A probabilistic power flow evaluation method and device based on edge graph attention network

Abstract

The application discloses a kind of based on edge diagram attention network's probabilistic power flow evaluation method and device, based on power system uncertainty factor, extraction sample value, the sample value is orthogonalized and sorted, obtain final sample, the final sample is modeled as graph data structure, the graph data structure is input by the power flow solver of multilayer edge diagram attention network graph filter stack, flow distribution is obtained by calculation, based on the flow distribution, realize probabilistic power flow evaluation.The application introduces the method of graph neural network in probabilistic power flow evaluation, by deep learning fitting mapping relationship between input and output, without iteration in calculation process, calculation speed is extremely fast, can effectively solve the problem of long time consumption of traditional power flow calculation, can retain the advantages of high calculation precision of Monte Carlo probabilistic power flow risk assessment method, while reducing its time consumption to some extent.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power grid evaluation, planning and scheduling, and particularly relates to a probabilistic power flow evaluation method and device based on an edge graph attention network. BACKGROUND

[0002] With the development of new power systems with high proportions of renewable energy and power electronic devices, new energy such as solar and wind energy is connected to the power grid, which brings stronger intermittency and randomness to the power grid. The probabilistic power flow calculation method is an important basis for solving the uncertainty factors of the power system, which helps to comprehensively evaluate the power grid under various operating conditions, and quantitatively analyze the weak links and potential dangers of the power grid (such as: over-limit of reactive power, line overload and power flow distribution, etc.), and provides a reference for reasonably determining the power system planning and design scheme and operation mode.

[0003] However, the traditional Monte Carlo probabilistic power flow calculation method has the problem of slow solution speed while having high calculation accuracy, which is particularly prominent when facing systems with high uncertainty. Therefore, how to improve the solution speed while ensuring the calculation accuracy of the probabilistic power flow has been a concern for people. SUMMARY

[0004] Therefore, the present application provides a probabilistic power flow evaluation method and device based on an edge graph attention network, so that when evaluating the probabilistic power flow, the solution speed can be improved to some extent while ensuring the calculation accuracy of the probabilistic power flow.

[0005] In order to achieve the above-mentioned purpose, the present scheme is as follows:

[0006] A probabilistic power flow evaluation method based on an edge graph attention network, comprising:

[0007] Based on the uncertainty factors of the power system, sample values are extracted;

[0008] The sample values are sorted orthogonally to obtain final samples;

[0009] The final samples are modeled as a graph data structure;

[0010] The graph data structure is input into a power flow solver composed of a multilayer edge graph attention network (Edge Aggregated Graph Attention Networks and Transfer Learning, EGAT) graph filter stack to calculate the power flow distribution;

[0011] Based on the power flow distribution, the probabilistic power flow evaluation is realized.

[0012] Optionally, it further comprises:

[0013] The risk probability is calculated based on the power flow distribution.

[0014] Optionally, the step of extracting sampled values ​​based on power system uncertainties includes:

[0015] Based on the uncertainties of the power system, a K×N dimensional random sequence matrix T and a K×N dimensional uniform random sampling matrix Q are generated.

[0016] Among them, t kn The nth sample value t from the kth sampling. kn ∈(0,1); q kn This represents the nth sample value of the kth random variable.

[0017] The cumulative probability distribution function for each sampled value is determined using the following formula:

[0018]

[0019] Based on the cumulative probability distribution function of each sample value, the sample value is extracted using the Latin hypercube method.

[0020] Optionally, the step of orthogonally sorting the sampled values ​​to obtain the final sample includes:

[0021] Generate a K×N matrix L = [L1, L2, ..., L K ] T The elements of each row of the matrix are random permutations of integers 1, 2, ..., N;

[0022] The matrix is ​​iterated in both forward and reverse directions by alternating forward and reverse operations.

[0023] When the root mean square value ρ represents the correlation of matrix columns ms The iteration stops when the number of iterations ceases or the number of forward and reverse iterations reaches a preset value.

