A method of power system situation awareness based on Chebyshev graph neural network
Through the Chebishev graph neural network method and combined with multi-attribute decision-making, the integrity and progressiveness of situation awareness of the power system are achieved, the problem of situation understanding and situation prediction in the existing technology is solved, and large-scale development is improved The situational understanding and prediction capabilities of complex power systems.
Patent Information
- Application Number
- CN202411435006.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-15
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2044-10-15
AI Technical Summary
The existing situational awareness methods of power systems are carried out in a split manner, lacking integrity and progressiveness, and cannot effectively combine situational understanding and situation prediction. The method based on graph convolution network has high computational complexity on large-scale complex graphs, and cannot capture comprehensiveness. graph structure information.
Using the Chebischev graph neural network method, the operating parameters and adjacency matrix are obtained through the power system trend simulation, the graph data set is constructed, and the graph neural network model is built using the Chebischev graph convolutional layer, the model is trained to achieve situational understanding, and Build multi-dimensional risk situation evaluation indicators from the perspectives of topology, attackers and chain failures, and combine multi-attribute decision-making methods to predict risk situations.
It realizes the integrity and progressiveness of situational awareness in the power system, improves the accuracy and systematicity of situational prediction, reduces the computational complexity, can effectively process the global information of large-scale complex power systems, and adapt to the changing topological structure.
Smart Images

Figure CN118941098B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of power system situation awareness, and in particular to a power system situation awareness method based on a Chebyshev graph neural network. Background Art
[0002] As the scale of power systems gradually expands and the complexity gradually increases, ensuring the efficient, safe and stable operation of power systems has become one of the key issues of power infrastructure. The most basic of these is to achieve situational awareness of the power system. Only by understanding the current power system operation status and predicting future power system operation trends in real time during the operation of the power system can we guide the production, dispatching and operation strategies of power system managers, make up for the weak links in real time, and prevent future risk situations. This has important practical significance for preventing large-scale power outages in the power system.
[0003] With the development of smart grids, a large amount of operating data is generated in the system at any time. Traditional situational awareness technology based on manual analysis is difficult to handle such large and complex data, which has led to the application of situational awareness technology in power systems. Situational awareness refers to the ability to obtain a comprehensive understanding of the current state of the power system and accurately predict future development trends by collecting, monitoring and analyzing various data and information related to the operating status of the power system. It can be divided into three stages: situational awareness, situational understanding, and situational prediction.
[0004] Situation awareness focuses on obtaining real-time data from all aspects of the power system, and conducting preliminary integration and analysis to obtain a comprehensive understanding of the current state of the power system. Situation understanding is to further analyze and interpret the data on the basis of situation awareness to obtain an in-depth understanding of the overall state and interrelationships of the power system. Situation prediction is based on existing data and information, using established models and methods to predict and infer the future state and development trend of the power system. In the current power system, advanced data acquisition and sensing technology, data processing and analysis, data integration and other key technologies are fully capable of the task of situation awareness, so the current application of situation awareness in the power system mainly focuses on situation understanding and situation prediction. Situation understanding infers the current comprehensive and real operating state parameters of the system through a small amount of incomplete system quantity measurements with bad data, so as to perceive the real-time operating state of the current power system. Situation prediction can be understood as predicting the future operating trend of the system through the perceived system operating state, including weak link prediction, cascading failure prediction, etc. Situation awareness combines traditional static analysis methods with real-time monitoring and prediction of dynamic systems, bringing deeper insights and effective tools to the research and operation management of power systems.
[0005] At present, there are relatively few studies on power system situational awareness, and they are not comprehensive enough. At present, the perception of the operating status of the power system is mainly achieved through state estimation, that is, it is mainly based on algorithms, focusing on the physical state of the power system at a specific moment or operating state, and the data types and scope processed are relatively limited. The research focus is also on state measurement and estimation. Traditional power system state estimation methods are usually based on the least squares method. This method uses the data measured in the system to estimate the state of the power system by minimizing the sum of squares of the error between the measured value and the estimated value.
[0006] A study proposed a power system state estimation model based on graph neural network, using GCN (Graph Convolution Network) to build a graph neural network, taking system quantity measurements and connection matrices as input, and state quantities as output to train the graph neural network model, and finally inputting actual quantity measurements into the trained model to obtain the true state quantity. However, in situation awareness, situation understanding is to conduct a comprehensive analysis and interpretation of the real-time data and information of the power system after obtaining the data, so as to have a deeper understanding of the system's state, attributes, and operating characteristics. Therefore, situation awareness is at a higher level than power system state estimation in terms of time dimension expansion, breadth and depth of functions, real-time performance, and responsiveness.
[0007] Situation prediction is also an important part of situation awareness. Situation prediction mainly focuses on predicting the possible operating state and possible weak links of the future system based on the perception of the current system state, including risk assessment, vulnerability assessment, etc., so that grid managers can take avoidance measures in advance. This process needs to consider multiple factors, including changes in load demand, fluctuations in renewable energy generation, equipment failures and other external environmental factors. Therefore, the accuracy of situation prediction depends greatly on the accuracy of situation awareness.
[0008] In their research, Huang Yan et al. proposed that, under the premise of clarifying the current operating status, the transient energy function of the multi-machine power system based on the structure-preserving model be used to solve the potential energy of each branch, and combined with the directed electrical betweenness in the complex network, the operational vulnerability and structural vulnerability of the branch be calculated to identify the weak links that may exist in the system in the future.
[0009] In their research, Guo Chuangxin et al. proposed a graph neural network algorithm GNN-RL (Graph Neural Network-Reinforcement Learning) based on reinforcement learning, established a graph neural network model for extracting network features, and then input the extracted features into the reinforcement learning network to achieve adaptive selection of risk indicators and improve the efficiency of risk situation prediction. However, the above studies are all based on situation understanding and completion, that is, the data used for situation prediction are assumed to be real values, but there is a lot of bad data and noise in the actual power system, which cannot be directly applied to situation prediction.
