A power distribution network gridding method and device based on a graph attention network

Abstract

The application discloses a power distribution network gridding method and device based on a graph attention network, which comprises the following steps: performing grid division on a power distribution network according to a preset topological structure to obtain a grid area; constructing a grid feature vector of the grid area; inputting the grid feature vector into a preset graph attention network to perform feature prediction and obtain a predicted feature vector; judging whether the predicted feature vector meets a reliability index based on dynamic information of the grid area; if yes, solving a preset power demand model based on a preset power constraint condition; if no, taking the predicted feature vector as a current action state; performing reinforcement learning according to the current action state, taking an action value corresponding to a maximum probability as an updated feature vector, and returning to the judgment step; and optimizing a grid division mechanism based on a solution of the power demand model to obtain an optimized grid area. The application solves the technical problem that the existing power distribution network gridding method lacks pertinence and cannot meet the development demand of stable operation of a new energy power grid.

Description

Technical Field

[0001] The present application relates to the technical field of distribution network planning, and in particular to a distribution network gridding method and device based on a graph attention network. Background Art

[0002] With the scarcity of fossil fuels, renewable energy sources such as wind power and photovoltaics are being connected to the power grid on a large scale. The uncertainty in their output poses a serious threat to the reliability of distribution networks and impacts power quality. In power systems, the distribution network is the component closest to users and plays a crucial role in the distribution of electricity. Scientific optimization and planning of this network can ensure the economic and safe operation of the power system and improve power quality. Distribution network planning can be viewed as a multi-objective, dynamic, and uncertain nonlinear programming problem, which can be solved using mathematical optimization algorithms and artificial intelligence algorithms. Mathematical methods include the fuzzy analytic hierarchy process (AHP), the hierarchical analysis method (AHP), and the selection elimination method (SEM), as well as artificial intelligence algorithms such as genetic algorithms, ant colony algorithms, and evolutionary algorithms. However, with the integration of distributed energy resources and electrified transportation into the power system, traditional distribution network planning approaches and methods are no longer suitable for this ever-changing, large-scale system.

[0003] Grid planning involves dividing the planned area into several independent zones based on the distribution network's power supply area classification, land use characteristics, and development levels. This bottom-up approach aims to meet user needs and ensure the coordinated and optimized operation of multi-regional power systems. However, existing distribution network grid planning typically considers power flow and geographic information. This severely impacts the safe and stable operation of the distribution network when integrating emerging loads such as distributed energy and electrified transportation. As a result, the grid-based approach is no longer adaptable to the needs of next-generation energy development. Summary of the Invention

[0004] The present application provides a distribution network gridding method and device based on a graph attention network, which is used to solve the technical problem that the existing distribution network gridding method lacks specificity and cannot meet the development needs of stable operation of new energy power grids.

[0005] In view of this, the first aspect of the present application provides a distribution network gridding method based on a graph attention network, comprising:

[0006] Divide the distribution network into grids according to a preset topology structure to obtain multiple grid areas;

[0007] Constructing a grid feature vector for each of the grid areas, wherein the grid feature vector includes a topological proximity matrix and power flow information;

[0008] Inputting the grid feature vector into a preset graph attention network for feature prediction to obtain a predicted feature vector, wherein the preset graph attention network includes an attention mechanism;

[0009] Determining whether the predicted feature vector satisfies a preset reliability index based on the dynamic information of the grid area; if so, solving a preset power demand model corresponding to the distribution network based on preset power constraints; if not, using the predicted feature vector as the current action state;

[0010] Perform reinforcement learning based on the current action state, use the action value corresponding to the maximum probability as the updated feature vector, and return to the step of determining whether the predicted feature vector meets the preset reliability index based on the dynamic information of the grid area;

[0011] Based on the solution optimization grid division mechanism of the power demand model, an optimized grid area is obtained.

