0,θ1,...,θ n-1 ]) is the graph convolution kernel, and the graph convolution calculation formula in the spectral domain is obtained

[0030] x* G y=Ug θ U T x (5);

[0031] By using the Chebyshev polynomial expansion

[0032]

[0033] Substitute into formula (5) for the graph convolution kernel g θ Perform approximate operations to obtain the spectrum graph convolution formula

[0034]

[0035] In the formula, β k are the coefficients corresponding to the k-th order Chebyshev polynomial.

[0036] Preferably, in the process of generating the operation data by the simulation model:

[0037] The characteristic vectors of the power system input to the simulation model include the power angle, rotor speed, voltage amplitude and current value corresponding to the q axis in the dq coordinate system of each generator node;

[0038] Based on the input power system characteristic vector, the transient stability index is calculated by

[0039]

[0040] Get the stability calibration result of the operating data; where Δδ max is the maximum relative power angle difference between any two generators in the simulation process. When TSI>0, the sample of the operating data is determined to be stable. When TSI<0, the sample of the operating data is determined to be unstable.

[0041] Preferably, there are 4 groups of training data sets with different time window lengths, and the time windows of the four groups of training data sets are [t0, t0+0.05s], [t0, t0+0.10s], [t0, t0+0.15s] and [t0, t0+0.20s] respectively; there are 4 transient stability training sub-agents, and each group of training data sets is input into a transient stability training sub-agent.

[0042] Preferably, the historical operating data of the target power system and / or the operating data generated by the simulation model include the power angle, rotor speed, voltage amplitude and current value corresponding to the q-axis in the dq coordinate system of each generator node;

[0043] The offline training process also includes:

[0044]

[0045] Sort and process the operation data;

[0046] Pass-through

[0047]

[0048] The sorted running data were normalized.

[0049] Preferably, based on the measured data during the operation of the power system, the start judgment is performed by a start criterion device based on information entropy, including:

[0050] Pass-through

[0051]

[0052] Calculate the information entropy value of the measured data;

[0053] Pass-through

[0054] |H(I q )-h i |>0.1 or|H(|V|)-h v |>0.1 (12)

[0055] The information entropy value of the measured data is compared with the preset threshold. When the absolute value of the difference between any information entropy value calculated and the constant value under steady state is greater than 0.1, the start criterion is determined; where hi represents the information entropy value corresponding to the q-axis current in steady state, h v It represents the information entropy value corresponding to the voltage in the steady state.

[0056] Preferably, there are 4 groups of data tensors with different time window lengths, and the time windows of the four groups of data tensors are [t0, t0+0.05s], [t0, t0+0.10s], [t0, t0+0.15s] and [t0, t0+0.20s] respectively; each group of data tensors includes the power angle, rotor speed, voltage amplitude and current value corresponding to the q axis in the dq coordinate system of each generator node;

[0057] There are four transient stability assessment sub-agents. According to the time sequence of the time windows of the four groups of data tensors, the data tensors are input into one transient stability assessment sub-agent in sequence to obtain the assessment results.

[0058] Preferably, based on the evaluation results, a comprehensive judgment is made through decision logic to obtain a judgment result of transient stability of the power system, including:

[0059] When the four groups of transient stability assessment sub-agents continuously output instability judgment or alternately output instability assessment, the result of judging the transient stability of the power system is instability.

[0060] It can be seen from the technical solutions provided by the above embodiments of the present invention that the present invention provides a method for evaluating transient stability of a power system based on information entropy and a spectrum graph network, and designs a method for analyzing transient stability of a power system based on information entropy and spectrum graph convolution technology. Spectral graph convolution technology belongs to an artificial intelligence method, which avoids the problem of slow time domain simulation calculation speed and overcomes the problem of strong conservatism and modeling difficulties of methods such as energy equations. The use of spectrum graph neural network technology overcomes the problem that technologies such as deep belief networks, stacked autoencoders, and convolutional neural networks cannot utilize the topological structure information of the power system. In view of the problem that the convolution kernel of the spectrum graph convolution is too large, Chebyshev polynomial expansion is used to approximate the graph convolution. The number of polynomial expansion terms K can be adjusted according to the number of nodes in the power system and the accuracy of the intelligent agent calculation. It solves the problem that methods such as GCN and GAT can only use first-order adjacent node information, and in actual systems, when the measurement nodes are limited, the accuracy decreases. An online operation framework for transient stability assessment intelligent agents based on information entropy and multi-segment spectrum graph convolution neural network technology is proposed. The framework solves the problems of input data frame positioning, agent startup criterion selection, and online transient stability assessment speed and reliability selection.