[0024]

[0025] in:

[0026]

[0027] Where: cov(·) is the covariance; var(·) is the variance, representing the vector L. k With L j The correlation between them.

[0028] Optionally, the final sample is modeled as a graph data structure, including:

[0029] Using the final sample, a power network diagram with N buses and M transmission channels is constructed, with the bus as the node and the transmission channel as the edge. The power network diagram is defined as G = {V, E}.

[0030] node v i ∈V is the i-th node in the graph;

[0031] The initial power flow information corresponding to the i-th node is represented by the node feature vector;

[0032] Among them, U i Represents voltage amplitude, θ i Represents voltage phase angle, P Li Indicates active load, Q Li Represents reactive load, P Gi Indicates contribution, Q Gi This indicates that no effort was put in;

[0033] All node features can be represented by a node feature matrix.

[0034] Side e i ∈E represents the i-th branch in the graph;

[0035] The line conductance and susceptance information of the i-th branch are represented by the edge feature vector;

[0036] All edge features are represented by the edge feature matrix.

[0037] The power grid topology is constructed using the node adjacency matrix A. N×N Let A represent the element A in the i-th row and j-th column. ij satisfy:

[0038]

[0039] Optional, also includes:

[0040] All feature vectors are preprocessed according to the following formula to obtain the preprocessed feature vectors;

[0041]

[0042] Among them, h max h is the maximum value of the eigenvector. min The minimum value of the eigenvector. The feature line vectors are linearly normalized and then preprocessed. The elements in the set are dimensionless numbers with a range between [0,1].

[0043] Optionally, the step of inputting the graph data structure into a power flow solver composed of stacked multi-layer EGAT graph filters to calculate the power flow distribution includes:

[0044] Construct the neighborhood message matrix for each node, and node i obtains messages from its neighboring nodes j∈N(i). for:

[0045]

[0046] Among them, W v ∈Ρ 6×D W e ∈Ρ 2×D The parameter matrix is ​​the learnable parameter matrix, and D∈P is the feature dimension of the output node of the graph filter;

[0047] Calculate the attention coefficient of each node to its neighboring nodes, and the attention coefficient a of node i to its neighboring node j. ij for:

[0048]

[0049] Where N(i) are the neighbors of node i; For attention mechanisms;

[0050] A message passing aggregation mechanism based on edge features is used to update node features. The feature update calculation method for node i is as follows:

[0051]

[0052] In the formula, Let K be the feature vector of node i after the update; K is the number of multi-head attention used; σ(·) is the activation function.

[0053] The power flow distribution is calculated based on the updated node characteristics.

[0054] Optionally, the graph data structure contains several nodes, and the calculation of the risk probability based on the power flow distribution includes:

[0055] Based on the power flow distribution, determine the voltage value V of node i. i Whether it exceeds a predetermined limit value, said limit value including: the upper limit of allowable voltage fluctuation V imax and the lower limit of permissible voltage fluctuation V imin ;

[0056] If so, the power flow distribution is calculated n times, and the risk probability is calculated using the following formula;

[0057]

[0058] Where m is the number of times the voltage exceeds the limit in the n power flow distribution calculations, and n is the total number of power flow distribution samples.

[0059] Optionally, the graph data structure contains several nodes, and the calculation of the risk probability based on the power flow distribution includes:

[0060] Based on the power flow distribution, determine the terminal power flow value I of node i. i Does it exceed the predetermined upper limit value I of the branch power flow? imax ;

[0061] If so, the power flow distribution is calculated n times, and the risk probability is calculated using the following formula;

[0062]

[0063] Where x is the number of times the branch power flow exceeds the limit in n power flow distribution calculations, and n is the total number of power flow distribution samples.

[0064] A probabilistic power flow evaluation device based on a side graph attention network includes:

[0065] The sample value extraction module is used to extract sample values ​​based on uncertainties in the power system.

[0066] The final sample determination module is used to orthogonally sort the sampled values ​​to obtain the final sample.