[0010] In their research, Huang Manyun et al. proposed a distribution network situation awareness method based on historical data mining. First, they used historical data to model the operating status of the distribution network and provided virtual measurement information for the traditional least squares-based state estimator to obtain a state estimation result with higher accuracy. Then, they calculated the safety situation indicators based on the state estimation results and predicted the future trend of the distribution network. Although this method is a situation awareness method that combines situation understanding with situation prediction, it has certain limitations because it uses the traditional least squares state estimation and the indicators used are only selected from the perspective of system stability.
[0011] In summary, the main problems of the prior art are:
[0012] 1) At present, most of the research on the situation awareness methods of power systems is carried out separately. After the situation understanding, the risk situation of the system is not predicted, or the situation prediction is not based on the situation understanding, but directly assumes that the system state quantity is the real value, which makes the research on the situation awareness of the power system lack of rationality and systematicity.
[0013] 2) Regarding situation understanding, although the current situation awareness method based on graph convolutional network (GCN) effectively considers the topological relationship of the power system, the traditional GCN uses graph Laplacian operator for feature decomposition, which makes the calculation complexity high on large-scale graphs. In addition, GCN usually only considers the information of the first-order neighbors of the node and cannot capture more comprehensive graph structure information, so it performs poorly on highly complex graphs. Summary of the invention
[0014] In view of the above problems, the purpose of the present invention is to provide a power system situation awareness method based on Chebyshev graph neural network, which takes into account the integrity and progressiveness of situation awareness technology, combines situation understanding with situation prediction, avoids the limitation that the separated research cannot be applied to the actual operation of the power system; and fully extracts the characteristic information of high-order neighbor nodes, and more effectively processes the global information of the graph to solve the problem of situation understanding of large-scale complex power systems. The technical solution is as follows:
[0015] A method for power system situation awareness based on Chebyshev graph neural network includes the following steps:
[0016] Step 1: Perform power system flow simulation, obtain system operating parameters and adjacency matrix under different working conditions, perform normalization and construct a graph data set;
[0017] Step 2: Use the Chebyshev graph convolutional layer to build a graph neural network model, determine the input and output of the model, and use the graph dataset obtained in step 1 to train the graph neural network model;
[0018] Step 3: Input the real-time measurement data of the power system into the trained graph neural network model to obtain the real-time operating status of the system and realize the understanding of the power system situation;
[0019] Step 4: Construct multi-dimensional risk situation evaluation indicators of power system nodes from the perspective of system topology, attacker and cascading failure, and calculate the values of each evaluation indicator using the situation understanding results in step 3;
[0020] Step 5: Assign weights to each evaluation indicator, and then calculate the comprehensive risk situation value of the node. According to the comprehensive risk situation value of the node, the risk situation level of the node is classified to realize the situation prediction of the power system; finally, the situation awareness of the power system is realized through the inheritance and combination of situation understanding and situation prediction.
[0021] Furthermore, step 1 specifically includes:
[0022] Step 1.1: Select the fault types under normal topology, removing one branch, and removing two branches respectively. Under each topology, change the load distribution, use Matlab to perform AC power flow simulation, and obtain the original power system power flow data under different working conditions;
[0023] Step 1.2: Normalize the original power system flow data so that all data are mapped between 0 and 1, as shown in formula (1):
[0024] (1);
[0025] In the formula, is the original feature data value, and are the maximum and minimum values of the original feature data, respectively. is the normalized feature data value;
[0026] Step 1.3: Construct a graph data set. Each graph data includes node features, node labels, and adjacency matrix under a certain load distribution in a certain topology. The node features are shown in formula (2):
[0027] (2);
[0028] In the formula, represents the input feature data vector, is the node number, - is the input data of each node; - Input data - A characteristic of,representing the injected active power of each node; - Input data - Another feature of represents the injected reactive power of each node; the node label is shown in formula (3):
[0029] (3);
[0030] In the formula, represents the output vector, - is the output data of each node; - is a label for each node, representing the voltage amplitude of each node; - is another label for each node, representing the phase angle of each node;
[0031] Elements in the adjacency matrix A As shown in formula (4):
[0032] (4);
[0033] Based on the above graph data construction method, a simulation graph dataset for training and testing is constructed, and the simulation graph dataset is divided into a training set and a test set in proportion. The training set is used to determine various hyperparameters and weights in the Chebyshev graph neural network model, and the test set is used to evaluate the performance of the graph neural network model after training.
[0034] Furthermore, the calculation steps of the graph neural network built by the Chebyshev graph convolution layer in step 2 are as follows:
[0035] Step 2.1: Input the feature data and adjacency matrix in step 1 and calculate the degree matrix :
[0036] (5);
[0037] In the formula, is an element in the adjacency matrix A, representing the node With Node connection relationship; is the degree matrix The diagonal elements of The sum of the number of connected branches, that is, the node degree;
[0038] Step 2.2: Calculate the graph Laplacian matrix , and then calculate the normalized graph Laplacian matrix by the following formula :
[0039] (6);
[0040] (7);
[0041] In the formula, is the graph Laplacian matrix The maximum eigenvalue of is the identity matrix;
[0042] Step 2.3: Initialize Chebyshev polynomials, Chebyshev polynomials The recursive definition is as follows:
[0043] (8);
[0044] In the formula, is the order of the Chebyshev polynomial;
[0045] Step 2.4: Perform graph convolution operation. The graph convolution is defined in the spectral domain as follows:
[0046] (9);
[0047] In the formula, is the node feature matrix, is the filter function in graph convolution, is the eigenvector matrix of the graph Laplacian matrix, is the diagonal matrix of eigenvalues of the graph Laplacian matrix, is the representation of the filter function in the spectrum domain; T is the transpose symbol; Represents the convolution operation;
[0048] ChebNet uses Chebyshev polynomial approximation , as shown below:
[0049] (10);
[0050] In the formula, are the parameters that need to be learned, For the The Chebyshev polynomial of order is obtained, from which the ChebNet graph convolution formula is obtained as shown below:
[0051] (11);
[0052] In the formula, For the Layer node feature matrix.