[0061] Additional aspects and advantages of the present invention will be given in part in the following description, which will become obvious from the following description, or may be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0063] Figure 1 A logical framework diagram of the power system transient stability assessment method based on information entropy and spectrum graph network provided by the present invention;

[0064] Figure 2 A schematic diagram of a sub-agent based on spectrum graph convolution of the power system transient stability assessment method based on information entropy and spectrum graph network provided by the present invention;

[0065] Figure 3 A schematic diagram of the IEEE 16-machine 68-node system structure of a test embodiment of the power system transient stability assessment method based on information entropy and spectrum graph network provided by the present invention;

[0066] Figure 4 A schematic diagram of a curve showing a change in information entropy during a fault process and a curve showing a change in the q-axis current value of a generator in a method for evaluating transient stability of a power system based on information entropy and a spectrum graph network provided by the present invention;

[0067] Figure 5 A schematic diagram of a curve showing a change in information entropy during a fault process and a curve showing a change in voltage amplitude at a generator node of a method for evaluating transient stability of a power system based on information entropy and a spectrum graph network provided by the present invention;

[0068] Figure 6 A schematic diagram showing the stability judgment training results of four spectral graph neural network sub-agents of the power system transient stability assessment method based on information entropy and spectral graph network provided by the present invention and their comparison with the convolutional neural network algorithm. DETAILED DESCRIPTION

[0069] The embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be interpreted as limiting the present invention.

[0070] It will be understood by those skilled in the art that, unless expressly stated, the singular forms "one", "said", and "the" used herein may also include plural forms. It should be further understood that the term "comprising" used in the specification of the present invention refers to the presence of the features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof. It should be understood that when we refer to an element as being "connected" or "coupled" to another element, it may be directly connected or coupled to the other element, or there may be intermediate elements. In addition, the "connection" or "coupling" used herein may include wireless connection or coupling. The term "and / or" used herein includes any unit and all combinations of one or more associated listed items.

[0071] It will be understood by those skilled in the art that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as those generally understood by those skilled in the art in the art to which the present invention belongs. It should also be understood that terms such as those defined in common dictionaries should be understood to have meanings consistent with the meanings in the context of the prior art, and will not be interpreted with idealized or overly formal meanings unless defined as herein.

[0072] To facilitate understanding of the embodiments of the present invention, several specific embodiments will be further explained below with reference to the accompanying drawings, and each embodiment does not constitute a limitation on the embodiments of the present invention.

[0073] The present invention provides a method for evaluating transient stability of a power system based on information entropy and spectrum graph network, which is used to solve the following technical problems:

[0074] (1) Existing transient stability analysis mainly uses time-domain simulation methods to perform numerical iterative calculations on the power system of the partial differential-algebraic equation model to obtain the state trajectory of the system after the fault, observe and judge the stability of the system after the fault. Although the existing methods are reliable and accurate, they are slow to calculate and time-consuming. As a result, in the online safety analysis of the power system, the number of examples that can be analyzed per unit time is limited. As the power system develops towards power electronics and AC / DC hybrid, the mathematical model of the power system becomes more complex, the iterative step size of the numerical calculation of the partial differential equation is further reduced, and the total amount of calculation for a single example is further increased. If the time-domain simulation method uses an accurate model, it will face the problem of explosive calculation; if the model is greatly simplified, although the calculation speed can be improved, it will sacrifice accuracy and reliability;

[0075] (2) The energy equation method is a power grid transient stability analysis method based on Lyapunov's direct method. The main idea is to construct an energy equation model for the system, and then analyze the stability of the system by comparing the critical energy that the system can withstand with the fault energy. First, this method also faces the problem of computational complexity, because when looking for the dominant unstable equilibrium point of the system, it is still necessary to obtain the operating trajectory of the system during the fault process through time domain simulation. Secondly, this method also faces the problem of difficult construction of energy equations, especially for complex AC and DC power systems. Finally, the energy equation method also faces the problem of overly conservative calculation results. These technical shortcomings lead to limited accuracy of the energy equation method;

[0076] (3) The extended equal area rule is another transient stability analysis method. This method divides the multi-machine system of the power grid into two groups after the disturbance, calculates the stability energy margin, and determines the stability of the system. This method faces technical defects similar to the energy equation method. The extended equal area rule is mainly used for multi-machine power grid systems based on synchronous generators. It is difficult to accurately perform model analysis on complex AC / DC hybrid power systems that include DC power sources with converters as interfaces, resulting in limited analysis accuracy in this type of method.

