A method for risk propagation analysis of a multi-layer coupled power system based on a spatio-temporal graph neural network

Abstract

The application is a kind of multilayer coupled power system risk propagation analysis method based on space-time graph neural network, and belongs to the technical field of power system.The method constructs an information-power-traffic three-layer heterogeneous coupled network model, collects physical quantities in the heterogeneous coupled network, and constructs graph data based on neighborhood and connectivity; a risk propagation prediction model is built based on a space-time multi-graph neural network, the spatial learning network includes k parallel graph convolution layers, and the temporal learning network is realized by combining a long short-term memory network LSTM and an attention mechanism; historical graph data of normal operation and historical graph data of fault occurrence of the heterogeneous coupled network are acquired, input into the prediction model for training, and the graph structure can be adaptively updated; the trained risk propagation prediction model is used to predict the risk propagation path in the three-layer heterogeneous coupled network.The application realizes the risk propagation path in the information-power-traffic three-layer heterogeneous complex network structure, and can reflect real-time space-time relationship.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of power systems, and particularly relates to a risk propagation analysis method for a multi-layer complex coupled power system based on a space-time graph neural network. BACKGROUND

[0002] In order to better meet the needs of economic and social development and energy transformation, high-proportion distributed energy access has gradually become a basic feature and development form of power systems. Among them, electric vehicles (EVs) as flexible resources on the load side have developed rapidly in recent years, and a large number of electric vehicle users make decisions on charging and discharging, driving, and other aspects under the guidance of various information, which will have a certain impact on the operation state of the power grid and the traffic flow of the road network. In addition, with the construction and deployment of 5G communication base stations, considering that base station energy storage can participate in economic dispatch, it provides more flexible adjustment capability for the power system.

[0003] Electricity, transportation and communication are important links and basic guarantees for the normal operation of society. Today, these three heterogeneous networks are no longer independent individuals, but constitute an information-power-transportation ternary coupled network that penetrates each other, is associated with each other and is deeply coupled, and involves frequent interactions between information, electricity and transportation. In this multi-layer complex system, the information communication network is responsible for the connection of various information communication devices (including 5G base stations) and further realizes functions such as calculation, sensing and control. The physical power grid mainly refers to the network connection between various power equipment (including charging piles) and is responsible for functions such as power transmission and supply. The transportation network is closely related to both of them and is mainly realized through electric vehicles, an important traffic flow component, such as signal reception and transmission, charging and discharging behavior, etc. The coupling network diagram of information-power-transportation is shown in Figure 1 The power grid is the power source of the transportation network and the information network; the transportation network is an important load of the power grid and an important user of the information network; and the information network is the communication foundation for efficient and stable operation of the power grid and the transportation network.

[0004] It is the strong coupling and dependence between heterogeneous systems that leads to the failure of a certain link, which may affect other networks. The deep coupling and interaction between information-power-transportation networks not only brings convenience to production and life, but also creates favorable conditions for risk propagation and hides safety hazards.

[0005] Currently, for complex coupled power systems, most of the research is on double-layer network models such as "cyber-physical power system", "power + traffic", etc., focusing on analyzing the impact of information network attacks on the safety performance of the power grid and the optimal dispatching operation of the power grid. For example, in 2010, Professor Buldyrev proposed the concept of interdependent networks and a vulnerability analysis framework in document 1 (Buldyrev S V, Parshani R, Paul G, et al. Catastrophic cascade of failures in interdependent networks[J]. Nature, 2010, 464(7291): 1025.), which provides a new way of thinking for analyzing the interaction between coupled systems and promotes the trend of studying cyber-physical power systems from the perspective of interdependent networks. In document 2 (Chen Shiming, Zou Xiaogun, Lv Hui, et al. Robustness of interdependent networks for cascading failures[J]. Acta Physica Sinica, 2014, 63(02): 432-441.), the critical characteristics of the interdependent network model are analyzed by adjusting the weight parameters and inter-network coupling strength of the two sub-networks. In document 3 (Zhang Yawei, Liu Wenxia, Liu Gengming, Huang Shaofeng. Modeling and vulnerability analysis of power information-physical system considering topological correlation and double coupling[J]. Proceedings of the CSEE, 2021, 41(16): 5486-5500.), the topological correlation between the double-layer networks is considered, and a double-coupled power information-physical system model is established for vulnerability assessment. In document 4 (Ji Xingpei, Wang Bo, Liu Dicun, et al. Review of interdependent network theory and its application in structural vulnerability analysis of power information-physical system[J]. Proceedings of the CSEE, 2016, 36(17): 4521-4532.), considering the actual situation of China's power system, an information node and power node partial one-to-one model is established. In document 5 (Wang Zhiwei, Wang Xiuli, Yong Weizhen, et al. Trading strategy of 5G base station virtual power plant with backup energy storage participating in power market[J]. Power System Technology, 2024, 48(03): 968-982.), the use of 5G base stations and wind farms and photovoltaic power plants to form virtual power plants (VPP) to participate in the optimal bidding and auxiliary frequency modulation of the power market is proposed. In document 6 (Zhang Wei, Zhu Tongtong, Su Jin. Demand response strategy for power-information-traffic coupled network considering electric vehicles and 5G base stations[J / OL]. Power System Automation, 2024.), the impact of electric vehicles and 5G base stations on the power grid is analyzed, and a two-stage demand response optimization scheduling strategy is proposed. In document 7 (Sheng Yujie, Guo Qinglai, Xue Yixun, et al. Modeling and collaborative optimization of power-traffic coupled network from the perspective of information-physical-social[J / OL]. Power System Automation, 2024.), related research on modeling and collaborative optimization of power-traffic coupled networks is conducted.