[0053] Furthermore, step 4 specifically includes:
[0054] Step 4.1: From a topological perspective, the improved directed electrical betweenness is used as the first indicator, which is defined as follows:
[0055] (12);
[0056] In the formula, is the set of power generation nodes, is the set of load nodes, For power generation nodes Power generation, For power generation nodes Load Node The transmitted active power, Active power Passing Node The amount of
[0057] Step 4.2: From the attacker’s perspective, use the node attack risk value As the second indicator, its definition is as follows:
[0058] (13);
[0059] In the formula, is the attack probability, which refers to the probability that the attacker attacks each target considering limited attack resources; is the probability of successful attack, which is related to the vulnerability of the node; The impact after the node is destroyed is represented by the load borne by the node;
[0060] Step 4.3: From the perspective of cascading failures, the cascading failure load loss is used as the third indicator, which means that in the current system, if the node Out of service, the total amount of load lost in the system due to a cascading failure caused by power flow transfer , defined as follows:
[0061] (14);
[0062] In the formula, For Node The set of failed nodes caused by chain failure after exiting the operation. for The nodes in For a node collection Middle The load borne by each node;
[0063] Node Collection It is calculated by the accident chain search algorithm, and the calculation formula is as follows:
[0064] (15);
[0065] In the formula, For transmission lines Disconnect the line The trend, For transmission lines Line after fault The probability of failure, For line Initial trend, For line Maximum capacity, is the tripping probability due to protection, is a multiple of the capacity limit.
[0066] Furthermore, in step 4.2, for the attack probability , success probability And the impact after the node is destroyed The calculation method is:
[0067] Step 4.2.1: Calculate the load weight and electrical betweenness weight of each node to determine the attacker resource allocation and system defense resource allocation, as shown in the following formula:
[0068] (16);
[0069] In the formula, For Node The load weight; For Node Load; For Node Load; is the total number of nodes in the system; For Node The electrical betweenness weight of For Node The electrical betweenness of For Node The electrical betweenness of
[0070] Step 4.2.2: Determine the allocation threshold of attacker resources and system defense resources;
[0071] Step 4.2.3: Calculate the attack probability , the node numbers are assigned according to the node load weights Sort in descending order and calculate the attack probability of each node according to the order of node numbers = , and accumulate ; When the attack probability After being greater than or equal to the attacker's resource allocation threshold, the attack probability of the remaining nodes =0;
[0072] Calculating the probability of attack success , the node numbers are assigned according to the node electrical betweenness weights Sort in descending order and calculate the attack success probability of each node according to the node number order =1- , and accumulate ; When the attack success probability The probability of attack success of the remaining nodes after it is greater than or equal to the system defense resource allocation threshold =1;
[0073] Step 4.2.4: The impact of node destruction directly uses node load As a measure, ,in Represents the node Load Numeric value.
[0074] Furthermore, in step 5, the weighting of each evaluation index is performed by a multi-attribute decision-making method, specifically including:
[0075] Step 5.1: Determine the subjective weight of the indicator through fuzzy analytic hierarchy process:
[0076] Using triangular fuzzy numbers To express the relative importance of the three evaluation indicators, the fuzzy judgment matrix is constructed using triangular fuzzy numbers. , as shown below:
[0077] (17);
[0078] In the formula, , s=1,2…r, t=1,2…r; and ; , and are fuzzy numbers, which are the minimum importance, most likely importance and maximum importance of the sth evaluation index relative to the tth evaluation index;
[0079] Calculate the fuzzy consistency matrix , extract the fuzzy weight vector from the fuzzy consistency matrix , as shown below:
[0080] (18);
[0081] In the formula, , and is the first Fuzzy number of rows; , and is the first Fuzzy number of rows; is the fuzzy weight vector The elements in; then use the center value method to convert the fuzzy weight vector Defuzzification is performed as shown below:
[0082] (19);
[0083] So as to obtain the subjective weight ;
[0084] Step 5.2: Determine the objective weight of the indicator by entropy weight method:
[0085] The entropy of the sth evaluation index is calculated by the following formula:
[0086] (20);
[0087] In the formula, is the number of evaluation indicators, , , For the The sth index value on the node; assuming that = 0, =0; is the number of nodes; so the entropy weight of the sth evaluation index is:
[0088] (twenty one);
[0089] In the formula, For the The entropy of the evaluation index;
[0090] So as to obtain objective weight ;
[0091] Step 5.3: The calculated subjective weight With objective weight Combined, calculate the comprehensive weight of the evaluation index :
[0092] (twenty two);
[0093] In the formula, and are the subjective weight and objective weight of the sth evaluation index respectively, and Respectively The subjective weight and objective weight of each evaluation indicator; is the comprehensive weight of the sth evaluation index;
[0094] Step 5.4: Calculate the node comprehensive situation score according to the comprehensive weight of the evaluation index; first normalize the index values of all nodes, and then calculate the node comprehensive situation score :
[0095] (twenty three);
[0096] In the formula, For the The sth index value of the node;
[0097] Step 5.5: Get the comprehensive situation score of each node Afterwards, the risk level is divided to obtain the risk situation level of each node and realize the power system situation prediction.
[0098] The beneficial effects of the present invention are:
[0099] 1) The present invention fully considers the integrity and progressiveness of the power system situation awareness technology, combines situation understanding with situation prediction, and the results of situation understanding serve situation prediction. Situation prediction is based on the results of situation understanding to form a complete situation awareness system of power system measurement data → situation understanding → situation prediction, avoiding the limitation of viewing each part separately and being unable to be applied to the actual operation of the power system, thereby increasing the availability of the situation awareness system in the power system.