[0077] (4) Currently, a small number of studies have applied deep learning technology to transient stability analysis of power systems. These studies use the electrical parameters such as phase angle, frequency, voltage, active power, reactive power, fault information, etc. of the transient process of the power system generated by the simulation model and the stability judgment results to form the training samples of the deep learning method, and obtain the transient stability assessment agent by offline training of the deep learning classifier. Existing deep learning technologies such as deep belief networks, stacked autoencoders, and convolutional neural networks use the tensor constructed by the system operation status obtained by the measurement of the power system as input, and perform convolution, pooling and other operations based on the arrangement structure of the tensor in the neural network. The adjacent relationship and aggregation process of the elements do not consider the important characteristic of the topological structure of the power system itself. In the application of graph neural network technology, some papers have studied the graph attention neural network (GAT) and message passing network (MPNN) evolved from the spatial domain graph convolutional neural network (GCN). GAT uses the attention coefficient to replace the Laplacian matrix in GCN to aggregate the information of adjacent nodes, and the MPNN technology considers the characteristics of the edges in the topology. Methods such as GCN, GAT, and MPNN all aggregate information based on the first-order neighboring nodes of each node in the topological structure, and assume that when the system stability can be evaluated, the electrical measurement parameter information of all nodes in the power system can be obtained. However, in actual systems, the number of nodes that can provide real-time measurement information during transient stability assessment is limited, and the phase angle, voltage and other parameters of most nodes cannot be obtained, resulting in a large number of nodes in the topology without input information. Therefore, it is difficult to update the parameters of each node in the neural network and provide effective message transmission by only aggregating the node information of the first-order neighborhood. In addition, existing artificial intelligence-based methods all assume that the transient stability assessment agent can obtain the system operating parameters within a specific time interval after the fault occurs (0.1 seconds or 0.2 seconds after the fault occurs). However, the existing artificial intelligence-based technical framework does not propose how to accurately and quickly locate the data input box according to the time of the fault occurrence in the continuous measurement data time series when the system fails. It does not propose how to use appropriate criteria to start the transient stability analysis agent. Finally, many existing studies have not considered the impact of the signal-to-noise ratio on the results during the transmission of the measured electrical quantity.

[0078] See also Figure 1 The present invention provides a method for evaluating transient stability of a power system based on information entropy and spectrum graph network, including an offline training process and an online evaluation process.

[0079] The offline training process includes:

[0080] Constructing multiple sets of training data sets with different time window lengths based on historical operating data of the target power system and / or operating data generated by a simulation model;

[0081] According to the timing of the time window of the training data set, the training data set is input into a transient stability training sub-agent in sequence; the output results obtained by the transient stability training sub-agent are used to train and iteratively update the transient stability assessment sub-agent in the online assessment process.

[0082] The online assessment process includes:

[0083] Based on the measured data during the operation of the power system, a startup judgment is made through a startup criterion device based on information entropy. If the judgment result is a startup criterion, multiple groups of data tensors with different time window lengths are generated;

[0084] According to the time sequence of the time windows of the multiple groups of data tensors, the data tensors are sequentially input into a transient stability assessment sub-agent to obtain an assessment result;

[0085] Based on the evaluation results, a comprehensive judgment is made through decision logic to obtain the transient stability judgment result of the power system.

[0086] In the preferred embodiment provided by the present invention, the structures of the transient stability training sub-agent and the transient stability assessment sub-agent are the same. The core of the overall technical route is the spectrum graph convolution technology, and the Chebyshev polynomial expansion is used for regional approximation processing. The specific structural block diagram is as follows: Figure 2 shown.

[0087] Each sub-agent structure includes three parts: input layer, hidden layer, pooling and output layer. The number of nodes in the input layer is consistent with the number of nodes in the object system of transient stability assessment, and can receive the input measurement vector of each node. The pooling layer connects the output of the hidden layer of the graph neural network with the fully connected network of the output layer. The output layer contains two fully connected layers and an export Softmax selector. The design of the input and output parts adopts the conventional neural network design in the existing transient stability assessment method based on artificial intelligence. The difference between the transient stability assessment method in the present invention and the existing method in terms of agent design lies in the hidden layer design scheme, which is as follows:

[0088] The hidden layer design uses spectral graph convolution technology. Define a graph G = (V, E, W) with n nodes and m edges. V and E are the node set and edge set of G respectively, and W is the adjacency matrix. Then,

[0089] L=ID -1 / 2 WD -1 / 2 , (1)

[0090] Where L is the normalized Laplace matrix, I is the identity matrix, and D is a diagonal matrix whose diagonal elements are,

[0091] D ii =∑ j Wij . (2)

[0092] Among them, i and j represent the node numbers, D ii Represents the sum of the weights of the edges connected to node i.