[0067] The modeling module is used to model the final sample as a graph data structure;

[0068] The power flow distribution calculation module is used to input the graph data structure into a power flow solver composed of stacked multi-layer EGAT graph filters to calculate the power flow distribution.

[0069] The probabilistic power flow assessment module is used to perform probabilistic power flow assessment based on the power flow distribution.

[0070] As can be seen from the above technical solutions, the probabilistic power flow assessment method and apparatus based on edge graph attention networks provided in this application extracts sampled values ​​based on power system uncertainties, orthogonally sorts the sampled values ​​to obtain final samples, models the final samples as a graph data structure, inputs the graph data structure into a power flow solver composed of stacked multi-layer EGAT graph filters, calculates the power flow distribution, and realizes probabilistic power flow assessment based on the power flow distribution. This application introduces a graph neural network method into probabilistic power flow assessment, fitting the mapping relationship between input and output through deep learning. The calculation process requires no iteration and is extremely fast, effectively solving the problem of long calculation time in traditional power flow calculations. It retains the high computational accuracy advantage of the Monte Carlo probabilistic power flow risk assessment method while reducing its time consumption to a certain extent. Attached Figure Description

[0071] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0072] Figure 1 A flowchart of a probabilistic power flow evaluation method based on a side graph attention network provided in an embodiment of this application;

[0073] Figure 2 A structural diagram of a probabilistic power flow evaluation device based on a side graph attention network is provided for an embodiment of this application;

[0074] Figure 3 This is a hardware structure block diagram of a probabilistic power flow evaluation device based on a side graph attention network, provided in an embodiment of this application. Detailed Implementation

[0075] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0076] Figure 1 A flowchart of a probabilistic power flow evaluation method based on a side graph attention network is provided for embodiments of this application. (Refer to...) Figure 1 The method may include the following steps:

[0077] Step S100: Extract sampled values ​​based on uncertainties in the power system.

[0078] Specifically, uncertainties in the power system are treated as random variables to establish a probability model, and samples of the probability distribution are drawn.

[0079] Step S101: Orthogonally sort the sampled values ​​to obtain the final sample.

[0080] Specifically, during the sampling process, the sampling matrix also needs to be arranged and Gram-Schmidt orthogonalized.

[0081] Step S102: Model the final sample as a graph data structure.

[0082] Specifically, in order to input sample data into a power flow solver composed of stacked multi-layer EGAT graph filters, the samples can be pre-modeled as a graph data structure.

[0083] Step S103: Input the graph data structure into a power flow solver composed of stacked multi-layer EGAT graph filters to calculate the power flow distribution.

[0084] Specifically, the power flow distribution is calculated by inputting the graph data structure into a power flow solver composed of stacked multi-layer EGAT graph filters. The power flow solver composed of stacked multi-layer EGAT graph filters fits the mapping relationship between input and output through deep learning. No iteration is required during the calculation process, and the calculation speed is extremely fast.

[0085] Step S104: Based on the power flow distribution, perform probabilistic power flow assessment.

[0086] As can be seen from the above technical solutions, the probabilistic power flow assessment method based on a side graph attention network provided in this application extracts sampled values ​​based on power system uncertainties, orthogonally sorts these sampled values ​​to obtain final samples, models these final samples as a graph data structure, and inputs this graph data structure into a power flow solver composed of stacked multi-layer EGAT graph filters to calculate the power flow distribution. Based on this power flow distribution, probabilistic power flow assessment is achieved. This application introduces a graph neural network method into probabilistic power flow assessment, fitting the mapping relationship between input and output through deep learning. The calculation process requires no iteration and is extremely fast, effectively solving the problem of long calculation time in traditional power flow calculations. It retains the high computational accuracy advantage of the Monte Carlo probabilistic power flow risk assessment method while reducing its time consumption to a certain extent.