[0100] 2) In view of the high computational complexity of GCN on large-scale graphs and the fact that GCN usually only considers the information of the first-order neighbors of nodes and cannot capture more comprehensive graph structure information, the present invention performs poorly on highly complex graphs. The present invention proposes to use ChebNet to understand the power system situation, fully extract the feature information of high-order neighbor nodes, and more effectively process the global information of the graph to solve the problem of understanding the situation of large-scale complex power systems. Thanks to the property that Chebyshev polynomials approximate the Laplace matrix of graphs, ChebNet can capture node information in a larger range in one layer of convolution. High-order Chebyshev polynomials can aggregate the information of multi-order node neighbors in one layer, making the information propagation range wider and being able to more effectively process the global information of the graph. At the same time, ChebNet has good adaptability under changing topologies, so ChebNet is more suitable for real-time operation situation perception of large-scale complex power systems with time-varying operation states. BRIEF DESCRIPTION OF THE DRAWINGS
[0101] Figure 1 This is a flowchart of power system situation awareness based on Chebyshev graph neural network and multi-attribute decision making.
[0102] Figure 2 This is a flowchart of the computational steps of the Chebyshev graph neural network.
[0103] Figure 3 Comparison chart of GCN and ChebNet training losses.
[0104] Figure 4 (a) is a comparison of the model prediction results of GCN and ChebNet with non-training data input - under normal topology.
[0105] Figure 4 (b) is a comparison of the model prediction results of GCN and ChebNet with non-training data input - after removing branches 72 and 154.
[0106] Figure 4 (c) is a comparison of the model prediction results of GCN and ChebNet with non-training data input - after removing branches 1, 54, and 168.
[0107] Figure 4 (d) is a comparison of the model prediction results of GCN and ChebNet with non-training data input - after removing branch 23. DETAILED DESCRIPTION
[0108] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.
[0109] A power system situation awareness method based on Chebyshev graph neural network is proposed. The situation understanding of the power system is realized by using the trained Chebyshev graph neural network. On this basis, the risk situation prediction of the power system is realized by constructing multi-dimensional risk situation indicators. The flow chart of power system situation awareness based on Chebyshev graph neural network and multi-attribute decision making is shown in the figure. Figure 1 As shown, the following steps are included:
[0110] Step 1: Perform power system flow simulation to obtain data for training.
[0111] Consider the fault types under normal topology, removing one branch (N-1), and removing two branches (N-2) to increase the robustness and scalability of the model. At the same time, under each topology, change the load distribution, use Matlab to perform AC power flow simulation, and obtain the power system power flow data under different working conditions and the corresponding adjacency matrix. Thus, a graph data set is constructed. In order to avoid the influence of dimension on model training, the original power system power flow data is normalized so that all data are mapped between 0 and 1, as shown in formula (1):
[0112] (1);
[0113] In the formula, is the original feature data value, , are the maximum and minimum values of the original feature data, respectively. is the normalized feature data value.
[0114] Each graph data includes node features, node labels and adjacency matrix under a certain load distribution in a certain topology. The node features are shown in formula (2):
[0115] (2);
[0116] In the formula, is the input vector, is the node number, - is the input data of each node; - Input data - A characteristic of,representing the injected active power of each node; - Input data - Another feature of represents the injected reactive power of each node.
[0117] The node label is shown in formula (3):
[0118] (3);
[0119] In the formula, is the output vector, - is the output data of each node; - is a label for each node, representing the voltage amplitude of each node; - is another label for each node, representing the phase angle of each node.
[0120] Adjacency Matrix As shown in formula (4):
[0121] (4);
[0122] Based on the above graph data construction method, a graph dataset for training and testing is constructed. The dataset is divided into a training set and a test set in a ratio of 7:3. The training set is used to determine various hyperparameters and weights in the graph neural network model, and the test set is used to evaluate the performance of the graph neural network model after training.
[0123] Step 2: Use the Chebyshev graph convolutional layer to build a graph neural network, use the graph dataset built in step 1 to train the network, put the trained network into the actual power system, input the current moment collection quantity, obtain the system's global real-time operating status quantity, and realize the understanding of the power system situation.
[0124] First, the graph convolutional network GCN uses a convolution operation based on the first-order nearest neighbor, as shown in formula (5):
[0125] (5);
[0126] In the formula, is the adjacency matrix with self-loops added, where is the adjacency matrix, is the identity matrix, yes The degree matrix of It is The feature matrix of the layer nodes, It is The weight matrix of the layer, is the activation function.
[0127] It can be seen that the convolution operation of GCN normalizes the adjacency matrix and multiplies it with the feature matrix, and then undergoes nonlinear transformation. It mainly aggregates the first-order neighbor information of the node. This means that the graph convolution operation of each layer can only propagate information to directly adjacent nodes. In order to capture information at a longer distance, multiple layers of GCN need to be stacked. In large-scale power systems, this will seriously affect the computing efficiency. However, the Chebyshev graph neural network can capture more extensive neighbor information in one network layer by recursively calculating the Chebyshev polynomial. The calculation step flow chart of the Chebyshev graph neural network is shown below. Figure 2 shown.
[0128] First enter the and , the degree matrix is calculated by formula (6):
[0129] (6);
[0130] Then the graph Laplacian matrix is calculated by formula (7): , and then calculate the normalized graph Laplacian matrix using formula (8) :
[0131] (7);
[0132] (8);
[0133] In the formula, yes The maximum eigenvalue of is the identity matrix.
[0134] Then, initialize the Chebyshev polynomial, Chebyshev polynomial The recursive definition is shown in formula (9):
[0135] (9);
[0136] In the formula, is the order of the Chebyshev polynomial.