[0093] Define the eigenvector matrix of the normalized Laplace matrix L as U = [u1,u2,...,u n ], the eigenvalue matrix is ​​Λ=diag(λ1,λ2,...,λ n ), L=UΛU T , the Fourier transform of the graph is defined as,

[0094]

[0095] In the formula is the original signal x, after the orthogonal basis U T , the expression projected onto the spectral domain; diag represents the diagonal matrix, u represents the eigenvector of the Laplace transform, and λ represents the corresponding eigenvalue.

[0096] Define the graph convolution operator as “* G ", the matrix dot product operator is "☉", then the spectrum graph convolution of two signals x and y can be defined as,

[0097] x* G y=U((U T x)☉(U T y)). (4)

[0098] Therefore, in actual signal processing, for the input signal x, U T y=[θ0,θ1,...,θ n-1 ] T The whole can be regarded as a convolution kernel in the spectral domain, that is, the training parameters of the model. From this, we can define the diagonal matrix θ to represent the free parameters and define g θ =diag([θ0,θ1,...,θ n-1 ]) is the graph convolution kernel, then the graph convolution (Spectral graph CNN) in the spectral domain is obtained as follows:

[0099] x* G y=Ug θ U T x. (5)

[0100] In this form of graph convolution, the information of all nodes on the graph needs to be aggregated during the convolution process of each node on the graph, resulting in low computational efficiency for large networks. In addition, in the transient stability assessment of power systems, the reference value of node information that is too far away is limited. Therefore, the Chebyshev polynomial expansion is used here to calculate the graph convolution kernel g.θ Approximation. The size of the node information aggregation range can be changed by adjusting the number of expansion terms K retained in the operation. And the optimal K value is selected according to the actual network topology and evaluation performance feedback results. The Chebyshev expansion of the graph convolution kernel is,

[0101]

[0102] Substituting equation (6) into equation (5), we get the spectrum graph convolution equation approximated by Chebyshev expansion:

[0103]

[0104] In formula (6), β k are the coefficients corresponding to the k-th order Chebyshev polynomial.

[0105] Formula (7) is the design scheme of the final hidden layer single-layer graph neural network. The hidden layer contains three single-layer graph neural networks. The output of each layer of the graph neural network is activated with the Relu activation function. The structure is as follows: Figure 2 shown.

[0106] Furthermore, the data sample set used for offline training of the transient stability assessment agent is obtained by time domain simulation of the simulation model of the power system being assessed. Graph neural networks are supervised learning, so each data sample in the data set needs to include the characteristic vector of the power system input to the agent and the stability calibration results of each sample used to calculate the loss function during training. In the design of this patent, the characteristic vectors of the power system are the power angle, rotor speed, voltage amplitude and current value corresponding to the q-axis in the dq coordinate system of each generator node in the system. The stability calibration results of each sample are obtained by using the transient stability index (TSI) as the stability criterion, by performing a time domain simulation of each sample for at least 15 seconds and making a judgment. The expression of TSI is,

[0107]

[0108] Where Δδ max is the maximum relative power angle difference between any two generators during the simulation. When TSI>0, the sample is considered stable, otherwise it is considered unstable.

[0109] exist Figure 1In the design scheme of the transient stability assessment framework for power systems described in , the offline training part contains four sub-agents. The internal structure of each sub-agent is the same, and the design scheme is shown in 4.2, but the length of the input data window used by each sub-agent is different. Let the time of fault occurrence be t0, and the input data windows corresponding to the four sub-agents are [t0, t0+0.05s], [t0, t0+0.10s], [t0, t0+0.15s], [t0, t0+0.20s] (corresponding to Figure 1 t1~t4 in ). The four sub-agents are trained with different input data in offline training. In the online evaluation, the evaluation results are output in sequence according to the order of the obtained data, and the final evaluation decision is made by the evaluation result comprehensive decision module. Therefore, in the part of compiling the training set with different time window lengths, the power angle, rotor speed, voltage amplitude of each generator node under different simulation fault samples obtained by time domain simulation and the current value corresponding to the q axis in the dq coordinate system are sorted into four sample sets according to the above four lengths of time windows as the training sets of the four sub-agents.

[0110] In the training set, the measurement data in each sample needs to be standardized and sorted. The sampling interval of the four parameters of each generator power angle, speed, voltage and current is 0.01s. Taking the [t0, t0+0.05s] time window as an example, the measurement parameters include six sampling points in total. in Corresponding to the sampling value at the time of fault occurrence t0, Corresponding to the moment t0+0.05s, that is, the sampling value 0.05s after the fault occurs. The normalized expression of the measured sampling value at the moment t is:

[0111]

[0112] The number of input signal channels of the spectral graph convolution agent is consistent with the number of nodes of the evaluation object system, and each node corresponds to an input signal channel. The input vector of the generator node also includes the power angle Speed Voltage Amplitude and q-axis current Four vectors, sorted as follows The other nodes are normalized by substituting the steady-state measurement values ​​into formula (10) and then sent to the spectral graph neural network through the data input channel of each node as supplementary input.