[0006] In summary, for the complex system of information-power-transportation, a multi-level coupled network model has not been established from the whole, and the risk evolution dynamic process brought by coupling interaction has not been analyzed. SUMMARY

[0007] The present application proposes a risk propagation analysis method for a multi-layer complex coupled power system based on a spatio-temporal graph neural network, which establishes a three-layer heterogeneous network model of information-power-transportation from the complexity of network structure and the dynamic characteristics of nodes, describes the interlayer coupling relationship by referring to the graph theory, describes the complex spatio-temporal dependence relationship between networks by using a spatio-temporal graph neural network (STGNN) for the time-varying characteristics of data in each network, and further analyzes the dynamic characteristics of risk propagation.

[0008] The risk propagation analysis method for a multi-layer complex coupled power system based on a spatio-temporal graph neural network of the present application comprises the following steps:

[0009] Step 1: Establish a three-layer heterogeneous coupled network model of information-power-transportation, which includes an upper information network, a middle power network and a lower transportation network; the information network is constructed based on the communication network of the Ethernet, the nodes in the graph include base stations and communication equipment, and the edges are established when the nodes are connected in communication; the power network is constructed based on the power system, the nodes in the graph are busbars, and the electrical lines connected between the busbars are constructed as edges in the graph; the transportation network is constructed using a dynamic road network model, the nodes in the graph are intersections, and the edges in the graph are established according to the road lanes; if there is a connection relationship between the nodes in different layers, the edges between the nodes in different layers are established;

[0010] Step 2: Obtain the node data in the three-layer heterogeneous coupled network, construct the graph structure data based on the neighborhood and connectivity; collect the physical quantities in the three-layer heterogeneous coupled network, construct the graph data of network G, including node features and edge relationships; the collected physical quantities include the signal strength of each node in the information network, the current, voltage and voltage phase angle of the busbar in the power network, and the traffic flow of each road in the transportation network; after normalizing the collected data in the same dimension, use the pre-set unified representation to represent the node features; establish the adjacency matrix A and the topological relationship graph A tp of network G according to the edge relationship between the nodes in different layers and the edge relationship between the nodes in each layer; tp characterize the edge relationship of G;

[0011] Step three: a risk propagation prediction model is built based on a spatio-temporal multi-graph neural network; the spatio-temporal multi-graph neural network of the risk propagation prediction model comprises a spatial learning network and a temporal learning network; the spatial learning network comprises k parallel graph convolution layers; the temporal learning network is realized by combining a long short-term memory network (LSTM) and an attention mechanism; the graph data of the three-layer heterogeneous coupled network collected at the first k time points is input into the risk propagation prediction model after normalization processing; the k parallel graph convolution layers are used to extract the spatial dimension features of the graph data at the k time points respectively; then the extracted features are input into the LSTM with the attention mechanism in sequence to extract the time dimension features; the extracted features are represented by the hidden state of the LSTM; finally, the features extracted at the k time points are fused to output the graph data at the k+1 time point; the risk propagation prediction model slides the input graph data by k steps to sequentially output the graph data at multiple prediction time points; k>=3.

[0012] Step four: historical data of the three-layer heterogeneous coupled network is acquired, including historical graph data of the network nodes in a normal operating state and historical graph data of the power system, the traffic network and the information network in different fault states; the historical data is input into the risk propagation prediction model for training until the model reaches the set accuracy requirement; in the training process, the graph structure of the three-layer heterogeneous coupled network is adaptively updated according to the collected time point data.

