Optimal power flow solving method and device based on graph neural network and extreme learning machine

Abstract

The invention discloses an optimal power flow solving method and device based on a graph neural network and an extreme learning machine, and relates to the field of power grid power flow calculation, and the method comprises the steps: obtaining the node topology of a to-be-solved power system and the historical operation data of each power node, and constructing a system topological graph according to the node topology and the historical operation data; extracting features of a system topological graph based on a multilayer graph convolutional neural network to obtain first node features; taking the minimum total output of the system as a target, screening first node features based on regularization, and fusing the first node features to obtain second node features; performing multiple times of iterative training on the initial extreme learning machine model according to the second node features, optimizing model parameters based on mean square error loss during iteration, completing training until the model parameters meet preset values, and obtaining a first extreme learning machine model; and obtaining real-time operation data of the to-be-solved power system, and obtaining an optimal power flow solving result of the to-be-solved power system according to the real-time operation data and the first extreme learning machine model.

Description

Technical Field

[0001] The present application relates to the field of power grid flow calculation, and in particular to a method and device for solving optimal power flow based on graph neural network and extreme learning machine. Background Art

[0002] Solving the Optimal Power Flow (OPF) problem is a key task in power system analysis. Its solution is conducive to optimizing the operating state of the power system, minimizing the operating cost, and improving the economy and reliability of the power system under the premise of meeting the safe operation constraints. To this end, it is necessary to continuously improve the accuracy and efficiency of solving the optimal power flow problem.

[0003] The existing traditional OPF solution methods mainly rely on mathematical optimization techniques based on physical models, which usually use linear programming, nonlinear programming or dynamic programming methods. However, with the continuous expansion of the scale of the power system and the successive access of grid-connected nodes, the grid topology has become increasingly complex. The computational complexity of the traditional OPF solution method has increased exponentially, and the computational efficiency of the traditional OPF solution method has been difficult to adapt to the current grid system topology. At the same time, the continuous expansion of the scale of the power system also requires the efficiency of the OPF solution to be continuously improved, and it is necessary to maintain the reliability of grid regulation based on the OPF solution results. Therefore, how to solve the optimal power flow of the power system more efficiently and accurately is still a technical problem that needs to be solved urgently in the existing technology. Summary of the invention

[0004] The present application provides an optimal power flow solution method and device based on graph neural network and extreme learning machine to solve the technical problem that traditional optimal power flow solution methods are not efficient and accurate.

[0005] According to a first aspect of the implementation of the present application, a method for solving an optimal power flow based on a graph neural network and an extreme learning machine is provided, comprising:

[0006] Obtaining the node topology of the power system to be solved and the historical operation data of each power node, and constructing a system topology diagram according to the node topology and the historical operation data; wherein the historical operation data includes node operation characteristics and line electrical characteristics;

[0007] Based on a preset multi-layer graph convolutional neural network, feature extraction is performed on the system topology graph layer by layer to obtain a first node feature;

[0008] Taking the minimum total output of the system as the screening target, each node feature in the first node feature is screened and merged based on regularization to obtain a second node feature;

[0009] According to the second node feature, a preset initial extreme learning machine model is iterated for multiple times, and model parameters of the initial extreme learning machine model are optimized based on mean square error loss during iteration, until the model parameters meet preset values, the training is completed, and a first extreme learning machine model is obtained;

[0010] Real-time operation data of the power system to be solved is obtained, and based on the real-time operation data and the first extreme learning machine model, an optimal power flow solution result of the power system to be solved is obtained.

[0011] The present application first obtains the node topology and the historical operation data of each power node, and constructs a system topology map, and then extracts features layer by layer based on a multi-layer graph convolutional neural network to obtain a first node feature, which can gradually enhance the characterization ability of the extracted features for the system topology map layer by layer, and improve the matching degree between the features and the actual topology of the power system, and then takes the minimum total system output as the screening target, and based on regularization, screens the node features in the first node feature and fuses them to obtain the second node feature, which can screen out unnecessary node feature types, and by reducing the node feature types, reduce the computational complexity, and improve the computational efficiency and solution accuracy of the optimal power flow solution, and then iterates the initial extreme learning machine model for multiple times according to the second node feature, and optimizes the model parameters based on the mean square error loss to obtain the first extreme learning machine model, which can improve the accuracy of the first extreme learning machine model, thereby improving the accuracy of the optimal power flow solution when the optimal power flow solution result is obtained according to the real-time operation data.