[0087] In some embodiments of this application, the process of extracting sampled values ​​based on power system uncertainties in step S100 is described, which may specifically include the following steps:

[0088] S11. Based on the uncertainties of the power system, generate a K×N dimensional random sequence matrix T and a K×N dimensional uniform random sampling matrix Q.

[0089] Among them, t kn The nth sample value t from the kth sampling. kn ∈(0,1); q kn This represents the nth sample value of the kth random variable.

[0090] S12. Determine the cumulative probability distribution function of each sampled value according to the following formula.

[0091]

[0092] S13. Based on the cumulative probability distribution function of each sample value, the sample value is extracted using the Latin hypercube method.

[0093] Latin hypercube sampling represents a recent advancement in sampling techniques. Compared to Monte Carlo methods, it is designed to accurately reconstruct the input distribution with fewer iterations. The key to Latin hypercube sampling is stratifying the input probability distribution. Stratification divides the cumulative curve into equal intervals on a cumulative probability scale (0 to 1). Samples are then randomly drawn from each interval or "stratum" of the input distribution.

[0094] In some embodiments of this application, the process of orthogonally sorting the sampled values ​​to obtain the final sample in step S101 is described, which may specifically include the following steps:

[0095] S21. Generate a K×N matrix L = [L1, L2, ..., L K ] T The elements in each row of the matrix are random permutations of integers 1, 2, ..., N.

[0096] S22. The matrix is ​​iterated in both forward and reverse directions by alternating forward and reverse operations.

[0097] S23, When the root mean square value ρ representing the correlation of matrix columns ms The iteration stops when the number of iterations no longer decreases or the number of forward and reverse iterations reaches a preset value.

[0098]

[0099] in:

[0100]

[0101] Where: cov(·) is the covariance; var(·) is the variance, representing the vector L. k With L j The correlation between them.

[0102] In some embodiments of this application, the process of modeling the final sample as a graph data structure in step S102 is described, which may specifically include the following steps:

[0103] Using the final sample, a power network diagram with N buses and M transmission channels is constructed, with the bus as the node and the transmission channel as the edge. The power network diagram is defined as G = {V, E}.

[0104] node v i ∈V is the i-th node in the graph;

[0105] Let be the feature vector of the initial power flow information node corresponding to the i-th node. Let be the feature vector of the initial power flow information node corresponding to the j-th node;

[0106] Among them, U i Represents voltage amplitude, θ i Represents voltage phase angle, P Li Indicates active load, Q Li Represents reactive load, P Gi Indicates contribution, Q Gi This indicates that no effort was put in;

[0107] All node features can be represented by a node feature matrix.

[0108] Side e i ∈E represents the i-th branch in the graph;

[0109] Let be the edge feature vector containing the line conductance and susceptance information of the i-th branch. Let be the edge feature vector of the edge between the i-th node and the j-th node;

[0110] All edge features are represented by the edge feature matrix.

[0111] The power grid topology is constructed using the node adjacency matrix A. N×N Let A represent the element A in the i-th row and j-th column. ij satisfy:

[0112]

[0113] Since the node feature vectors involve values ​​of three different dimensions—voltage, phase angle, and power—linear normalization can be used to preprocess the feature vectors to eliminate the influence of these dimensions. Specifically, all feature vectors can be preprocessed according to the following formula to obtain the preprocessed feature vectors.

[0114]

[0115] Among them, h max h is the maximum value of the eigenvector. min The minimum value of the eigenvector. The feature line vectors are linearly normalized and then preprocessed. The elements in the set are dimensionless numbers with a range between [0,1].

[0116] In some embodiments of this application, the process of inputting the graph data structure into a power flow solver composed of stacked EGAT graph filters to calculate the power flow distribution in step S103 is described, and may specifically include the following steps:

[0117] S31. Construct the neighborhood message matrix for each node, and node i obtains the messages of its neighboring node j∈N(i). for:

[0118]

[0119] Among them, W v ∈Ρ 6×D W e ∈Ρ 2×D Let P be the learnable parameter matrix, and D∈P be the feature dimension of the output node of the graph filter.