[0113] The training of the spectral graph neural network is carried out using the back propagation and gradient descent method. The direction of descent is the direction in which the corresponding network parameters reduce the loss function value, and momentum is added during the descent process to avoid falling into the local minimum and improve the computational efficiency. The loss function is the cross entropy function. In the specific program implementation, the tf.train.MomentumOptimizer function and the sparse_softmax_cross_entropy_with_logits function in TensorFlow are used.

[0114] During online operation, the synchronous phasor unit (PMU) based on the satellite clock signal transmits the measured operating status of each generator node in the system to the startup criterion module. The startup criterion model uses the generator q-axis current value and voltage amplitude at time t of each generator node as input features, calculates the information entropy value, and compares it with the setting threshold. The calculation formula of information entropy is:

[0115]

[0116] Where X=[x1,x2,…,x i ,…x n ] is the discretized generator q-axis current value or voltage amplitude, P(x1), P(x2),…, P(x i ),…P(x n ) is the occurrence probability corresponding to each discrete variable, and H(X) is the information entropy value. The startup criterion based on information entropy is

[0117] |H(I q )-h i |>0.1 or|H(|V|)-h v |>0.1 (12)

[0118] In the formula, h i represents the information entropy value corresponding to the q-axis current in steady state, h v It represents the information entropy value corresponding to the voltage in steady state. When a disturbance caused by a fault occurs in the system, the two information entropy values ​​H(I q ) and H(|V|) change relative to the information entropy value in the steady state. According to formula (11), when the absolute value of the difference between any of the two information entropy values ​​and the fixed value in the steady state exceeds 0.1, the transient stability intelligent assessment is started.

[0119] According to the proposed transient stability assessment framework design scheme, as shown in Figure (1), after the judgment is started, the system will generate tensors from the real-time data obtained through the PMU and distribute them to each sub-agent. Speed Voltage Amplitude and q-axis current The length of the vector varies according to the length of the time window input by each sub-agent, and the time window lengths are [t0, t0+0.05s], [t0, t0+0.10s], [t0, t0+0.15s], [t0, t0+0.20s] respectively. The tensor with the shortest time window will be generated the fastest and sent to sub-agent 1 first. The issuance time of subsequent tensors will be delayed as the length of the time window increases, and they will be sent to sub-agents 2, 3, and 4 respectively. After receiving the input tensor, the sub-agent uses the spectrum graph neural network structure designed by this patent to perform calculations and judgments.

[0120] In the preferred embodiment provided by the present invention, the logic design of the sub-agent result collection and comprehensive decision output work is as follows:

[0121] Since each sub-agent receives the input tensor at different times, after taking into account the calculation delay, sub-agents 1 to 4 output the stability judgment results in turn. The longer the time window of the sub-agent input tensor, the higher the accuracy of the output stability judgment. In actual calculations, the input tensors of each algorithm use the PMU measurement data of all generator nodes, rather than the measurement data of all bus nodes. The graph neural network consists of three hidden layers. The Chebyshev polynomial expansion parameters K of each layer of graph spectrum convolution are 4, 3, and 2 respectively. Under the above parameters, the stability judgment results of the sub-agents are collected by the comprehensive decision module in the order in which the results are generated. In the comprehensive decision module, only when the "stable" or "instability" judgment results output by two consecutive sub-agents of a certain sample are collected continuously, the comprehensive decision module will output the "stable" or "instability" result, that is, the two consecutive judgment results are consistent. If the results output by the four sub-agents alternate continuously, such as "stable", "instability", "stable", and "instability", the judgment results appear. Then, following the conservative principle, the comprehensive decision module outputs the result of "instability".

[0122] The present invention also provides an embodiment for exemplarily showing the process and effect of evaluating the transient stability of a power system by applying the method provided by the present invention.

[0123] The following takes the IEEE 16-machine 68-node system as an example. The system structure diagram is as follows: Figure 3 When a two-phase short circuit fault with a duration of 0.719s occurs in the connection line between node 19 and node 20, the q-axis current value and voltage amplitude of the generator node and the corresponding information entropy value are as follows: Figure 4 , Figure 5 shown.

[0124] from Figure 4 , Figure 5It can be seen that at the moment of fault occurrence, with the change of the generator q-axis current and node voltage amplitude, the information entropy value has a significant jump at the same time, exceeding the setting threshold and correctly triggering the startup criterion.