[0012] In certain embodiments of the present application, constructing a system topology diagram according to the node topology and the historical operation data specifically includes:

[0013] According to the node topology, a system topology graph is constructed with the node operation characteristics as nodes of the system topology graph and the line electrical characteristics as edges of the system topology graph.

[0014] This application is based on node topology, with node operation characteristics as nodes and line electrical characteristics as edges, which can make the constructed system topology diagram closer to the actual power system.

[0015] In certain embodiments of the present application, the system topology graph is extracted layer by layer based on a preset multi-layer graph convolutional neural network to obtain the first node feature, specifically including:

[0016] Based on the multi-layer graph convolutional neural network, feature extraction is performed on the system topology map layer by layer to obtain first node features; wherein, during feature extraction, each layer of the graph convolutional neural network updates the node features of the previous layer based on graph convolution to obtain the node features of the current layer; each current layer node feature is connected with the initial features of the system topology map before being output to the next layer to obtain the node features of the updated layer.

[0017] The present application performs feature extraction layer by layer based on a multi-layer graph convolutional neural network to obtain the first node feature, which can gradually enhance the characterization ability of the extracted features for the system topology map layer by layer, and improve the matching degree between the features and the actual topology of the power system.

[0018] In certain embodiments of the present application, each layer of the graph convolutional neural network updates the node features of the previous layer based on the graph convolution to obtain the node features of the current layer during feature extraction, specifically including:

[0019]

[0020] Among them, H (l) , H (l-1) are the node features of the lth layer and the node features of the (l-1)th layer, respectively, W (l) is the weight matrix of the lth layer, σ(·) is the activation function, A is the adjacency matrix corresponding to the system topology graph, D is the degree matrix composed of the degrees of each node in the system topology graph, and D ii =Σ j A ij , D ii is the element value of the i-th row and i-th column of the degree matrix, A ij is the element value in the i-th row and j-th column of the adjacency matrix.

[0021] In certain embodiments of the present application, the minimum total system output is used as the screening target, and each node feature in the first node feature is screened and merged based on regularization to obtain the second node feature, which specifically includes:

[0022] Taking the minimum total output of the system as the screening target, each node feature in the first node feature is screened based on regularization to obtain an intermediate node feature;

[0023] Based on the preset node weights, the node features of the intermediate node features are fused to obtain the third node feature.

[0024] The present application first takes the minimum total system output as the screening target, and obtains the intermediate node features based on regularization screening of each node feature in the first node feature, which can screen out unnecessary node feature types, reduce calculation complexity, and improve the calculation efficiency and solution accuracy of the optimal power flow solution. Then, based on the preset node weights, the node features of the intermediate node features are fused to obtain the third node feature. Through the fusion of node features, the global features can be captured, thereby more accurately characterizing the actual topology of the power system.

[0025] In certain embodiments of the present application, the screening target is to minimize the total system output, specifically:

[0026]

[0027] Among them, Y is the system target power flow, X is the first node feature matrix, β is the feature weight of each node feature of the first node feature, λ is the regularization parameter, ||·||2, ||·||1 are the 2-norm and 1-norm, respectively.

[0028] In certain embodiments of the present application, the method of performing multiple iterative training on a preset initial extreme learning machine model according to the second node feature, optimizing the model parameters of the initial extreme learning machine model based on the mean square error loss during iteration, and completing the training until the model parameters meet the preset values ​​to obtain the first extreme learning machine model, specifically includes:

[0029] According to the second node feature, the preset initial extreme learning machine model is iteratively trained for multiple times. During the iteration, the output layer weights of the output layer are updated according to the mean square error loss between the current output of the output layer of the initial extreme learning machine model and the actual output, until the training is completed when the mean square error loss meets the preset value, thereby obtaining the first extreme learning machine model; wherein the current output of the output layer is calculated based on the hidden layer.

[0030] The present application performs multiple iterative training on the initial extreme learning machine model according to the second node characteristics, optimizes the model parameters based on the mean square error loss to obtain the first extreme learning machine model, and can improve the accuracy of the first extreme learning machine model.