[0012] Preferably, inputting the grid feature vector into a preset graph attention network for feature prediction to obtain a predicted feature vector comprises:

[0013] The preset activation function is used to calculate the attention coefficient according to the correlation between nodes;

[0014] Based on the attention mechanism, feature prediction is performed according to the attention coefficient and the grid feature vector in the input preset graph attention network to obtain a predicted feature vector.

[0015] Preferably, the method of performing feature prediction based on the attention mechanism and the grid feature vector in the attention network of the input preset graph to obtain a predicted feature vector further includes:

[0016] Based on the multi-head attention mechanism, de-averaged integrated feature prediction is performed according to the attention coefficient and the grid feature vector in the input preset graph attention network to obtain a predicted feature vector.

[0017] Preferably, the step of performing reinforcement learning according to the current action state, taking the action value corresponding to the maximum probability as the updated feature vector, and returning the step of judging whether the predicted feature vector meets the preset reliability index based on the dynamic information of the grid area includes:

[0018] Perform state update calculations on the Q value matrix and the P value matrix based on the current action state and the initial reward value, and record the probabilities corresponding to different action values;

[0019] An action value corresponding to the maximum probability is obtained as an updated feature vector, and the step of determining whether the predicted feature vector meets a preset reliability index based on the dynamic information of the grid area is returned.

[0020] Preferably, the method further comprises: determining whether the predicted characteristic vector satisfies a preset reliability index based on the dynamic information of the grid area; if so, solving a preset power demand model corresponding to the distribution network based on preset power constraints; if not, taking the predicted characteristic vector as the current action state; and

[0021] Obtain relevant operating data of the distribution network;

[0022] A preset power demand model of the distribution network is constructed according to the relevant operation data, wherein the preset power demand model includes preset power constraints, and the preset power constraints include node voltage constraints, branch flow constraints, and node current constraints.

[0023] The second aspect of the present application provides a distribution network gridding device based on a graph attention network, comprising:

[0024] A grid division module is used to divide the distribution network into grids according to a preset topology structure to obtain multiple grid areas;

[0025] A vector construction module, configured to construct a grid feature vector for each of the grid areas, wherein the grid feature vector includes a topological proximity matrix and power flow information;

[0026] A feature prediction module, configured to input the grid feature vector into a preset graph attention network for feature prediction to obtain a predicted feature vector, wherein the preset graph attention network includes an attention mechanism;

[0027] a feature judgment module, configured to judge whether the predicted feature vector satisfies a preset reliability index based on the dynamic information of the grid area; if so, solving a preset power demand model corresponding to the distribution network based on preset power constraints; if not, using the predicted feature vector as the current action state;

[0028] A reinforcement learning module is used to perform reinforcement learning based on the current action state, use the action value corresponding to the maximum probability as an updated feature vector, and trigger the feature judgment module;

[0029] The grid optimization module is used to optimize the grid partitioning mechanism based on the solution of the power demand model to obtain an optimized grid area.

[0030] Preferably, the feature prediction module includes:

[0031] The coefficient calculation submodule is used to calculate the attention coefficient according to the correlation between nodes using a preset activation function;

[0032] The feature prediction submodule is used to perform feature prediction based on the attention mechanism according to the attention coefficient and the grid feature vector in the input preset graph attention network to obtain a predicted feature vector.

[0033] Preferably, it also includes:

[0034] The optimized feature prediction module is used to perform de-averaged integrated feature prediction based on the attention coefficient and the grid feature vector in the input preset graph attention network based on a multi-head attention mechanism to obtain a predicted feature vector.

[0035] Preferably, the reinforcement learning module includes:

[0036] A matrix calculation submodule is used to perform state update calculations on the Q value matrix and the P value matrix according to the current action state and the initial reward value, and record the probabilities corresponding to different action values;

[0037] The state screening submodule is used to obtain the action value corresponding to the maximum probability as an updated feature vector to trigger the feature judgment module.