[0137] Then, the graph convolution operation is performed. The graph convolution is defined in the spectrum domain as shown in formula (10):
[0138] (10);
[0139] In the formula, is the node feature matrix, is the filter function in graph convolution, is the eigenvector matrix of the graph Laplacian matrix, is the diagonal matrix of eigenvalues of the graph Laplacian matrix, It is the representation of the filter function in the spectral domain. Since the complexity of directly calculating the spectral convolution is high, ChebNet uses Chebyshev polynomial approximation , as shown in formula (11):
[0140] (11);
[0141] In the formula, are the parameters that need to be learned, For the The Chebyshev polynomial of order is obtained, and the ChebNet graph convolution formula is obtained, as shown in formula (12):
[0142] (12);
[0143] In the formula, For the Layer node feature matrix.
[0144] by =3 as an example, the ChebNet graph convolution process is:
[0145] (1) Initialize input features = ;
[0146] (2) Calculate the graph Laplacian matrix and ;
[0147] (3) Input the first convolutional layer and calculate , , and ;
[0148] (4) Apply formula (12) to perform convolution operation and obtain ;
[0149] (5) Apply activation function , get the feature matrix after the first convolution ;
[0150] (6) Repeat steps (1) to (5) in the remaining convolutional layers, and finally output by the fully connected layer = .
[0151] From the above steps, it can be seen that when = 3, ChebNet can capture node information in a wider range in one layer of convolution through the high-order approximation of Chebyshev polynomials. High-order Chebyshev polynomials can aggregate the information of multiple node neighbors in one layer, making the information propagation range wider and more efficient in processing the global information of the graph. Therefore, ChebNet is more suitable for the problem of understanding the situation of large-scale complex power systems.
[0152] Step 3: Construct multi-dimensional risk situation evaluation indicators of power system nodes from the perspective of topology, attacker, and cascading failure, respectively. Use the situation understanding results in step 2 to calculate the values of each indicator, use fuzzy analytic hierarchy process and entropy weight method to subjectively and objectively weight each indicator, and then calculate the comprehensive risk situation value of the node. According to the comprehensive risk situation value of the node, the node is divided into risk situation levels to realize the situation prediction of the power system. Finally, through the inheritance and combination of situation understanding and situation prediction, the situation awareness of the power system is realized.
[0153] First, from the perspective of system topology, many studies currently use complex network theory to analyze the topological vulnerability of the system. In complex networks, the topological structure of the network is considered. In most models, it is assumed that the flow propagates along the shortest path. Both weighted betweenness and electrical betweenness can characterize structural vulnerability. From the perspective of betweenness, it is concluded that high-betweenness nodes or branches of the system will become very vulnerable under disturbance. The traditional betweenness is defined as the number of times the shortest path passes between all generators and loads in the network. However, the traditional betweenness regards the system as an undirected weighted network, without considering the current operating status of the system and the directionality of the flow. Since the shortest path belongs to the concept of graph theory, it does not combine the physical characteristics of the power system. Therefore, the improved directed electrical betweenness is used as the first indicator, and its definition is shown in formula (13):
[0154] (13);
[0155] In the formula, is the set of power generation nodes, is the set of load nodes, For power generation nodes Power generation, For power generation nodes Load Node The transmitted active power, Active power Passing Node The definition shows that the node Electrical Betweenness It reflects the importance of the node in the system structure. The larger the value, the greater the impact of the node's decommissioning on the system.
[0156] Secondly, from the attacker's perspective, since the attacker will consider multiple factors when attacking a node, such as its own attack resources, the target's defense resources, and the impact of the target being destroyed, the node attack risk value is used as the second indicator, which is defined as shown in formula (14):
[0157] (14);
[0158] In the formula, is the attack probability, which refers to the probability that the attacker attacks each target considering limited attack resources; In order to increase the probability of successful attack, since there are naturally weak points in the power system, more defense resources are often applied to the weaker points during system planning to avoid damage. Related to node vulnerability; The impact after a node is destroyed is usually represented by the load borne by the node.
[0159] Further, for , and The calculation method is:
[0160] (1) Calculate the load weight and electrical betweenness weight of each node to determine the attacker resource allocation and system defense resource allocation, as shown in formula (15):
[0161] (15);
[0162] (2) Determine the allocation threshold of attacker resources and system defense resources. For example, resources can be allocated to nodes with a cumulative weight of 80%.
[0163] (3) Calculate the attack probability , the node numbers are assigned according to the node load weights Sort in descending order and calculate the attack probability of each node according to the order of node numbers = And accumulate ,when After it is greater than or equal to 80%, the attack probability of the remaining nodes =0; Calculate the probability of attack success , the node numbers are assigned according to the node electrical betweenness weights Sort in descending order and calculate the attack success probability of each node according to the node number order =1- And accumulate ,when After it is greater than or equal to 80%, the attack success probability of the remaining nodes =1.
[0164] Among them, the attack probability It is the probability of the attacker attacking each target node when the attack resources are limited. When the resources are limited, the attacker will give priority to the nodes with high load as the attack target to maximize the load loss of the power system. That is, the higher the load of the node, the higher the attack probability. At the same time, the power system itself has certain defense resources to protect each node from malicious damage. In the allocation of defense resources, system managers tend to deploy more defense resources at more critical nodes to protect the key links in the system, so the probability of successful attack is It is related to the defense resources applied to the node. That is, the more critical the node is, the more defense resources it will have, and the probability of successful attack is higher. It will get smaller.
[0165] In this embodiment, the attack probability The proportion of the load directly borne by the node to the total load To quantify, = , that is, node The higher the load, the higher the node Probability of attack At the same time, the probability of successful attack Directly use the proportion of node electrical intermediates to total node electrical intermediates However, since the higher the electrical betweenness of a node is, the more critical the node is, which means that the defense resources of this node are higher, the probability of success of attacking this node is should be smaller, so use =1- Characterize the probability of attack success and node electrical betweenness weight The relationship between.
[0166] Since in the actual environment, the attacker's attack resources and the power system's defense resources are limited, the attack and defense often cannot cover all nodes. In this case, the order in which the attacker invests attack resources should be based on the node load weight. Attack in descending order until the attack resources are exhausted. At the same time, the order in which the system applies defense resources should be based on the proportion of the node electrical betweenness. The load weights of all nodes that have invested in attack resources are The sum is the attacker's resource allocation threshold, and the proportion of the electrical betweenness of all nodes that apply defense resources is The sum is the system defense resource allocation threshold.