[0031] In some implementations of the present application, the mean square error loss is specifically:

[0032] min||Y (pred) Y (true) || 2 ;

[0033] Among them, Y (pred) is the current output of the output layer, Y (true) is the actual output, and ||·|| is the matrix norm.

[0034] In some embodiments of the present application, the current output of the output layer is calculated based on the hidden layer, specifically including:

[0035] The output of the hidden layer is specifically:

[0036] H hidden =σ(H fued W hiden ));

[0037] Among them, H hidden is the output of the hidden layer, H fused is the third node feature matrix, W hiddenis the weight of the hidden layer, σ(·) is the activation function;

[0038] The current output of the output layer is specifically:

[0039] Y (pred) =H hidden W output ;

[0040] Among them, Y (pred) is the current output of the output layer, W output is the weight of the output layer.

[0041] According to a second aspect of the implementation of the present application, there is provided an optimal power flow solving device based on a graph neural network and an extreme learning machine, comprising a data acquisition construction module, a node feature extraction module, a node feature screening module, a model iteration training module and an optimal power flow solving module;

[0042] The data acquisition construction module is used to obtain the node topology of the power system to be solved and the historical operation data of each power node, and construct a system topology diagram according to the node topology and the historical operation data; wherein the historical operation data includes node operation characteristics and line electrical characteristics;

[0043] The node feature extraction module is used to extract features from the system topology graph layer by layer based on a preset multi-layer graph convolutional neural network to obtain a first node feature;

[0044] The node feature screening module is used to screen and fuse each node feature in the first node feature based on regularization to obtain a second node feature with the minimum total system output as the screening target;

[0045] The model iteration training module is used to perform multiple iteration training on the preset initial extreme learning machine model according to the second node feature, and optimize the model parameters of the initial extreme learning machine model based on the mean square error loss during iteration, until the training is completed when the model parameters meet the preset values, so as to obtain the first extreme learning machine model;

[0046] The optimal power flow solving module is used to obtain real-time operating data of the power system to be solved, and obtain the optimal power flow solving result of the power system to be solved based on the real-time operating data and the first extreme learning machine model.

[0047] The present application first obtains the node topology and the historical operation data of each power node, and constructs a system topology map, and then extracts features layer by layer based on a multi-layer graph convolutional neural network to obtain a first node feature, which can gradually enhance the characterization ability of the extracted features for the system topology map layer by layer, and improve the matching degree between the features and the actual topology of the power system, and then takes the minimum total system output as the screening target, and based on regularization, screens the node features in the first node feature and fuses them to obtain the second node feature, which can screen out unnecessary node feature types, and by reducing the node feature types, reduce the computational complexity, and improve the computational efficiency and solution accuracy of the optimal power flow solution, and then iterates the initial extreme learning machine model for multiple times according to the second node feature, and optimizes the model parameters based on the mean square error loss to obtain the first extreme learning machine model, which can improve the accuracy of the first extreme learning machine model, thereby improving the accuracy of the optimal power flow solution when the optimal power flow solution result is obtained according to the real-time operation data. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 : A schematic diagram of a flow chart of an optimal power flow solution method based on a graph neural network and an extreme learning machine shown in certain embodiments of the present application;

[0049] Figure 2 : A module structure diagram of an optimal power flow solving device based on a graph neural network and an extreme learning machine shown in certain embodiments of the present application. DETAILED DESCRIPTION

[0050] The embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals represent the same or similar elements or elements having the same or similar functions from beginning to end. The embodiments described below in conjunction with the accompanying drawings are exemplary and are only used to explain some embodiments of the present application, and should not be construed as limitations on the embodiments of the present application. Based on the embodiments shown in the present application, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present application.

[0051] In the description of this application, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include at least one of the features. In the description of this application, unless otherwise clearly and specifically defined, "multiple" and "several" mean two or more.

[0052] Traditional OPF solving methods mainly rely on mathematical optimization techniques based on physical models. As the scale of power systems continues to expand and the grid topology becomes increasingly complex, the computational complexity of traditional OPF solving methods increases exponentially, and the computational efficiency is difficult to adapt to the current grid system topology. At the same time, the continuous expansion of the scale of power systems also requires the efficiency of OPF solving to continue to improve, and it is necessary to maintain the reliability of grid regulation based on the OPF solution results. Therefore, how to solve the optimal power flow of power systems more efficiently and accurately is still a technical problem that needs to be solved urgently in the existing technology.