[0038] Preferably, it also includes:

[0039] Data acquisition module, used to obtain relevant operating data of the distribution network;

[0040] A model construction module is used to construct a preset power demand model of the distribution network based on the relevant operating data, wherein the preset power demand model includes preset power constraints, and the preset power constraints include node voltage constraints, branch flow constraints and node current constraints.

[0041] It can be seen from the above technical solutions that the embodiments of the present application have the following advantages:

[0042] In the present application, a distribution network gridding method based on a graph attention network is provided, comprising: gridding the distribution network according to a preset topological structure to obtain a plurality of grid areas; constructing a grid feature vector for each grid area, the grid feature vector including a topological proximity matrix and flow information; inputting the grid feature vector into a preset graph attention network for feature prediction to obtain a predicted feature vector, the preset graph attention network including an attention mechanism; judging whether the predicted feature vector meets a preset reliability index based on dynamic information of the grid area, and if so, solving a preset power demand model corresponding to the distribution network based on preset power constraints, and if not, taking the predicted feature vector as the current action state; performing reinforcement learning based on the current action state, taking the action value corresponding to the maximum probability as the updated feature vector, and returning the step of judging whether the predicted feature vector meets the preset reliability index based on the dynamic information of the grid area; and obtaining an optimized grid area based on the solution of the power demand model.

[0043] The distribution network gridding method based on the graph attention network provided in this application uses a graph neural network with an attention mechanism to analyze and learn the power characteristics of each grid area in the distribution network to obtain prediction features. The graph attention network can consider the intrinsic correlation between nodes and emphasize the correlation between nodes, making the results more reliable and targeted; and the use of preset reliability indicators for screening and judgment makes the model solution process more accurate and more robust. Therefore, this application can solve the technical problem that the existing distribution network gridding method lacks specificity and cannot meet the development needs of the stable operation of new energy power grids. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 A schematic diagram of a flow chart of a distribution network gridding method based on a graph attention network provided in an embodiment of the present application;

[0045] Figure 2 A schematic diagram of the structure of a distribution network gridding device based on a graph attention network provided in an embodiment of the present application;

[0046] Figure 3 Schematic diagram of the gridding of a distribution network based on graph attention network and reinforcement learning provided in an embodiment of the present application. DETAILED DESCRIPTION

[0047] In order to help those skilled in the art better understand the present invention, the following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of this application.

[0048] For easier understanding, see Figure 1 , the present application provides a first embodiment of a distribution network gridding method based on a graph attention network, comprising:

[0049] Step 101: Divide the distribution network into grids according to a preset topology structure to obtain multiple grid areas.

[0050] The pre-configured topology is constructed based on the distribution network. The specific process can be referenced in existing technologies and will not be detailed here. Grid division can be based on factors such as the nature of power flow or geographic characteristics. The specific division mechanism needs to be tailored to the actual grid supply area. Reasonable regional gridding can improve the coordination performance of the power system and ensure its stable operation.

[0051] Step 102: construct a grid feature vector for each grid area, where the grid feature vector includes a topological proximity matrix and power flow information.

[0052] The information elements in the constructed grid feature vector can be expressed as G(A,X), where A is the proximity matrix of the topological relationship between the distribution network nodes, and X is the power flow information, specifically including the three-phase voltage and current amplitudes, reactive power, and active power. The proximity matrix A is expressed as:

[0053]

[0054] Among them, n is the number of sampling points and m is the number of nodes.

[0055] The three-phase voltage and current amplitudes, reactive power, and active power are expressed as: X = [V1, V2, V3, …, I1, I2, I3, …, P1, P2, P3, …, Q1, Q2, Q3, …] T .

[0056] Step 103: Input the grid feature vector into a preset graph attention network for feature prediction to obtain a predicted feature vector. The preset graph attention network includes an attention mechanism.

[0057] Furthermore, step 103 includes:

[0058] The preset activation function is used to calculate the attention coefficient according to the correlation between nodes;

[0059] Based on the attention mechanism, feature prediction is performed according to the attention coefficient and the grid feature vector in the input preset graph attention network to obtain the predicted feature vector.