[0125] The effectiveness of the startup criterion based on information entropy has also been verified in a large number of sample experiments. Taking the IEEE16-machine 68-node system as an example, the fault type is randomly selected from the six faults of three-phase fault, single-phase ground fault, two-phase short circuit fault, two-phase ground short circuit fault, and node load loss; the fault location is randomly selected from the lines and nodes in the 68-node system; the fault removal time is randomly selected from the Gaussian distribution with a mean of 0.1s and a variance of 0.0334s. A total of 30,000 test samples are generated, of which 8,508 are unstable samples. The two information entropy criteria based on the generator q-axis current value and the node voltage amplitude can start correctly on all samples.

[0126] According to the proposed transient stability assessment framework design scheme, Figure 1 After the judgment is started, the system will generate tensors from the real-time data obtained through PMU and distribute them to each sub-agent. Speed Voltage Amplitude and q-axis current The vector length varies according to the length of the time window input by each sub-agent, and the time window lengths are [t0, t0+0.05s], [t0, t0+0.10s], [t0, t0+0.15s], [t0, t0+0.20s] respectively. The tensor with the shortest time window will be generated the fastest and sent to sub-agent 1 first. The issuance time of subsequent tensors will be delayed as the time window length increases, and they will be sent to sub-agents 2, 3, and 4 respectively. After receiving the input tensor, the sub-agent uses the neural network structure designed in 4.2 to perform calculations and judgments.

[0127] Each sub-agent receives the input tensor in sequence. After taking into account the calculation delay, sub-agents 1 to 4 output the stability judgment results in turn. The longer the time window of the sub-agent input tensor, the higher the accuracy of the output stability judgment. In the comparative verification of the example, for the test of the above 30,000 samples in the IEEE 16-machine 68-node system, the input tensors of each algorithm use the PMU measurement data of all 16 generator nodes, rather than the measurement data of all 68 bus nodes. In the input data, Gaussian white noise with a signal-to-noise ratio of 40dB is added using the MATLAB awgn function. The graph neural network consists of three hidden layers. The Chebyshev polynomial expansion parameters K of each layer of graph spectrum convolution are 4, 3, and 2 respectively. Under the above parameters, the stability judgment accuracy of sub-agents 1, 2, 3, and 4 are 92.4%, 94.8%, 96.5%, and 98.6%, respectively. The above stability judgment results are collected by the comprehensive decision module in the order in which the results are generated. In the comprehensive decision-making module, the comprehensive decision-making module will output the result of "stable" or "instability" only after the "stable" or "instability" judgment results output by two consecutive sub-agents of a certain sample are collected continuously, that is, the two consecutive judgment results are consistent. If the results output by the four sub-agents alternate continuously, such as "stable", "instability", "stable", and "instability", the output result of the comprehensive decision-making module is "instability" in accordance with the conservative principle. Table 1 shows the comparison of the results of the Chebyshev spectral graph convolution method proposed in this patent and other artificial intelligence methods in transient stability analysis. Among them, the convolutional neural network method and the long short-term memory neural network method both use the longest time window of [t0, t0+0.20s], and the graph convolutional neural network (GCN) method uses 4 time windows from short to long, which is consistent with the Chebyshev spectral graph convolution method proposed in this patent. Figure 6 The process of gradually improving the accuracy of four sub-agents based on the Chebyshev spectral graph convolution method (CheNet) and four sub-agents based on the graph convolution neural network (GCN) during 100 generations of training is shown. It can be seen from the figure that the Chebyshev spectral graph convolution method has a higher accuracy and a faster convergence speed. The last column is the final output result of the comprehensive decision model. After passing through the comprehensive decision module, the output stability judgment result is the final output result of this evaluation method. After testing in the above-mentioned IEEE 16-machine 68-node system, the stability judgment accuracy of this discrimination method reached 99.2%. This result verifies the accuracy and effectiveness of the power system transient stability assessment method based on information entropy and spectral graph convolution protected by this patent. In terms of computing time, when this method runs on an Intel(R) Core(TM) i7-10750H CPU@2.60GHz processor, the average computing time for sample stability judgment is only 0.016 seconds.

[0128] Table (1) Comparison between the method proposed in this patent and other artificial intelligence methods

[0129]

[0130]

[0131] In summary, the present invention provides a method for transient stability assessment of power system based on information entropy and spectrum graph network, including an offline training part and an online assessment part. In the offline training part, the historical operation data of the target power system to be analyzed or the data generated by the simulation model are used to form tensor data sets of different length time windows to train four transient stability assessment sub-agents based on spectrum graph convolution respectively, and the trained sub-agents are used to iteratively update the sub-agents used for online assessment. In the online assessment stage, the measurement parameters collected by the wide-area phasor measurement equipment during the operation of the system are sent to the start criterion based on information entropy for judgment. When the calculated information entropy exceeds the set threshold, a start instruction is issued, and the data moment of triggering the start criterion is used as the starting point of the input data window. The input data tensor generation module is respectively made into multiple groups of data tensors according to the time sequence of data input, and is successively distributed to the four sub-agents for transient stability assessment. After the judgment results of the four sub-agents are collected successively, the final transient stability judgment result is output according to the design decision logic.