[0053] Based on the above technical background, please refer to Figure 1 The embodiment of the present application provides an optimal power flow solution method based on a graph neural network and an extreme learning machine, including steps S101 to S105, each of which is as follows:

[0054] Step S101: Obtain the node topology of the power system to be solved and the historical operation data of each power node, and construct a system topology diagram based on the node topology and the historical operation data; wherein the historical operation data includes node operation characteristics and line electrical characteristics.

[0055] In certain embodiments of the present application, the node operation characteristics include generator characteristics, transformer characteristics and transformer characteristics. Specifically, the generator characteristics include the rated power, reactive power, voltage level and upper and lower limits of the generator's power generation capacity.

[0056] In certain embodiments of the present application, the line electrical characteristics include line impedance and line capacity limitation.

[0057] In certain embodiments of the present application, constructing a system topology diagram according to the node topology and the historical operation data specifically includes:

[0058] According to the node topology, a system topology graph is constructed with the node operation characteristics as nodes of the system topology graph and the line electrical characteristics as edges of the system topology graph.

[0059] This application is based on node topology, with node operation characteristics as nodes and line electrical characteristics as edges, which can make the constructed system topology diagram closer to the actual power system.

[0060] Step S102: Based on a preset multi-layer graph convolutional neural network, feature extraction is performed on the system topology graph layer by layer to obtain a first node feature.

[0061] In certain embodiments of the present application, the system topology graph is extracted layer by layer based on a preset multi-layer graph convolutional neural network to obtain the first node feature, specifically including:

[0062] Based on the multi-layer graph convolutional neural network, feature extraction is performed on the system topology map layer by layer to obtain first node features; wherein, during feature extraction, each layer of the graph convolutional neural network updates the node features of the previous layer based on graph convolution to obtain the node features of the current layer; each current layer node feature is connected with the initial features of the system topology map before being output to the next layer to obtain the node features of the updated layer.

[0063] In certain embodiments of the present application, a preferred embodiment of the multi-layer graph convolutional neural network is a graph neural network (GNN).

[0064] In certain embodiments of the present application, each layer of the graph convolutional neural network updates the node features of the previous layer based on the graph convolution to obtain the node features of the current layer during feature extraction, specifically including:

[0065]

[0066] Among them, H (l) , H (l-1) are the node features of the lth layer and the node features of the (l-1)th layer, respectively, W (l) is the weight matrix of the lth layer, σ(·) is the activation function, A is the adjacency matrix corresponding to the system topology graph, D is the degree matrix composed of the degrees of each node in the system topology graph, and D ii =Σ j A ij , D ii is the element value of the i-th row and i-th column of the degree matrix, A ij is the element value in the i-th row and j-th column of the adjacency matrix.

[0067] In certain embodiments of the present application, a preferred implementation of the activation function σ(·) is a ReLU function.

[0068] In certain embodiments of the present application, when each layer of the graph convolutional neural network updates the previous layer node features based on graph convolution to obtain the current layer node features during feature extraction, each power node in the system topology graph aggregates its own current layer node features with the current layer node features of all adjacent nodes into its own current layer node features.

[0069] In certain embodiments of the present application, the node features of each current layer are connected with the initial features of the system topology map before being output to the next layer to obtain the node features of the update layer, which specifically includes:

[0070] H final =concat(H (l) ,H (O) );

[0071] Among them, H final is the update layer node feature, H(l) , H (0) are the current layer node features and the initial features respectively, and concat(·) represents the feature connection operation.

[0072] The present application performs feature extraction layer by layer based on a multi-layer graph convolutional neural network to obtain the first node feature, which can gradually enhance the characterization ability of the extracted features for the system topology map layer by layer, and improve the matching degree between the features and the actual topology of the power system.

[0073] Step S103: Taking the minimum total system output as the screening target, each node feature in the first node feature is screened and integrated based on regularization to obtain a second node feature.