[0013] Step five: when a fault occurs in a node of the network, the trained risk propagation prediction model is used to predict the risk propagation path in the information-power-traffic three-layer heterogeneous coupled network.

[0014] The advantages and positive effects of the present application are as follows:

[0015] (1) The present application constructs an information-power-traffic three-layer heterogeneous network model, which fully represents the coupling interaction relationship between nodes in a complex network structure. The present application method extends from a regular data structure to a dynamic model, and realizes dynamic prediction of risk propagation in an information-power-traffic three-layer heterogeneous complex network structure.

[0016] (2) The present application uses a multi-dimensional time series based STMGCN to predict the risk propagation path of the information-power-traffic three-layer heterogeneous network, so as to achieve the purpose of predicting fault nodes and detecting key nodes. The STMGCN of the present application introduces time series features based on a graph neural network, which can simultaneously obtain the correlation of feature data in the time domain and the spatial domain, and is suitable for the scenario where the node features of the present application change over time. For each graph, the prediction model of the present application respectively uses a graph convolutional neural network (GCN), a long short-term memory network (LSTM) and an attention mechanism to capture time and spatial correlation, and finally obtains the final prediction result by fusing the parallel multi-graph spatio-temporal hidden information.

[0017] (3) The method of the application uses a graph neural network GNN to model and predict the propagation dynamics process on a complex network, designs a training process and a suitable GNN architecture, and can represent a wide range of dynamics with few assumptions. Previously, graph neural networks GNN have been widely used in various fields such as biological networks, social networks, recommendation systems, etc. However, these models are limited to static graph data, and the graph structure is fixed. The method of the application makes full use of time-varying graph data, mines valuable time series information in various systems in addition to static topology, extracts complex spatio-temporal dependencies through a space-time graph neural network framework, and makes the prediction result more reasonable and accurate.

[0018] (4) The method of the application analyzes the correlation strength between nodes in each network through a graph attention mechanism, focuses on the time correlation characteristics between data, and accurately reflects the real-time spatio-temporal relationship. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 is an information-power-transportation network interaction graph;

[0020] Figure 2 is a risk propagation analysis method for a multi-layer complex coupled power system of the application;

[0021] Figure 3 is a whole framework diagram of a risk propagation prediction model built by the method of the application;

[0022] Figure 4 is an ST structure diagram for capturing spatio-temporal correlation of the application;

[0023] Figure 5 is an intra-layer node load redistribution schematic diagram of the application. DETAILED DESCRIPTION

[0024] The technical solutions of the application will be described below in conjunction with the drawings and examples.

[0025] The application fully learns from the related theory of graph neural network, and realizes a risk propagation analysis method of a multi-layer complex coupled power system based on a space-time graph neural network. The application considers that the propagation of fault risk in the power grid not only has time sequence characteristics, but also depends on the topological structure of the power grid. In order to fully utilize the spatial topological information of the power grid and the time sequence characteristics of the risk propagation, the application proposes a risk propagation prediction method based on a space-time graph convolution neural network. First, the graph data is constructed according to the node characteristics and the topological information, and the spatial correlation is extracted for each graph by the graph convolution neural network (GCN). The Long Short-Term Memory (LSTM) and the attention mechanism are used to capture the time sequence association of multiple node data, and the space-time graph convolution unit is used to fuse the spatial and time sequence characteristics. Then, the risk propagation prediction is modeled as a regression problem, and the space-time graph convolution neural network model is trained. The graph neural network is trained based on the historical data, the learned propagation dynamic characteristics are generalized, and the existing data set is used for test analysis. The propagation dynamics of the complex network structure can be predicted in the unknown propagation dynamics and a group of unknown network structures.