[0120] S32. Calculate the attention coefficient of each node to its neighboring nodes, and the attention coefficient a of node i to its neighboring node j. ij for:

[0121]

[0122] Where N(i) are the neighbors of node i; This is an attention mechanism.

[0123] S33. A message passing aggregation mechanism based on edge features is used to update node features. The feature update calculation method for node i is as follows:

[0124]

[0125] In the formula, Let be the feature vector of node i after the update; K is the number of multi-head attention used; σ(·) is the activation function.

[0126] S34. Based on the updated node characteristics, the power flow distribution is calculated.

[0127] Current solutions for uncertainties in new power systems have improved computational accuracy and speed, making them more suitable for meeting the requirements of power system operation, planning, protection, and stability analysis. Furthermore, risk probabilities can be calculated based on power flow distribution.

[0128] There are multiple ways to calculate the risk probability based on power flow distribution in the above embodiments. This embodiment introduces several optional implementation methods as follows:

[0129] The first method is to use the node voltage values ​​for calculation.

[0130] S41. Based on the power flow distribution, determine the voltage value V of node i.i Whether it exceeds a predetermined limit value.

[0131] Specifically, the aforementioned limits may include: the upper limit of permissible voltage fluctuation V imax and the lower limit of permissible voltage fluctuation V imin ;

[0132] S42. If so, then repeat the power flow distribution n times and use the following formula to calculate the risk probability;

[0133]

[0134] Where m is the number of times the voltage exceeds the limit in the n power flow distribution calculations, and n is the total number of power flow distribution samples.

[0135] The second method is to use the end-flow power values ​​of the nodes for calculation.

[0136] S51. Based on the power flow distribution, determine the terminal power flow value I of node i. i Does it exceed the predetermined upper limit value I of the branch power flow? imax .

[0137] S52. If so, then repeat the power flow distribution n times and use the following formula to calculate the risk probability;

[0138]

[0139] Where x is the number of times the branch power flow exceeds the limit in n power flow distribution calculations, and n is the total number of power flow distribution samples.

[0140] The following describes a probabilistic power flow assessment device based on a side graph attention network provided in an embodiment of this application. The probabilistic power flow assessment device based on a side graph attention network described below and the probabilistic power flow assessment method based on a side graph attention network described above can be referred to and correspond to each other.

[0141] Figure 2 A schematic diagram of a probabilistic power flow evaluation device based on a side graph attention network is provided for an embodiment of this application, as shown below. Figure 2 As shown, a probabilistic power flow evaluation device based on a side graph attention network may include:

[0142] The sampling value extraction module 10 is used to extract sampling values ​​based on the uncertainties of the power system;

[0143] The final sample determination module 20 is used to orthogonally sort the sampled values ​​to obtain the final sample;

[0144] Modeling module 30 is used to model the final sample as a graph data structure;

[0145] The power flow distribution calculation module 40 is used to input the graph data structure into a power flow solver composed of stacked multi-layer EGAT graph filters to calculate the power flow distribution.

[0146] The probabilistic power flow assessment module 50 is used to perform probabilistic power flow assessment based on the power flow distribution.

[0147] As can be seen from the above technical solution, the probabilistic power flow assessment device based on a side graph attention network provided in this application involves a sampling value extraction module 10 extracting sampling values ​​based on power system uncertainties, a final sample determination module 20 orthogonally sorting the sampling values ​​to obtain final samples, a modeling module 30 modeling the final samples as a graph data structure, a power flow distribution calculation module 40 inputting the graph data structure into a power flow solver composed of stacked multi-layer EGAT graph filters to calculate the power flow distribution, and a probabilistic power flow assessment module 50 performing probabilistic power flow assessment based on the power flow distribution. This application introduces a graph neural network method into probabilistic power flow assessment, fitting the mapping relationship between input and output through deep learning. The calculation process requires no iteration and is extremely fast, effectively solving the problem of long calculation time in traditional power flow calculations. It retains the high computational accuracy advantage of the Monte Carlo probabilistic power flow risk assessment method while reducing its time consumption to a certain extent.