[0167] (4) Impact of node destruction on direct usage load As a measure, ,in Represents the node Load Numeric value.
[0168] Finally, from the perspective of cascading failures, studies have shown that many large blackouts are caused by cascading failures, specifically small local disturbances such as generator and transmission line failures that trigger a series of grid component failures through network connections. From the perspective of complex network theory, the grid topology is a small-world network with a high clustering coefficient and a short average path length. The network nodes or branches are closely connected. At the same time, due to the short average path length, the speed of fault propagation is increased. Therefore, when a node or branch in the network fails, the fault is easily propagated to adjacent nodes or even non-adjacent nodes, resulting in the breadth and depth of cascading failures. Therefore, the load loss due to cascading failures is used as the third indicator, which means that in the current system, if the node Out of operation, the total amount of load lost by the system in a cascading failure due to power flow transfer is defined as shown in formula (16):
[0169] (16);
[0170] In the formula, For Node The set of failed nodes caused by chain failure after exiting the operation. for The nodes in For a node collection Middle The load borne by a node. It is calculated by the accident chain search algorithm, and the calculation formula is shown in formula (17):
[0171] (17);
[0172] In the formula, For transmission lines Disconnect the line The trend, For line Line after fault The probability of failure, For line Initial trend, For line Maximum capacity, is the tripping probability due to protection, It is the capacity limit multiple, take 1.4.
[0173] Through the above steps, according to the system state value obtained in step 2, the three indicators of the node are calculated, and then the indicator weighting is realized through the multi-attribute decision-making method.
[0174] Firstly, the subjective weight of the index is determined by fuzzy hierarchical analysis method. To express the relative importance of the three evaluation indicators, the fuzzy judgment matrix is constructed using triangular fuzzy numbers. , as shown in formula (18):
[0175] (18);
[0176] In the formula , s=1,2…r, t=1,2…r; and ; , and is a fuzzy number, which is the minimum importance, most likely importance, and maximum importance of the sth indicator relative to the tth indicator. Calculate the fuzzy consistency matrix , extract the fuzzy weight vector from the fuzzy consistency matrix , as shown in formula (19):
[0177] (19);
[0178] In the formula, , and is the first Fuzzy number of rows; , and is the first Fuzzy number of rows; is the fuzzy weight vector Then use the center value method to convert the fuzzy weight vector Defuzzification is performed.
[0179] Among them, the fuzzy consistency matrix is calculated The process is Keep multiplying until the consistency ratio is met <0.1. Consistency index From the fuzzy consistency matrix The calculation formula is:
[0180] ;
[0181] in, is the fuzzy consistency matrix The maximum eigenvalue of the average random consistency index is a The standard value of the order correlation is 0.58 in this embodiment. is the number of evaluation indicators.
[0182] Then use the center value method to Defuzzification is performed as shown in formula (20):
[0183] (20);
[0184] Finally, a consistency check is performed to calculate the fuzzy consistency ratio to ensure the consistency of the judgment matrix, as shown in formula (21):
[0185] (twenty one);
[0186] In the formula, the consistency index Random Consistency Index It is determined according to the size of the judgment matrix. It is generally believed that when the consistency ratio When <0.1, the consistency of the judgment matrix is acceptable.
[0187] Secondly, the objective weight of the indicator is determined by the entropy weight method. The entropy of the sth evaluation indicator is calculated by formula (22):
[0188] (twenty two);
[0189] In the formula, is the number of evaluation indicators, , , For the The value of the sth evaluation index on the node; assuming that = 0, =0.
[0190] Therefore, the entropy weight of the sth evaluation index is:
[0191] (twenty three);
[0192] Finally, the subjective weight obtained by calculation With objective weight Combined, calculate the comprehensive weight of the evaluation index :
[0193] (twenty four);
[0194] After obtaining the comprehensive weight of the indicators, the comprehensive status score of the node can be calculated. First, the indicator values of all nodes are normalized by formula (1), and then the comprehensive status score of the node is calculated by formula (25): :
[0195] (25);
[0196] In the formula, Representative The sth index value of the node.
[0197] To each node Finally, the risk level is divided according to the method in Table 1, and the risk situation level of each node is obtained, realizing the power system situation prediction.
[0198] Table 1 Classification of node comprehensive risk situation levels
[0199] .
[0200] Case analysis:
[0201] The IEEE-118 node system is used for example simulation test. The IEEE-118 node system contains 186 branches, with a total power generation and total load of 4374.86MW and 4242MW respectively. The graph dataset for training is constructed using the method in step 1. The composition of the graph dataset is shown in Table 2:
[0202] Table 2 Composition of graph dataset
[0203] .
[0204] The graph dataset is divided into a training set and a test set in a ratio of 7:3. The training set is input into the Chebyshev graph neural network for training. The node feature information is mined based on multiple ChebNet layers. The full connection layer outputs the system node amplitude and phase angle. The loss function is calculated by the predicted value and the actual value, and the model weight is updated using the back propagation algorithm. Finally, through multiple iterations, a situation understanding model suitable for different topologies can be obtained. Due to the training status and parameters of ChebNet is closely related, that is, the order of captured neighbor node information, so The selection of needs to be compared experimentally. Table 3 shows the different ChebNet training under value:
[0205] Table 3 Different The impact of values on ChebNet training
[0206] .
[0207] It can be seen from Table 3 that with As the loss of the training set and the loss of the validation set gradually increase, the loss of the training set and the loss of the validation set gradually decrease. This shows that the higher the order of neighbor node information capture, the more global information is obtained, and the higher the learning accuracy. At the same time, the difference between the loss of the validation set and the loss of the training set is getting smaller and smaller, indicating that as the The overfitting phenomenon is significantly alleviated with the gradual increase of The larger the value, the higher the polynomial order that needs to be fitted in each iteration, the more neighbor nodes that need to be captured, and the training time will increase significantly. It can be found that = 3, the change in loss value is very small, indicating that there is not much room for optimization, but the training time is still increasing, so we choose =3 as the situation understanding model, and its model architecture and parameters are shown in Table 4:
[0208] Table 4 ChebNet situation understanding model architecture and parameters
[0209] .