[0060] Since most distribution networks are tree-like and radial, the internal component information can be ignored and the distribution network can be abstracted as a graph consisting of vertices and edges. Each node represents an intelligent agent, and the edges between nodes represent the interaction relationship between two intelligent agents.

[0061] The attention mechanism is a weighted summation of all node features, where the weight is the correlation between the central node and the neighboring node features. Graph attention networks can aggregate the proximity matrix through the attention mechanism to achieve adaptive weight allocation for adjacent regions, greatly improving the expressive power of graph neural network models. The core of the attention mechanism is to assign weights to given information, as shown below:

[0062] A=(X,Y)=∑<X,Y> ·Y

[0063] Where Y is the characteristic phasor of all neighboring nodes, X is the three-phase voltage and current amplitude, reactive power, and active power information, and A is the proximity matrix of the topological relationship between distribution network nodes. The correlation of all neighboring nodes is calculated, and then normalized using the ReLu activation function to calculate the attention coefficient. The specific attention coefficient calculation can be expressed as:

[0064]

[0065] Among them, α(·) is the function for calculating the correlation between two nodes, W is the weight matrix, and v k is the kth node, N(v k ) is the total number of nodes, X i 、X j and X k are all different feature vectors in the preset graph attention network. The input network feature vector can be expressed as X = {X1, X2, ..., X i},X i ∈R F , the feature vector of the output network can be expressed as X′={X′1,X′2,...,X′ i},X′ i ∈R F′ .

[0066] Based on the weighted idea of ​​attention mechanism, node v i The predicted feature vector can be expressed as:

[0067]

[0068] Among them, σ(·) is the activation function, β ij is the weight coefficient. Since the graph attention network can extract the internal connections between nodes, the predicted feature vector is targeted and has strong expressive power.

[0069] Furthermore, based on the attention mechanism, feature prediction is performed according to the attention coefficient and the grid feature vector in the input preset graph attention network to obtain a predicted feature vector, which also includes:

[0070] Based on the multi-head attention mechanism, the feature prediction of de-averaged integration is performed according to the attention coefficient and the grid feature vector in the input preset graph attention network to obtain the predicted feature vector.

[0071] In addition to the basic attention mechanism, in order to further improve the expressiveness of the attention layer, a multi-head attention mechanism can be added to perform feature prediction using an average integration approach:

[0072]

[0073] Where H is the number of attention heads, H ≥ 2.

[0074] Both methods of obtaining the prediction feature vector are acceptable and can be selected based on the specific situation.

[0075] Step 104: determine whether the predicted feature vector meets the preset reliability index based on the dynamic information of the grid area. If so, solve the preset power demand model corresponding to the distribution network based on the preset power constraint condition. If not, use the predicted feature vector as the current action state.

[0076] The average power outage frequency f and the average power outage duration t s And the average power supply availability rate η is used as the preset reliability index, and the index calculation formula is as follows:

[0077]

[0078]

[0079]

[0080] Among them, N i is the number of users at node i, λ i is the failure rate of node i, T i is the annual average power outage time of node i.

[0081] The model can be solved only when it is determined that the predicted feature vector meets the preset reliability index; otherwise, the subsequent feature analysis operation process needs to be entered.

[0082] Furthermore, before step 104, the following steps are also included:

[0083] Obtain relevant operating data of the distribution network;

[0084] A preset power demand model of the distribution network is constructed based on relevant operating data. The preset power demand model includes preset power constraints, which include node voltage constraints, branch flow constraints, and node current constraints.

[0085] The preset power demand model is a power demand model constructed based on the principle of urban-rural integration and the actual power demand of each power supply unit.

[0086] L n =L n-1 ×(1+g)+L b ×bL D1 +L D2

[0087] Among them, L n is the highest load in the nth year, L n-1is the peak load in year n-1, g is the natural growth rate of power load, L b is the reported load, b is the practical coefficient, L D1 With L D2 They are respectively the incoming load and the outgoing load.