[0074] In certain embodiments of the present application, the minimum total system output is used as the screening target, and each node feature in the first node feature is screened and merged based on regularization to obtain the second node feature, which specifically includes:

[0075] Taking the minimum total output of the system as the screening target, each node feature in the first node feature is screened based on regularization to obtain an intermediate node feature;

[0076] Based on the preset node weights, the node features of the intermediate node features are fused to obtain the third node feature.

[0077] In certain embodiments of the present application, the screening target is to minimize the total system output, specifically:

[0078]

[0079] Among them, Y is the system target power flow, X is the first node feature matrix, β is the feature weight of each node feature of the first node feature, λ is the regularization parameter, ||·||2, ||·||1 are the 2-norm and 1-norm, respectively.

[0080] The reason why the minimum total system output is considered as the screening target is that by penalizing unnecessary features, it can ensure that the subsequent model focuses on the most relevant node features when solving the optimal power flow, thereby reducing the computational complexity and improving the computational efficiency and solution accuracy of the optimal power flow solution.

[0081] In certain embodiments of the present application, the third node feature is obtained by fusing the node features of the intermediate node features based on the preset node weight, specifically:

[0082] H fused =∑ i w i ·H i ;

[0083] Among them, H fusedis the third node feature, w i is the weight of the i-th node feature, H i is the feature of the i-th node.

[0084] The fusion of node features is considered because it is possible to capture global features while maintaining the representation capabilities of each necessary node feature, thereby enhancing the representation capability to improve the actual matching degree with the node topology of the power system.

[0085] The present application first takes the minimum total system output as the screening target, and obtains the intermediate node features based on regularization screening of each node feature in the first node feature, which can screen out unnecessary node feature types, reduce calculation complexity, and improve the calculation efficiency and solution accuracy of the optimal power flow solution. Then, based on the preset node weights, the node features of the intermediate node features are fused to obtain the third node feature. Through the fusion of node features, the global features can be captured, thereby more accurately characterizing the actual topology of the power system.

[0086] Step S104: According to the second node feature, the preset initial extreme learning machine model is iterated for multiple trainings, and the model parameters of the initial extreme learning machine model are optimized based on the mean square error loss during the iterations, until the training is completed when the model parameters meet the preset values, so as to obtain the first extreme learning machine model.

[0087] In certain embodiments of the present application, the method of performing multiple iterative training on a preset initial extreme learning machine model according to the second node feature, optimizing the model parameters of the initial extreme learning machine model based on the mean square error loss during iteration, and completing the training until the model parameters meet the preset values ​​to obtain the first extreme learning machine model, specifically includes:

[0088] According to the second node feature, the preset initial extreme learning machine model is iteratively trained for multiple times. During the iteration, the output layer weights of the output layer are updated according to the mean square error loss between the current output of the output layer of the initial extreme learning machine model and the actual output, until the training is completed when the mean square error loss meets the preset value, thereby obtaining the first extreme learning machine model; wherein the current output of the output layer is calculated based on the hidden layer.

[0089] It should be understood that the preferred embodiment of the extreme learning machine (ELM) described in the present application is a single hidden layer feedforward network, and the ELM structure includes an input layer, a hidden layer and an output layer.

[0090] In some implementations of the present application, the mean square error loss is specifically:

[0091] min||Y (pred) Y (true) || 2 ;

[0092] Among them, Y (pred) is the current output of the output layer, Y (true) is the actual output, and ||·|| is the matrix norm.

[0093] In some embodiments of the present application, the current output of the output layer is calculated based on the hidden layer, specifically including:

[0094] The output of the hidden layer is specifically:

[0095] H hidden =σ(H fused W hidden ));

[0096] Among them, H hidden is the output of the hidden layer, H fused is the third node feature matrix, W hidden is the weight of the hidden layer, σ(·) is the activation function;

[0097] The current output of the output layer is specifically:

[0098] Y (pred) =H hidden W output ;

[0099] Among them, Y (pred) is the current output of the output layer, W output is the weight of the output layer.

[0100] The present application performs multiple iterative training on the initial extreme learning machine model according to the second node characteristics, optimizes the model parameters based on the mean square error loss to obtain the first extreme learning machine model, and can improve the accuracy of the first extreme learning machine model.