[0026] Firstly, the coupling relationship among information-power-transportation network is analyzed. The coupling relationship of power grid, transportation network and information network is as follows Figure 1As shown in the figure. In this coupling, the power, information and transportation systems realize synergy through the coupling between the charging station, 5G base station and EV. The charging station serves as the coupling point between the power grid and the transportation network, providing charging services for electric vehicles EV. The use of EV requires the charging station to provide power supply, and the operation of the charging station also depends on reliable power supply. Therefore, the stability and availability of the power system are crucial for the normal operation of the charging station and the transportation system. At the same time, EV has the dual attributes of source and load based on the vehicle-to-grid (V2G) technology, adjusts its charging and discharging behavior under the influence of charging price, reduces the impact on the power grid and provides power support. The 5G base station serves as the coupling point between the power grid and the information network, providing high-speed and reliable communication services, which is mainly coupled with the power system through power supply and communication interconnection. The power supply of the 5G base station needs the support of a reliable power system, and the power system can monitor and control the energy consumption of the base station through the Internet energy management system to optimize energy utilization efficiency. At the same time, the high-speed communication service provided by the 5G base station provides an important basis for information transmission of the transportation system and EV. The intelligent traffic system (ITS) supported by the V2G technology enables EVs, charging stations and power companies to share real-time information. EV serves as the coupling point between the transportation network and the information network, with a dual role. On the one hand, as a mobile energy storage unit, EV can participate in energy scheduling and energy storage of the power system, and the popularity and use of EV have an impact on the transportation system, such as reducing emissions of traditional fuel vehicles and alleviating traffic congestion. On the other hand, EV can be used as a communication load to transfer the communication load of the base station by adjusting its driving path, thereby reducing the running cost of the communication side.

[0027] In order to fully utilize the topology of the power grid, the application adopts graph convolution to extract the spatial features of the nodes in the power grid, and realizes the prediction of the risk propagation. When applying graph neural network to extract features, the data to be processed is first abstracted and converted into graph structure data. In the real world, the structure information and node features of the graph will change over time, which can be reflected in information networks, power networks and transportation networks, etc. The state of a node at a certain position is not only affected by the nodes at adjacent positions, but also affected by the historical state information of the current node. Therefore, the risk propagation model needs to fully consider the dynamic characteristics of the graph structure and node attributes, and predict the network state at time t+1 according to the changes of the network at time 0 to t.

[0028] The risk propagation analysis method of the multi-layer complex coupled power system based on the spatio-temporal graph neural network of the embodiment of the application is as shown in the figure, which is explained in the following five steps. Figure 2

[0029] Step 1, establish an information-power-transportation three-layer heterogeneous coupled network model. ​

[0030] Based on the above coupling relationship, the mathematical model of the multi-layer coupling system of information-power-transportation is constructed. In the multi-layer coupling system, there is generally weak correlation between heterogeneous network nodes, and the three-layer network architecture is modeled respectively. This step constructs the graph structure of each layer and establishes the relationship between the nodes in the layer, describing the spatial dependence relationship of the nodes in the same layer.

[0031] (1) Construct the upper information network. The information layer is an Ethernet-based communication network. The 5G base station, communication equipment, data channel, etc. are abstractly represented by information nodes, and the signal strength such as reference signal received power, communication range, etc. can be obtained as the characteristics of the nodes. The connection relationship between nodes can be simplified as an adjacency matrix:

[0032] C=(a ij ) k×k (1)

[0033] Wherein, when the information nodes i and j are connected in communication, a ij =1, otherwise a ij =0; k is the number of information nodes.

[0034] (2) Construct the intermediate power network. The bus in the power system is constructed as a graph node, and the electrical line connected between the bus is constructed as the edge in the graph. The electrical data such as current, voltage, voltage phase angle of the bus node can be extracted as the characteristics of the nodes in the graph. The power system is abstracted into a network with m bus nodes, and the adjacency matrix of each bus node is represented as:

[0035] P=(a ij ) m×m (2)

[0036] Wherein, when the bus nodes i and j are connected, a ij =1, otherwise a ij =0; m is the number of bus nodes.

[0037] (3) Construct the lower layer traffic network. The dynamic road network model is used for analysis, which is specifically represented as follows:

[0038]

[0039] Wherein, R is the traffic network, Node is the set of all nodes in the traffic network, which contains n road intersection nodes; P is the set of all directed arcs in R, a directed arc represents a passing road; T is the time series set; U is the set of road traffic flow, u ij,t represents the road traffic flow from node i to node j at time t; D is the set of road lengths, d ijLet L be the road length from node i to node j. For a traffic network with n intersection nodes, assign quantitative values ​​and establish the elements L in the road adjacency matrix L. ij The value can be:

[0040]

[0041] Step 2 involves acquiring node data from the information network, power network, and transportation network. Based on neighborhood and connectivity, a multi-graph structure is constructed, transforming network characteristic data into graph data and abstracting the data to be processed into graph-structured data to represent various temporal and spatial dependencies. Step 1 above constructed the graph structure for each layer and established intra-layer node relationships, describing the spatial dependencies between nodes within the same layer. This step establishes inter-layer node relationships based on neighborhood and connectivity, further describing the inter-layer spatial dependencies.