[0088] The preset power constraints include node voltage constraints, branch flow constraints, and node current constraints, specifically:

[0089]

[0090] P L ≤P Gmax

[0091] I i ≤I imax

[0092] in, V i and V i are the lower and upper limits of the node voltage, P L is the total active load of the substation, P Gmax is the maximum active load allowed by the substation, I i is the load current flowing through node i, I imax is the maximum allowable current of node i, and Φ is the set of distribution network nodes.

[0093] Step 105 : Perform reinforcement learning based on the current action state, use the action value corresponding to the maximum probability as the updated feature vector, and return to the step of judging whether the predicted feature vector meets the preset reliability index based on the dynamic information of the grid area.

[0094] Furthermore, step 105 includes:

[0095] Perform state update calculations on the Q-value matrix and the P-value matrix based on the current action state and the initial reward value, and record the probabilities corresponding to different action values;

[0096] The action value corresponding to the maximum probability is obtained as the updated feature vector, and the process returns to step 104 .

[0097] Reinforcement learning is a model-free and online strategy-updating technology. The agent obtains the initial or current state s and reward value r from the environment; updates the Q-value matrix and P-value matrix under the current state, and finally selects the action value with high probability to output the update formula as follows:

[0098]

[0099]

[0100] Where s is the current state, s′ is the state at the next moment, α, β, and γ are the learning rate, probability distribution factor, and discount factor for reinforcement learning, respectively. Q(s,a) and P(s,a) represent the Q-value matrix and P-value matrix, respectively. R(s,s′,a) represents the immediate reward for the agent transitioning from the current state s to the next state s′ by taking action a, expressed as a probability. Based on the current action state, the reward values ​​for executing different action values ​​are obtained. The action value corresponding to the maximum value is selected as the updated feature vector. The preset reliability indicator judgment operation is then continued. If the conditions are met, the model is solved. Otherwise, reinforcement learning continues to refine the feature vector.

[0101] Step 106: Optimize the grid division mechanism based on the solution of the power demand model to obtain an optimized grid area.

[0102] See also Figure 3 A schematic diagram of a gridded distribution network based on a graph attention network and reinforcement learning is presented. The solution to the power demand model is derived through an accurate judgment mechanism, making it more targeted and reflecting the regional power consumption and supply characteristics of the distribution network. Grid planning based on this solution is more accurate and reliable, achieving the optimization goal of grid planning.

[0103] The distribution network gridding method based on the graph attention network provided in the embodiment of the present application uses a graph neural network that introduces an attention mechanism to analyze and learn the power characteristics of each grid area in the distribution network to obtain prediction features. The graph attention network can consider the intrinsic associations between nodes and emphasize the correlation between nodes, making the results more reliable and targeted; and the use of preset reliability indicators for screening and judgment makes the model solution process more accurate and more robust. Therefore, the embodiment of the present application can solve the technical problem that the existing distribution network gridding method lacks specificity and cannot meet the development needs of stable operation of new energy power grids.

[0104] For easier understanding, see Figure 2 , the present application also provides an embodiment of a distribution network gridding device based on a graph attention network, comprising:

[0105] A grid division module 201 is used to divide the distribution network into grids according to a preset topology structure to obtain multiple grid areas;

[0106] A vector construction module 202 is used to construct a grid feature vector for each grid area, where the grid feature vector includes a topological proximity matrix and power flow information;

[0107] A feature prediction module 203 is configured to input the grid feature vector into a preset graph attention network for feature prediction to obtain a predicted feature vector, wherein the preset graph attention network includes an attention mechanism;

[0108] A feature determination module 204 is configured to determine whether the predicted feature vector satisfies a preset reliability index based on the dynamic information of the grid area. If so, a preset power demand model corresponding to the distribution network is solved based on preset power constraints. If not, the predicted feature vector is used as the current action state.