[0210] Figure 3 When the number of layers and neurons is the same, GCN and = 3. In addition, in order to verify the scalability of ChebNet, the state quantities calculated using the load distribution conditions that were not used for training under different topologies were input into the trained model to obtain the voltage amplitude prediction values. The absolute errors between the predicted values of different models and the true values at each node were calculated and the absolute values were taken for comparison. The comparison results are shown in Figure 4 (a) to Figure 4 (d).
[0211] Depend on Figure 3It can be seen that the final loss value of GCN converges to 0.0238, while the loss value of ChebNet converges to 0.0049, which is about one-fifth of GCN, and its accuracy has been significantly improved. At the same time, from the comparison of the model prediction results of GCN and ChebNet with non-training data input under normal topology in Figure 4 (a), after removing branches 72 and 154 in Figure 4 (b), after removing branches 1, 54, and 168 in Figure 4 (c), and after removing branch 23 in Figure 4 (d), it can be seen that ChebNet also maintains good fitting accuracy for data that does not appear in the training set, especially in Figure 4 (c) for the case of removing branches 1, 54, and 168. This topology does not exist in the training set, but the ChebNet prediction value still accurately fits the true value. This shows that ChebNet has strong adaptability to unknown topologies and verifies its scalability for large-scale complex systems. When measurements are collected from real-time systems and input into trained ChebNet, the global state of the system can be obtained quickly, accurately and in real time, providing a data basis for subsequent situation prediction.
[0212] On the basis of situation understanding, that is, after obtaining the global state quantity, the three indicators in the state quantity calculation step 3 are used to calculate the three indicator values of the state quantity calculation predicted under a certain load distribution in Figure 4 (a) to Figure 4 (d) using formula (13) to formula (17), as shown in Table 5. Due to limited space, only part of the results are shown.
[0213] Table 5 Index values of each node
[0214] .
[0215] After obtaining the three indicator values of each node, the subjective weight, objective weight and comprehensive weight of the indicator are first calculated by formula (18)-formula (24), and then the indicator is normalized using formula (1). Finally, the comprehensive risk situation score of the node is calculated by formula (25). The results are arranged in descending order according to the score value, as shown in Table 6:
[0216] Table 6 Node comprehensive risk situation score
[0217] .
[0218] The results in Table 5 and the risk situation level classification method in Table 1 can be used to obtain the risk situation level of each node in the system under the current operating state in real time. In this example, nodes 59 and 116 are the nodes with the highest risk situation level in the current operating state, which is level 4. All redundancy and other low-risk node defense resources need to be called immediately for defense. From the above process, it can be seen that through the situation awareness method based on ChebNet and multi-attribute decision-making, the operating status and risk situation of the system can be perceived in real time in a dynamically running system.
Claims
1. A method for power system situation awareness based on Chebyshev graph neural network, characterized in that: The following steps are involved: Step 1: Perform power system flow simulation, obtain the simulation system operating parameters and adjacency matrix under different working conditions, perform normalization and construct a graph data set; Step 2: Use the Chebyshev graph convolutional layer to build a graph neural network model, determine the input and output of the graph neural network model, and use the graph dataset obtained in step 1 to train the graph neural network model; Step 3: Input the real-time measurement data of the power system into the trained graph neural network model to obtain the real-time operation status of the power system and realize the understanding of the power system situation; Step 4: Construct multi-dimensional risk situation evaluation indicators of power system nodes from the perspective of topology, attacker and cascading failure, and calculate the value of each evaluation indicator using the situation understanding results in step 3; Step 5: Assign weights to each evaluation index, and then calculate the comprehensive risk situation value of the node. According to the comprehensive risk situation value of the node, the node is classified into risk situation levels to realize the situation prediction of the power system; Finally, through the integration and combination of situation understanding and situation prediction, the situation awareness of the power system can be achieved; Step 4 Specific include: Step 4.1: From a topological perspective, the improved directed electrical betweenness is used as the first indicator, which is defined as follows: (12); In the formula, is the set of power generation nodes, is the set of load nodes, For power generation nodes Power generation, For power generation nodes Load Node The transmitted active power, Active power Passing Node The amount of Step 4.2: From the attacker’s perspective, use the node attack risk value As the second indicator, its definition is as follows: (13); In the formula, is the attack probability, which refers to the probability of the attacker attacking each target considering limited attack resources; is the probability of successful attack, which is related to the vulnerability of the node; The impact after the node is destroyed is represented by the load borne by the node; Step 4.3: From the perspective of cascading failures, the cascading failure load loss is used as the third indicator, which means that in the current system, if the node Out of service, the total amount of load lost in the system due to a cascading failure caused by power flow transfer , defined as follows: (14); In the formula, For Node The set of failed nodes caused by chain failure after exiting the operation. for The nodes in For a node collection Middle The load borne by each node; Node Collection It is calculated by the accident chain search algorithm, and the calculation formula is as follows: (15); In the formula, For transmission lines Disconnect the line The trend, For transmission lines Line after fault The probability of failure, For line Initial trend, For line Maximum capacity, is the tripping probability due to protection, is the capacity limit multiple; In step 5, the weights of the evaluation indicators are assigned through the multi-attribute decision-making method, which specifically includes: Step 5.1: Determine the subjective weight of the indicator through fuzzy analytic hierarchy process: Using triangular fuzzy numbers To express the relative importance of the three evaluation indicators, the fuzzy judgment matrix is constructed using triangular fuzzy numbers. , as shown below: (17); In the formula, , s=1,2…r, t=1,2…r; and ; , and are fuzzy numbers, which are the minimum importance, most likely importance and maximum importance of the sth evaluation index relative to the tth evaluation index; Calculate the fuzzy consistency matrix , extract the fuzzy weight vector from the fuzzy consistency matrix , as shown below: (18); In the formula, , and is the first Fuzzy number of rows; , and is the first Fuzzy number of rows; is the fuzzy weight vector The elements in; then use the center value method to convert the fuzzy weight vector Defuzzification is performed as shown below: (19); So as to obtain the subjective weight ; Step 5.2: Determine the objective weight of the indicator by entropy weight method: The entropy of the sth evaluation index is calculated by the following formula: (20); In the formula, is the number of evaluation indicators, , , For the The sth index value on the node; assuming that = 0, =0; is the number of nodes; Therefore, the entropy weight of the sth evaluation index is: (21); In the formula, For the The entropy of the evaluation index; So as to obtain objective weight ; Step 5.3: The calculated subjective weight With objective weight Combined, calculate the comprehensive weight of the evaluation index : (22); In the formula, and are the subjective weight and objective weight of the sth evaluation index respectively, and Respectively The subjective weight and objective weight of each evaluation indicator; is the comprehensive weight of the sth evaluation index; Step 5.4: Calculate the node comprehensive situation score according to the comprehensive weight of the evaluation index; first normalize the index values of all nodes, and then calculate the node comprehensive situation score : (23); In the formula, For the The sth index value of the node; Step 5.5: Get the comprehensive situation score of each node Afterwards, the risk level is divided to obtain the risk situation level of each node and realize the power system situation prediction.