[0042] For a three-layer heterogeneous coupled network, if there is a direct connection between nodes in different layers, an edge is established between nodes in the two layers, thus creating an adjacency matrix for the nodes between layers. Let the three-layer heterogeneous coupled network established in this invention be graph G, with N nodes in G. In this invention, the connections between nodes in the three-layer complex coupled system, including connections between nodes within and between layers, can be represented by an adjacency matrix, i.e., adjacency matrix A. In adjacency matrix A, node A... ij for:

[0043]

[0044] A neighbor graph defined based on neighbor node connectivity can only reflect a node's first-order neighbors. A node can be connected to geographically distant but reachable locations; topological connectivity is used to represent whether a path exists between a node and other nodes. The method of this invention defines the topological relationship graph of graph G as A. tp Node v i With node v j The edges between them are Internal elements of a topological graph The calculation method is as follows:

[0045]

[0046] In the formula, conn(v i ,v j ) represents node v in graph G i With node v j The connectivity indicator function indicates whether a path exists between two nodes.

[0047] The embodiment of the present application respectively acquires the signal strength of each information node in the information network, the current, voltage and voltage phase angle of each bus node in the power network, and the traffic flow of each road in the traffic network. A unified representation method of node characteristics is set in advance, the characteristics of each node in the three-layer heterogeneous coupled network are represented by using the unified method, the adjacency matrix A and the topological relationship graph A tp characterize the connection relationship of nodes in the three-layer heterogeneous coupled network. The graph data constructed by the embodiment of the present application contains node characteristics and edge connection relationship. The edge connection relationship is represented by using an adjacency matrix or a topological relationship graph, which can be selected according to the data situation in the following processing.

[0048] The method of the present application also normalizes the collected node data before inputting the graph data into the risk propagation prediction model. After normalizing the data in each dimension, the node characteristics are constructed, and the final graph data is generated.

[0049] Step three: build a risk propagation prediction model, which is realized based on a spatio-temporal graph neural network (STGCN).

[0050] The STGCN generally combines a spatial learning network and a temporal learning network together. The risk propagation prediction model of the present application is realized by using a multi-STGCN (MSTGCN). The MSTGCN introduces a dynamic graph structure and a learning mechanism into the STGCN, that is, a data-driven way is used to obtain in-depth insights into the properties of input graph data from an unknown dynamic process. The MSTGCN respectively uses a graph convolutional neural network GCN, a long short-term memory network LSTM and an attention mechanism to capture the time and space correlation for each graph, learns the meta-knowledge of nodes and edges in the multi-layer graph G from the attribute graph, and extracts the spatio-temporal features. The spatial learning network and the temporal learning network are organically combined together by using a spatio-temporal fusion neural network architecture.

[0051] The risk propagation prediction model built by the embodiment of the present application is as shown in Figure 3 The risk propagation prediction model contains a spatial learning network and a temporal learning network. The present application introduces a spatio-temporal multigraph convolutional network (STMGCN) as the spatial learning network to extract the time correlation of node characteristics in the multi-layer coupled network. Figure 3 The temporal learning network is realized by using a TCN+self-attention mechanism. The present application combines a long short-term memory network LSTM and an attention mechanism to realize the temporal learning network to extract the spatial correlation of node characteristics in the multi-layer coupled network, as shown in Figure 4 Figure 3 and 4 ​The risk propagation prediction model of the application is illustrated, the model space learning network comprises k parallel graph convolution layers GCN, the graph data of the three-layer heterogeneous coupled network collected at the first k moments is input into the risk propagation prediction model, the spatial dimension features of the graph data at the k moments are extracted by the k parallel graph convolution layers respectively, then the extracted features are input into the LSTM with attention mechanism in sequence to extract the time dimension features, the extracted features are represented by the hidden state of the LSTM, finally the multi-graph spatio-temporal hidden features extracted at the k moments are fused, and the graph data at the k+1 moment is output, through the sliding window, the graph data of the continuous collection time is continuously input, and the graph data at multiple prediction moments is continuously output.

[0052] The GCN operation of the application is realized by spectral graph theory, and the spatial information of the network G is learned through the topological connection relationship of the graph data.

[0053] L=D-A=I-D -1 / 2 AD -1 / 2 (8)

[0054] In the formula, I is a unit matrix, D∈R N×N is the degree matrix of the graph G, and A is the adjacency matrix of the graph G.

[0055] The result of L through characteristic decomposition is:

[0056] L=UΛU T (9)

[0057] In the formula, Λ is a diagonal matrix composed of eigenvalues; U is a Fourier base; and the upper index T represents transposition. Taking the graph node feature x at the t moment as an example, the graph Fourier transform of the signal is defined as:

[0058]

[0059] The graph convolution is realized by replacing the classical convolution operator with a diagonalized linear operator in the Fourier domain. Therefore, the convolution formula of the graph signal is as follows:

[0060] g θ * x=g θ (L)x=g θ (UΛU T )x=Ug θ (Λ)U T x (11)

[0061] In the formula, g θ represents the graph signal convolution operation.