[0109] Reinforcement learning module 205, used to perform reinforcement learning based on the current action state, use the action value corresponding to the maximum probability as the updated feature vector, and trigger the feature judgment module;

[0110] The grid optimization module 206 is used to optimize the grid division mechanism based on the solution of the power demand model to obtain the optimized grid area.

[0111] Furthermore, the feature prediction module 203 includes:

[0112] The coefficient calculation submodule 2031 is used to calculate the attention coefficient according to the correlation between nodes using a preset activation function;

[0113] The feature prediction submodule 2032 is used to perform feature prediction based on the attention mechanism, according to the attention coefficient and the grid feature vector in the input preset graph attention network, to obtain a predicted feature vector.

[0114] Furthermore, it also includes:

[0115] The optimized feature prediction module 2033 is used to perform de-averaged integrated feature prediction based on the attention coefficient and the grid feature vector in the input preset graph attention network based on the multi-head attention mechanism to obtain a predicted feature vector.

[0116] Furthermore, the reinforcement learning module 205 includes:

[0117] The matrix calculation submodule 2051 is used to perform state update calculations on the Q-value matrix and the P-value matrix according to the current action state and the initial reward value, and record the probabilities corresponding to different action values;

[0118] The state screening submodule 2052 is used to obtain the action value corresponding to the maximum probability as an updated feature vector to trigger the feature judgment module.

[0119] Furthermore, it also includes:

[0120] Data acquisition module 207, used to obtain relevant operation data of the distribution network;

[0121] The model building module 208 is used to build a preset power demand model of the distribution network based on relevant operating data. The preset power demand model includes preset power constraints, which include node voltage constraints, branch flow constraints, and node current constraints.

[0122] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0123] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0124] In addition, the functional units in the various embodiments of the present application may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0125] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a number of instructions for executing all or part of the steps of the method described in each embodiment of the present application through a computer device (which can be a personal computer, server, or network device, etc.). The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (English full name: Read-Only Memory, English abbreviation: ROM), random access memory (English full name: Random Access Memory, English abbreviation: RAM), disk or optical disk and other media that can store program code.

[0126] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A distribution network gridding method based on graph attention network, characterized in that: include: Divide the distribution network into grids according to a preset topology structure to obtain multiple grid areas; Constructing a grid feature vector for each of the grid areas, wherein the grid feature vector includes a topological proximity matrix and power flow information; Inputting the grid feature vector into a preset graph attention network for feature prediction to obtain a predicted feature vector, wherein the preset graph attention network includes an attention mechanism; Determining whether the predicted feature vector satisfies a preset reliability index based on the dynamic information of the grid area; if so, solving a preset power demand model corresponding to the distribution network based on preset power constraints; if not, using the predicted feature vector as the current action state; The average power outage frequency f and the average power outage duration t s and the average power supply availability rate η as the preset reliability index, and the index calculation formula is as follows: Among them, N i is the number of users at node i, λ i is the failure rate of node i, T i is the annual average power outage time of node i; Perform reinforcement learning based on the current action state, use the action value corresponding to the maximum probability as the updated feature vector, and return to the step of determining whether the predicted feature vector meets the preset reliability index based on the dynamic information of the grid area; Based on the solution optimization grid division mechanism of the power demand model, an optimized grid area is obtained.

2. The distribution network gridding method based on graph attention network according to claim 1 is characterized in that: The step of inputting the grid feature vector into a preset graph attention network for feature prediction to obtain a predicted feature vector comprises: The preset activation function is used to calculate the attention coefficient according to the correlation between nodes; Based on the attention mechanism, feature prediction is performed according to the attention coefficient and the grid feature vector in the input preset graph attention network to obtain a predicted feature vector.

3. The distribution network gridding method based on graph attention network according to claim 2 is characterized in that: The method further comprises: performing feature prediction based on the attention coefficient and the grid feature vector in the input preset graph attention network to obtain a predicted feature vector based on the attention mechanism; Based on the multi-head attention mechanism, de-averaged integrated feature prediction is performed according to the attention coefficient and the grid feature vector in the input preset graph attention network to obtain a predicted feature vector.