[0132] The method provided by the present invention designs an online transient stability assessment scheme for power systems based on information entropy and spectrum graph convolution technology. The scheme uses the q-axis current and voltage of each generator node to calculate the information entropy respectively, and compares it with the preset setting threshold value as the starting criterion of the transient stability assessment agent and the starting point of the input data window. The problem of locating the starting time of the transient stability assessment agent and the input data window is solved. The transient stability intelligent assessment framework design scheme consists of multiple sub-agents based on spectrum graph convolution technology, and different sub-agents have different input data window lengths and different response speeds. The transient stability assessment results obtained by integrating different sub-agents are improved. The accuracy and speed of online transient stability analysis are improved. A transient stability assessment method based on spectrum graph convolution neural network technology is proposed, which not only utilizes the important characteristic of the topological structure of the power system, but also overcomes the limitation that the existing spatial domain graph convolution method must use the measurement parameters of all nodes. Only the measurement parameters of some nodes are required to effectively aggregate information. By using the Chebyshev polynomial expansion to approximate the convolution kernel, the computational efficiency problem of spectral convolution is solved, and the scope of information aggregation can be controlled according to the algorithm performance and system topology size. Finally, in the verification experiment, noise designed according to the signal-to-noise ratio of the actual power system wide-area phasor measurement device was added to the input data to verify the anti-interference ability of this scheme. In addition, it has the following advantages:

[0133] 1. The information entropy is calculated using the q-axis current and voltage of each generator node, and compared with the preset setting threshold value as the starting criterion of the transient stability assessment agent and the starting point of the input data window. This solves the problem of locating the starting time of the transient stability assessment agent and the input data window.

[0134] 2. The transient stability intelligent assessment framework design consists of multiple sub-agents based on spectrum graph convolution technology. Different sub-agents have different input data window lengths and response speeds. The transient stability assessment results obtained by combining different sub-agents improve the accuracy and speed of online transient stability analysis.

[0135] 3. A transient stability assessment method based on spectral graph convolutional neural network technology is proposed. It not only utilizes the important characteristic of the power system topology structure, but also overcomes the limitation that the existing spatial domain graph convolution method must use the measurement parameters of all nodes. Only the measurement parameters of some nodes are needed to effectively aggregate information.

[0136] 4. By using the Chebyshev polynomial expansion to approximate the convolution kernel, the computational efficiency problem of spectral convolution is solved, and the scope of information aggregation can be controlled according to the algorithm performance and system topology size.

[0137] Those skilled in the art can understand that the accompanying drawings are only schematic diagrams of an embodiment, and the modules or processes in the accompanying drawings are not necessarily required to implement the present invention.

[0138] It can be known from the description of the above implementation methods that those skilled in the art can clearly understand that the present invention can be implemented by means of software plus a necessary general hardware platform. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which can be stored in a storage medium such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in the various embodiments of the present invention or certain parts of the embodiments.

[0139] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device or system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. The device and system embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0140] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. Power system transient stability assessment method based on information entropy and spectrum graph network, It is characterized in that Includes offline training process and online evaluation process; The offline training process includes: Constructing multiple sets of training data sets with different time window lengths based on historical operating data of the target power system and / or operating data generated by a simulation model; According to the time sequence of the time window of the training data set, the training data set is sequentially input into a transient stability training sub-agent; the output result obtained by the transient stability training sub-agent is used to train and iteratively update the transient stability assessment sub-agent in the online assessment process; The online assessment process includes: Based on the measured data during the operation of the power system, the startup judgment is performed through the startup criterion device based on information entropy. If the judgment result is a startup criterion, multiple groups of data tensors with different time window lengths are generated; including: Pass-through Calculate the information entropy value of the measured data; Pass-through |H(I q )-h i |>0.1 or|H(|V|)-h v |>0.1 (12) The information entropy value of the measured data is compared with the preset threshold. When the absolute value of the difference between any information entropy value calculated and the constant value under steady state is greater than 0.1, the start criterion is determined; where h i represents the information entropy value corresponding to the q-axis current in steady state, h v It represents the information entropy value corresponding to the voltage in steady state; According to the time sequence of the time windows of the multiple groups of data tensors, the data tensors are sequentially input into a transient stability assessment sub-agent to obtain an assessment result; Based on the evaluation results, a comprehensive judgment is made through decision logic to obtain the transient stability judgment result of the power system; The transient stability training sub-agent and the transient stability assessment sub-agent are constructed based on spectral graph convolution and Chebyshev polynomial expansion methods.