[0062] AsFigure 4 As shown, the graph data of the multi-layer heterogeneous coupling network is collected in time sequence and normalized, and then the normalized data is input into the risk propagation prediction model implemented by the MSTGCN, in which the graph data is first subjected to graph convolution operation to extract spatial features, and the extracted graph features are input into the time sequence prediction module in time sequence in series to model the time correlation. The time sequence prediction model is realized by combining the long short-term memory network (LSTM) and the attention mechanism, and the long short-term time correlation pattern of the time sequence is captured by the LSTM. The combination of the two can obtain a spatio-temporal graph neural network model with spatial and temporal modeling capabilities.

[0063] The time sequence prediction module of the present application, i.e. the time learning network, has the input gate, the forget gate and the output gate of the LSTM, but is derived from the graph convolution operator, and the attention mechanism is introduced. The purpose of introducing the attention mechanism is to enhance the information of key nodes, as shown in formula (12):

[0064]

[0065] Wherein, sigma(·) is a sigmoid function, and is a same or operator, i, f, o and c represent the input gate, the forget gate, the output gate and the cell state vector, respectively, and when each of them is updated, there are corresponding trainable weights W and bias vectors b, such as the weight W f and the bias vector b f of f att represent the attention network, which can enhance the information of key nodes while ensuring the integrity of the information, The result is an attention matrix, is the hidden state of the LSTM output at time t, and h t is the hidden state of the LSTM output at time t with the introduction of the attention mechanism.

[0066] The attention matrix is set as V=(V1, V2, V3, … V N ), V t is a column vector, and the calculation formula is shown in formula (13).

[0067]

[0068] In formula (13), the attention matrix V is finally obtained by normalization through the softmax(·) activation function. t In semantics, it is understood as the degree of interdependence between nodes at output time t, K and b are the weight matrix and bias vector of the attention network, and tanh is the activation function.

[0069] Step four, using multi-dimensional time series, train risk propagation prediction model based on historical data for dynamic behavior prediction. By coupling node dynamics propagation model to describe the process of fault propagation in heterogeneous network. That is, the network can adaptively update the graph structure according to the change of spatio-temporal data, and then analyze the dynamic propagation process in heterogeneous network.

[0070] The load in many practical systems is in a dynamic process of continuous change, especially when nodes are added or removed, the load in the network will be redistributed. In the intra-layer nodes, the load-capacity model is selected as the basis for research. The carrying capacity C(i) of a node i can be represented as:

[0071] C(i) = (1 + a)L0 (14) In the formula, a is the redundancy parameter, and L0 represents the initial load at node i. The greater the value of redundancy a, the stronger the ability to resist risks and failures.

[0072] Assume that a node i in network A is attacked, and after the node fails, the load is redistributed, as shown in Figure 5 , node i will allocate its load to adjacent nodes j in network A according to the proportion , as follows

[0073]

[0074] In the formula, n ∈ T i represents that n belongs to the adjacent nodes of node i, T i is the set of adjacent nodes of node i, are the carrying capacities of nodes j and n in network A, respectively.

[0075] Based on the virus propagation model, the risk propagation dynamic process between inter-layer nodes in heterogeneous network is described as follows:

[0076] p 1,i (t+1) = (1-p 1,i (t))(1-q 1,i (t)) + (1-β1)p 1,i (t) + γ1p 2,i (t)(1-p 1,i (t)) (16)

[0077] p 2,i (t+1) = (1-p 2,i (t))(1-q 2,i (t)) + (1-β2)p 2,i (t) + γ2p 1,i (t)(1-p 2,i (t)) (17)

[0078] wherein q 1,i (t), q 2,i (t) respectively represent the probability that node i is not infected by the neighboring nodes in network A and network B in a failure state, and 1-q 1,i (t), 1-q 2,i (t) respectively represent the probability that node i is infected by the neighboring nodes in network A and network B; p 1,i (t+1), p 2,i (t+1) respectively represent the probability that node i in network A and network B is in an abnormal state at time t+1, p 1,i (t), p 2,i (t) respectively represent the probability that node i in network A and network B is in a failure state at time t. Therefore, (1-p 1,i (t)), (1-p 2,i (t)) respectively represent the probability that node i is in a normal operation state, and the probabilities of being infected by the neighboring nodes in the layer are (1-q 1,i (t)) and (1-q 2,i (t)), β1, γ1, β2, γ2 are weight parameters, (1-β1)p 1,i (t), (1-β2)p 2,i (t) respectively represent the probability that node i is not recovered and operated after being infected with a failure at time t, γ1p 2,i (t)(1-p 1,i (t)), γ2p 1,i (t)(1-p 2,i (t)) respectively represent the probability that node i is infected by the neighboring nodes in another layer network in a failure state. In the formula, it is defined that