2. A method for power system situation awareness based on Chebyshev graph neural network according to claim 1, characterized in that: Step 1 specifically includes: Step 1.1: Select the fault types under normal topology, removing one branch, and removing two branches respectively. Under each topology, change the load distribution, use Matlab to perform AC power flow simulation, and obtain the original power system power flow data under different working conditions; Step 1.2: Normalize the original power system flow data so that all data are mapped between 0 and 1, as shown in formula (1): (1); In the formula, is the original feature data value, and are the maximum and minimum values of the original feature data, respectively. is the normalized feature data value; Step 1.3: Construct a graph dataset. Each graph dataset includes node features, node labels, and adjacency matrix under a certain load distribution in a certain topology. The node characteristics are shown in formula (2): (2); In the formula, represents the input feature data vector, is the node number, - is the input data of each node; - Input data - A characteristic of,representing the injected active power of each node; - Input data - Another characteristic of,represents the injected reactive power of each node; The node label is shown in formula (3): (3); In the formula, represents the output vector, - is the output data of each node; - is a label for each node, representing the voltage amplitude of each node; - is another label for each node, representing the phase angle of each node; Elements in the adjacency matrix A As shown in formula (4): (4); Based on the above graph data construction method, a simulation graph dataset for training and testing is constructed, and the simulation graph dataset is divided into a training set and a test set in proportion. The training set is used to determine various hyperparameters and weights in the Chebyshev graph neural network model, and the test set is used to evaluate the performance of the graph neural network model after training.
3. A method for power system situation awareness based on Chebyshev graph neural network according to claim 2, characterized in that: The calculation steps of the graph neural network built by the Chebyshev graph convolution layer in step 2 are as follows: Step 2.1: Input the feature data and adjacency matrix in step 1 and calculate the degree matrix : (5); In the formula, is an element in the adjacency matrix A, representing the node With Node connection relationship; is the degree matrix The diagonal elements of The sum of the number of connected branches, that is, the node degree; Step 2.2: Calculate the graph Laplacian matrix , and then calculate the normalized graph Laplacian matrix by the following formula : (6); (7); In the formula, is the graph Laplacian matrix The maximum eigenvalue of is the identity matrix; Step 2.3: Initialize Chebyshev polynomials, Chebyshev polynomials The recursive definition is as follows: (8); In the formula, is the order of the Chebyshev polynomial; Step 2.4: Perform graph convolution operation. The graph convolution is defined in the spectral domain as follows: (9); In the formula, is the node feature matrix, is the filter function in graph convolution, is the eigenvector matrix of the graph Laplacian matrix, is the diagonal matrix of eigenvalues of the graph Laplacian matrix, is the representation of the filter function in the spectrum domain; T is the transpose symbol; Represents the convolution operation; ChebNet uses Chebyshev polynomial approximation , as shown below: (10); In the formula, are the parameters that need to be learned, For the The Chebyshev polynomial of order is obtained, from which the ChebNet graph convolution formula is obtained as shown below: (11); In the formula, For the Layer node feature matrix.
4. The method for power system situation awareness based on Chebyshev graph neural network according to claim 1 is characterized in that: In step 4.2, for the attack probability , success probability And the impact after the node is destroyed The calculation method is: Step 4.2.1: Calculate the load weight and electrical betweenness weight of each node to determine the attacker resource allocation and system defense resource allocation, as shown in the following formula: (16); In the formula, For Node The load weight; For Node Load; For Node Load; is the total number of nodes in the system; For Node The electrical betweenness weight of For Node The electrical betweenness of For Node The electrical betweenness of Step 4.2.2: Determine the allocation threshold of attacker resources and system defense resources; Step 4.2.3: Calculate the attack probability , the node numbers are assigned according to the node load weights Sort in descending order and calculate the attack probability of each node according to the node number order = , and accumulate ; When the attack probability After being greater than or equal to the attacker's resource allocation threshold, the attack probability of the remaining nodes =0; Calculating the probability of attack success , the node numbers are assigned according to the node electrical betweenness weights Sort in descending order and calculate the attack success probability of each node according to the node number order =1- , and accumulate ; When the attack success probability The probability of attack success of the remaining nodes after it is greater than or equal to the system defense resource allocation threshold =1; Step 4.2.4: The impact of node destruction directly uses node load As a measure, ,in Represents the node Load Numeric value.