2. The method according to claim 1, It is characterized in that The transient stability training sub-agent and the transient stability assessment sub-agent include an input layer, a hidden layer, a pooling layer, and an output layer connected in sequence; the hidden layer includes three single-layer graph neural networks connected in sequence, and the output of each single-layer graph neural network is activated by a Relu activation function; each single-layer graph neural network is constructed by the following process: Define a graph G = (V, E, W) with n nodes and m edges, and get the calculation formula L=I-D -1 / 2 WD -1 / 2 (1), Where V and E are the node set and edge set of G, respectively, W is the adjacency matrix, L is the normalized Laplace matrix, I is the identity matrix, and D is the diagonal matrix whose diagonal elements are D ii =∑ j W ij (2), In the formula, i and j represent the node numbers, D ii Represents the sum of the weights of the edges connected to node i; Define the eigenvector matrix of the normalized Laplace matrix L as U = [u 1 ,u 2 ,...,u n ], the eigenvalue matrix is ​​Λ=diag(λ 1 ,λ 2 ,...,λ n ), L=UΛU T , the Fourier transform of the graph is defined as In the formula, is the original signal x, after the orthogonal basis U T The expression projected onto the spectral domain, diag represents the diagonal matrix, u represents the eigenvector of the Laplace transform, and λ represents the corresponding eigenvalue; Define the graph convolution operator as *G, the matrix dot product operator as ☉, and define the spectrum graph convolution of two signals x and y as the calculation formula x* G y=U((U T x)☉(U T y)) (4); For the original signal x, U T y=[θ 0 ,θ 1 ,...,θ n-1 ] T The overall definition is the convolution kernel in the spectral domain. The training parameters have n vectors θ, defining the diagonal matrix g θ =diag([θ 0 ,θ 1 ,...,θ n-1 ]) is the graph convolution kernel, and the graph convolution calculation formula in the spectral domain is obtained x* G y=Ug θ U T x (5); By using the Chebyshev polynomial expansion Substitute into formula (5) for the graph convolution kernel g θ Perform approximate operations to obtain the spectrum graph convolution formula In the formula, β k are the coefficients corresponding to the k-th order Chebyshev polynomial.

3. The method according to claim 1, It is characterized in that In the process of generating operation data through the simulation model: The characteristic vectors of the power system input to the simulation model include the power angle, rotor speed, voltage amplitude and current value corresponding to the q axis in the dq coordinate system of each generator node; Based on the input power system characteristic vector, the transient stability index is calculated by Get the stability calibration result of the operating data; where Δδ max is the maximum relative power angle difference between any two generators in the simulation process. When TSI>0, the sample of the operating data is determined to be stable. When TSI<0, the sample of the operating data is determined to be unstable.

4. The method according to claim 1, It is characterized in that There are 4 sets of training data sets with different time window lengths, and the time windows of the four sets of training data sets are [t 0 ,t 0 +0.05s], [t 0 ,t 0 +0.10s], [t 0 ,t 0 +0.15s] and [t 0 ,t 0 +0.20s]; there are 4 transient stability training sub-agents, and each group of the training data sets is input into one of the transient stability training sub-agents.

5. The method according to claim 4, It is characterized in that The historical operation data of the target power system and / or the operation data generated by the simulation model include the power angle, rotor speed, voltage amplitude and current value corresponding to the q axis in the dq coordinate system of each generator node; The offline training process also includes: Sorting the operation data; Pass-through The sorted operation data are normalized.

6. The method according to claim 1, It is characterized in that There are 4 groups of data tensors with different time window lengths, and the time windows of the four groups of data tensors are [t 0 ,t 0 +0.05s], [t 0 ,t 0 +0.10s], [t 0 ,t 0 +0.15s] and [t 0 ,t 0 +0.20s]; each set of data tensors includes the power angle, rotor speed, voltage amplitude and current value corresponding to the q axis in the dq coordinate system of each generator node; There are four transient stability assessment sub-agents. According to the time sequence of the time windows of the four groups of data tensors, the data tensors are input into one transient stability assessment sub-agent in sequence to obtain the assessment results.

7. The method according to claim 6, It is characterized in that The above-mentioned comprehensive judgment based on the evaluation result and the decision logic to obtain the transient stability judgment result of the power system includes: When the four groups of transient stability assessment sub-agents continuously output instability judgment or alternately output instability assessment, the result of judging the transient stability of the power system is instability.

Images