4. The distribution network gridding method based on graph attention network according to claim 1 is characterized in that: The step of performing reinforcement learning according to the current action state, taking the action value corresponding to the maximum probability as the updated feature vector, and returning the step of determining whether the predicted feature vector meets the preset reliability index based on the dynamic information of the grid area includes: Perform state update calculations on the Q value matrix and the P value matrix based on the current action state and the initial reward value, and record the probabilities corresponding to different action values; An action value corresponding to the maximum probability is obtained as an updated feature vector, and the step of determining whether the predicted feature vector meets a preset reliability index based on the dynamic information of the grid area is returned.

5. The distribution network gridding method based on graph attention network according to claim 1 is characterized in that: The method further comprises: determining whether the predicted feature vector satisfies a preset reliability index based on the dynamic information of the grid area; if so, solving a preset power demand model corresponding to the distribution network based on a preset power constraint condition; if not, using the predicted feature vector as the current action state; and Obtain relevant operating data of the distribution network; A preset power demand model of the distribution network is constructed according to the relevant operation data, wherein the preset power demand model includes preset power constraints, and the preset power constraints include node voltage constraints, branch flow constraints, and node current constraints.

6. A distribution network gridding device based on graph attention network, characterized in that: include: A grid division module is used to divide the distribution network into grids according to a preset topology structure to obtain multiple grid areas; A vector construction module, configured to construct a grid feature vector for each of the grid areas, wherein the grid feature vector includes a topological proximity matrix and power flow information; A feature prediction module, configured to input the grid feature vector into a preset graph attention network for feature prediction to obtain a predicted feature vector, wherein the preset graph attention network includes an attention mechanism; a feature judgment module, configured to judge whether the predicted feature vector satisfies a preset reliability index based on the dynamic information of the grid area; if so, solving a preset power demand model corresponding to the distribution network based on preset power constraints; if not, using the predicted feature vector as the current action state; The average power outage frequency f and the average power outage duration t s and the average power supply availability rate η as the preset reliability index, and the index calculation formula is as follows: Among them, N i is the number of users at node i, λ i is the failure rate of node i, T i is the annual average power outage time of node i; A reinforcement learning module is used to perform reinforcement learning based on the current action state, use the action value corresponding to the maximum probability as an updated feature vector, and trigger the feature judgment module; The grid optimization module is used to optimize the grid partitioning mechanism based on the solution of the power demand model to obtain an optimized grid area.

7. The distribution network gridding device based on graph attention network according to claim 6, characterized in that: The feature prediction module includes: The coefficient calculation submodule is used to calculate the attention coefficient according to the correlation between nodes using a preset activation function; The feature prediction submodule is used to perform feature prediction based on the attention mechanism according to the attention coefficient and the grid feature vector in the input preset graph attention network to obtain a predicted feature vector.

8. The distribution network gridding device based on graph attention network according to claim 7 is characterized in that: Also includes: The optimized feature prediction module is used to perform de-averaged integrated feature prediction based on the attention coefficient and the grid feature vector in the input preset graph attention network based on a multi-head attention mechanism to obtain a predicted feature vector.

9. The distribution network gridding device based on graph attention network according to claim 6, characterized in that: The reinforcement learning module includes: A matrix calculation submodule is used to perform state update calculations on the Q value matrix and the P value matrix according to the current action state and the initial reward value, and record the probabilities corresponding to different action values; The state screening submodule is used to obtain the action value corresponding to the maximum probability as an updated feature vector to trigger the feature judgment module.

10. The distribution network gridding device based on graph attention network according to claim 6, characterized in that: Also includes: Data acquisition module, used to obtain relevant operating data of the distribution network; A model construction module is used to construct a preset power demand model of the distribution network based on the relevant operating data, wherein the preset power demand model includes preset power constraints, and the preset power constraints include node voltage constraints, branch flow constraints and node current constraints.

Images