[0148] This application also provides a probabilistic power flow evaluation device based on a side graph attention network. Figure 3 The hardware block diagram of the probabilistic power flow evaluation device based on the side graph attention network is shown. Figure 3 The hardware structure of the probabilistic power flow evaluation device based on edge graph attention network may include: at least one processor 1, at least one communication interface 2, at least one memory 3 and at least one communication bus 4.

[0149] In this embodiment of the application, the number of processor 1, communication interface 2, memory 3, and communication bus 4 is at least one, and processor 1, communication interface 2, and memory 3 communicate with each other through communication bus 4;

[0150] Processor 1 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention.

[0151] Memory 3 may include high-speed RAM, and may also include non-volatile memory, such as at least one disk storage device;

[0152] The memory stores a program, which the processor can call. The program is used to implement the various processing steps in the aforementioned probabilistic power flow evaluation method based on edge graph attention networks.

[0153] This application embodiment also provides a storage medium that can store a program suitable for processor execution, the program being used to implement each processing flow in the aforementioned probabilistic power flow evaluation method based on edge graph attention networks.

[0154] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0155] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The various embodiments can be combined with each other, and the same or similar parts can be referred to each other.

[0156] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A probabilistic power flow evaluation method based on edge graph attention networks, characterized in that, include: Based on the uncertainties in the power system, sampled values ​​are extracted; The sampled values ​​are orthogonally sorted to obtain the final samples; The final sample is modeled as a graph data structure; The graph data structure is input into a power flow solver consisting of stacked graph filters from a multi-layered edge-aggregated graph attention network (EGAT) to calculate the power flow distribution. Based on the power flow distribution, probabilistic power flow assessment is achieved; The step of modeling the final sample as a graph data structure includes: Using the final sample, a power network diagram with N buses and M transmission channels is constructed, with the bus as the node and the transmission channel as the edge. The power network diagram is defined as G = {V, E}. node v i ∈V is the i-th node in the graph; Let i be the power flow initial value information node feature vector corresponding to the i-th node; Among them, U i Represents voltage amplitude, θ i Represents voltage phase angle, P Li Indicates active load, Q Li Represents reactive load, P Gi Indicates contribution, Q Gi This indicates that no effort was put in; All node features can be represented by a node feature matrix. Side e i ∈E represents the i-th branch in the graph; Let be the edge feature vector containing the line conductance and susceptance information of the i-th branch; All edge features are represented by the edge feature matrix. The power grid topology is constructed using the node adjacency matrix A. N×N Let A represent the element A in the i-th row and j-th column. ij satisfy:

2. The method according to claim 1, characterized in that, Also includes: Based on the power flow distribution, the risk probability is calculated.

3. The method according to claim 1, characterized in that, The sampling values ​​extracted based on power system uncertainties include: Based on the uncertainties of the power system, a K×N dimensional random sequence matrix T and a K×N dimensional uniform random sampling matrix Q are generated. Among them, t kn The nth sample value t from the kth sampling. kn ∈(0,1), q kn This represents the nth sample value of the kth random variable. The cumulative probability distribution function for each sampled value is determined using the following formula: Based on the cumulative probability distribution function of each sample value, the sample value is extracted using the Latin hypercube method.

4. The method according to claim 1, characterized in that, The step of orthogonally sorting the sampled values ​​to obtain the final samples includes: Generate a K×N matrix L = [L1, L2, ..., L K ] T The elements of each row of the matrix are random permutations of integers 1, 2, ..., N; The matrix is ​​iterated in both forward and reverse directions by alternating forward and reverse operations. When the root mean square value ρ represents the correlation of matrix columns ms The iteration stops when the number of iterations ceases or the number of forward and reverse iterations reaches a preset value. in: Where: cov(·) is the covariance; var(·) is the variance, representing the vector L. k With L j The correlation between them.