[0079]

[0080] wherein α1, α2 respectively represent the influence factors of network A and network B, a ij represents the edge relationship between node i and node j in the adjacency matrix of network A, b ij represents the edge relationship between node i and node j in the adjacency matrix of network B, and N1 and N2 are the number of nodes in network A and network B, respectively.

[0081] Thus, for different failures, different parameters can be set to analyze the propagation dynamics process in the heterogeneous network.

[0082] Step five: the trained risk propagation prediction model can generalize the learned propagation dynamic characteristics, and can predict the propagation dynamics of complex systems in unknown propagation dynamics and a set of unknown network structures, such as predicting the charging and discharging behavior of electric vehicles based on graph time series network, and predicting the risk propagation path in complex coupled networks.

[0083] One implementation process of the multi-layer coupled power system risk propagation prediction method of the application includes the following four links:

[0084] (1) Sample acquisition: respectively acquire the voltage, current, phase angle, traffic volume, signal strength and other information X (X1, X2, …, X R ) of each bus node in the sample in the power system, traffic network and information network under normal operation and various fault states.

[0085] (2) Graph data processing: based on the analysis of network topology, inter-network coupling relationship, time dependence and spatial dependence described above, the data collected by the three coupled systems is converted into graph data information. When converting each information into node information of graph data, normalization is needed, and Z-score standardization is adopted.

[0086]

[0087] In the formula: is the mean value of the data information X r carried by the node; σ r is the standard deviation of the information X; r ∈ [1, R]. After normalization of the node characteristics, the graph data required by the model is composed in combination with the adjacency matrix.

[0088] (3) Model training: in the training stage, the risk propagation prediction model is built, and the hyperparameters for training are set, including batch size, learning rate and iteration number, etc. After training, the accuracy and prediction error of the propagation model are verified. In the application, convolution is adopted to realize parallelization at the input end, and the number of parameters is less and the training speed is faster.

[0089] The training target of the model is to minimize the error between the true value Y pred and the predicted value , and the corresponding loss function is

[0090]

[0091] In the formula, the first term ensures the minimization of error, and the second term L reg is a regularization term that can effectively avoid overfitting, and λ is a hyperparameter.

[0092] (4) Model application: in the model application stage, for the actual fault of the power grid, the voltage phase angle data of each bus node is obtained, the fault position is input, and then the voltage phase angle data of each bus node obtained is input into the corresponding risk propagation prediction model, the propagation path of the fault risk in the system in the future period of time is predicted, and the key risk node is located.

[0093] In addition to the technical features described in the specification, they are known to those skilled in the art. The present application omits the description of known components and known technologies to avoid redundancy and unnecessary limitation of the present application. The embodiments described in the above embodiments do not represent all embodiments consistent with the present application. Various modifications or variations made by those skilled in the art without creative labor on the basis of the technical solutions of the present application are still within the protection scope of the present application.