5. The method according to claim 1, characterized in that, Also includes: All feature vectors are preprocessed according to the following formula to obtain the preprocessed feature vectors; Among them, h max h is the maximum value of the eigenvector. min The minimum value of the eigenvector. The feature line vectors are linearly normalized and then preprocessed. The elements in the set are dimensionless numbers with a range between [0,1].

6. The method according to claim 1, characterized in that, The step of inputting the graph data structure into a power flow solver composed of stacked multi-layer EGAT graph filters to calculate the power flow distribution includes: Construct the neighborhood message matrix for each node, and node i obtains messages from its neighboring nodes j∈N(i). for: Among them, W v ∈Ρ 6×D W e ∈Ρ 2×D Let P be the learnable parameter matrix, and D∈P be the feature dimension of the output node of the graph filter. Let be the feature vector of the initial power flow information node corresponding to the i-th node. Let be the feature vector of the initial power flow information node corresponding to the j-th node. Let be the edge feature vector of the edge between the i-th node and the j-th node; Calculate the attention coefficient of each node to its neighboring nodes, and the attention coefficient a of node i to its neighboring node j. ij for: Where N(i) are the neighbors of node i; For attention mechanisms; A message passing aggregation mechanism based on edge features is used to update node features. The feature update calculation method for node i is as follows: In the formula, Let K be the feature vector of node i after the update; K is the number of multi-head attention used; σ(·) is the activation function. The power flow distribution is calculated based on the updated node characteristics.

7. The method according to claim 2, characterized in that, The graph data structure contains several nodes, and the risk probability is calculated based on the power flow distribution, including: Based on the power flow distribution, determine the voltage value V of node i. i Whether it exceeds a predetermined limit value, said limit value including: the upper limit of allowable voltage fluctuation V imax and the lower limit of permissible voltage fluctuation V imin ; If so, the power flow distribution is calculated n times, and the risk probability is calculated using the following formula; Where m is the number of times the voltage exceeds the limit in the n power flow distribution calculations, and n is the total number of power flow distribution samples.

8. The method according to claim 2, characterized in that, The graph data structure contains several nodes, and the risk probability is calculated based on the power flow distribution, including: Based on the power flow distribution, determine the terminal power flow value I of node i. i Does it exceed the predetermined upper limit value I of the branch power flow? imax ; If so, the power flow distribution is calculated n times, and the risk probability is calculated using the following formula; Where x is the number of times the branch power flow exceeds the limit in n power flow distribution calculations, and n is the total number of power flow distribution samples.

9. A probabilistic power flow evaluation device based on a side graph attention network, characterized in that, include: The sample value extraction module is used to extract sample values ​​based on uncertainties in the power system. The final sample determination module is used to orthogonally sort the sampled values ​​to obtain the final sample. The modeling module is used to model the final sample as a graph data structure; The power flow distribution calculation module is used to input the graph data structure into a power flow solver composed of stacked multi-layer EGAT graph filters to calculate the power flow distribution. The probabilistic power flow assessment module is used to perform probabilistic power flow assessment based on the power flow distribution. The modeling module performs the process of modeling the final sample as a graph data structure, including: Using the final sample, a power network diagram with N buses and M transmission channels is constructed, with the bus as the node and the transmission channel as the edge. The power network diagram is defined as G = {V, E}. node v i ∈V is the i-th node in the graph; Let i be the power flow initial value information node feature vector corresponding to the i-th node; Among them, U i Represents voltage amplitude, θ i Represents voltage phase angle, P Li Indicates active load, Q Li Represents reactive load, P Gi Indicates contribution, Q Gi This indicates that no effort was put in; All node features can be represented by a node feature matrix. Side e i ∈E represents the i-th branch in the graph; Let be the edge feature vector containing the line conductance and susceptance information of the i-th branch; All edge features are represented by the edge feature matrix. The power grid topology is constructed using the node adjacency matrix A. N×N Let A represent the element A in the i-th row and j-th column. ij satisfy:

Images