Claims

1. A method for risk propagation analysis of a multi-layer complex coupled power system based on a spatio-temporal graph neural network, comprising the following steps: Step 1: Establishing a three-layer heterogeneous coupled network G of information-power-transportation, including an upper-layer information network, a middle-layer power network and a lower-layer transportation network; the information network is constructed based on a communication network of an Ethernet, the nodes of the graph include base stations and communication devices, and edges are established when the nodes are connected for communication; the power network is constructed based on a power system, the nodes of the graph are busbars, and the electrical lines connecting the busbars are constructed as edges of the graph; the transportation network is constructed using a dynamic road network model, the nodes of the graph are intersections, and the edges of the graph are established according to the lanes of the road network; if there is a connection relationship between nodes of different layers, edges between the nodes of different layers are established; Step 2: Collecting physical quantities in the three-layer heterogeneous coupled network and constructing graph data of the network G, including node features and edge relationships; the collected physical quantities include signal strength of each node in the information network, current, voltage and voltage phase angle of the busbars in the power network, and vehicle flow of each road in the transportation network; after the collected data are normalized in the same dimension, a pre-set unified representation is used to represent the features of the nodes; According to the edge relationship between the nodes in each layer in the network G and the edge relationship between the nodes in each layer, an adjacency matrix A of G and a topological relationship graph A are established tp ; using the adjacency matrix A or the topological relation graph A tp characterizing the edge relation of G; Step 3: Building a risk propagation prediction model based on a spatio-temporal multi-graph neural network; the spatio-temporal multi-graph neural network of the risk propagation prediction model includes a spatial learning network and a temporal learning network, the spatial learning network includes k parallel graph convolution layers, and the temporal learning network is realized by combining a long short-term memory network (LSTM) and an attention mechanism; the graph data of the three-layer heterogeneous coupled network at the first k time points are input into the risk propagation prediction model, the k parallel graph convolution layers are used to extract spatial dimension features from the graph data at the k time points respectively, the extracted features are input into the LSTM with the attention mechanism in sequence to extract time dimension features, the extracted features are represented by hidden states of the LSTM, and finally the features extracted at the k time points are fused to output graph data at a (k+1) th time point; k≥3; Step 4: Obtaining historical data of the three-layer heterogeneous coupled network, including historical graph data of each node of the network in a normal operating state, and historical graph data of the power system, the transportation network and the information network in different fault states, inputting the historical data into the risk propagation prediction model for training until the model reaches a set accuracy requirement; during the training process, the graph structure of the three-layer heterogeneous coupled network is updated adaptively according to the collected time data; Step 5: When a fault occurs in a node of the network, the trained risk propagation prediction model is used to predict the risk propagation path in the three-layer heterogeneous coupled network of information-power-transportation.

2. The method of claim 1, wherein, The step one, when constructing the traffic network, obtaining all road set P and road length set U, calculating the element L of the adjacent matrix L ij The value as follows: where i, j represent the number of two intersections, ij represents the road from intersection i to intersection j, d ij represents the length of road ij.

3. The method of claim 1, wherein, In the second step, the collected data is normalized in the same dimension, including: setting the R physical quantity information of the acquisition node as X, and the rth physical quantity X r The Z-score is used for standardization is expressed as and σ r respectively, the mean and standard deviation of the rth physical quantity; r ∈ [1, R].

4. The method of claim 1, wherein, In the third step, the time learning network is realized by combining a long short-term memory network (LSTM) and an attention mechanism, the information of the key nodes is enhanced through the introduced attention mechanism, and f att The time learning network outputs the hidden state h t is represented as: wherein, is the hidden state of the LSTM output at time t, h t is the hidden state of the LSTM output at time t with the attention mechanism introduced; is the attention matrix, denoted by V, where the column vector V t is calculated as follows: Wherein, softmax and tanh are activation functions, K is a weight matrix of the attention network, and b is a bias vector of the attention network.

5. The method of claim 1, wherein, In step 4, the training network updates the graph structure adaptively according to the changes of the spatio-temporal data, and describes the fault propagation process in the heterogeneous network through a coupled node dynamics propagation model, including: (1) When a node is added or removed in the network, the load in the network will be redistributed, and the graph structure will be updated; when a node i in a certain layer network A is attacked, the node i will face the redistribution of the load after its failure, and the node i will proportionally distribute its own load to the adjacent nodes j in the network A as follows: Loadi = Loadi / (Loadi + Loadj) where T i is a set of neighboring nodes of node i, are the carrying capacities of node j, node n in network A, respectively; (2) Describing the fault propagation process in the heterogeneous network through a coupled node dynamics propagation model, as follows: p 1,i (t+1) = (1 - p 1,i (t))(1 - q 1,i (t)) + (1 - β1)p 1,i (t) + γ1p 2,i (t)(1 - p 1,i (t)); p 2,i (t+1) = (1 - p 2,i (t))(1 - q 2,i (t)) + (1 - β2)p 2,i (t) + γ2p 1,i (t)(1 - p 2,i (t)); where p 1,i (t+1) and p 2,i (t+1) represent the probability of node i in network A and network B being in a failure state at time t+1, respectively, p 1,i (t) and p 2,i (t) represent the probability of node i in network A and network B being in a failure state at time t, respectively, q 1,i (t) and q 2,i (t) represent the probability of node i not being infected by the neighbor nodes in a failure state in network A and network B, respectively; β1, γ1, β2, and γ2 are weight parameters; q 1,i (t) and q 2,i (t) are calculated as follows: wherein, a1, a2 represent the influence factor of network A, network B respectively, a ij represents the edge relationship between node i and node j in the adjacency matrix of network A, b ij represents the edge relationship between node i and node j in the adjacency matrix of network B.

Images