[0101] Step S105: acquiring real-time operating data of the power system to be solved, and obtaining an optimal power flow solution result of the power system to be solved based on the real-time operating data and the first extreme learning machine model.

[0102] In certain embodiments of the present application, obtaining an optimal power flow solution of the power system to be solved according to the real-time operation data and the first extreme learning machine model specifically includes:

[0103] Based on multiple preset task scenarios, multiple scenario feature vectors are constructed;

[0104] According to the multiple scene feature vectors, adjusting the model parameters of the first extreme learning machine model to construct a multi-scene extreme learning machine model;

[0105] According to the real-time operation data and the multi-scenario extreme learning machine model, an optimal power flow solution result of the power system to be solved is obtained.

[0106] In certain embodiments of the present application, after obtaining the optimal power flow solution of the power system to be solved according to the real-time operation data and the first extreme learning machine model, the method further includes:

[0107] According to the optimal power flow solution result, based on the incremental learning idea, the model parameters of the first extreme learning machine model are adjusted.

[0108] In certain embodiments of the present application, the model parameters of the first extreme learning machine model are adjusted based on the incremental learning idea according to the optimal power flow solution result, specifically:

[0109]

[0110] Among them, θ t+1 ,θ t are the updated model parameters and the current model parameters, η is the learning rate, L(θ t , D t ) is based on the new sample D t The loss function is λ, where λ is the regularization weight. corresponds to the current model parameter θ t and a regularization term for the initial model parameters θ0.

[0111] Specifically, a possible implementation of the regularization term is as follows:

[0112]

[0113] Through incremental learning, the new data and old knowledge received by the model can be integrated to maintain the model's adaptability to the power system.

[0114] Compared with the prior art, the present application first obtains the node topology and the historical operation data of each power node, and constructs a system topology map, and then extracts features layer by layer based on a multi-layer graph convolutional neural network to obtain the first node feature, which can gradually enhance the characterization ability of the extracted features for the system topology map layer by layer, and improve the matching degree between the features and the actual topology of the power system. Then, taking the minimum total system output as the screening target, the node features in the first node feature are screened based on regularization and fused to obtain the second node feature, which can screen out unnecessary node feature types, reduce the computational complexity by reducing the node feature types, and improve the computational efficiency and solution accuracy of the optimal power flow solution. Then, according to the second node feature, the initial extreme learning machine model is iteratively trained for multiple times, and the model parameters are optimized based on the mean square error loss to obtain the first extreme learning machine model, which can improve the accuracy of the first extreme learning machine model, thereby improving the accuracy of the optimal power flow solution when the optimal power flow solution result is obtained according to the real-time operation data.

[0115] Corresponding to the previous method, see Figure 2 , the embodiment of the present application provides an optimal power flow solving device based on graph neural network and extreme learning machine, including a data acquisition construction module 210, a node feature extraction module 220, a node feature screening module 230, a model iteration training module 240 and an optimal power flow solving module 250;

[0116] The data acquisition construction module 210 is used to obtain the node topology of the power system to be solved and the historical operation data of each power node, and construct a system topology diagram according to the node topology and the historical operation data; wherein the historical operation data includes node operation characteristics and line electrical characteristics;

[0117] The node feature extraction module 220 is used to extract features from the system topology graph layer by layer based on a preset multi-layer graph convolutional neural network to obtain a first node feature;

[0118] The node feature screening module 230 is used to screen and fuse each node feature in the first node feature based on regularization to obtain a second node feature with the minimum total system output as the screening target;

[0119] The model iteration training module 240 is used to perform multiple iterations of training on the preset initial extreme learning machine model according to the second node feature, and optimize the model parameters of the initial extreme learning machine model based on the mean square error loss during iteration until the training is completed when the model parameters meet the preset values, thereby obtaining the first extreme learning machine model;

[0120] The optimal power flow solving module 250 is used to obtain real-time operating data of the power system to be solved, and obtain the optimal power flow solving result of the power system to be solved according to the real-time operating data and the first extreme learning machine model.

[0121] The present application first obtains the node topology and the historical operation data of each power node, and constructs a system topology map, and then extracts features layer by layer based on a multi-layer graph convolutional neural network to obtain a first node feature, which can gradually enhance the characterization ability of the extracted features for the system topology map layer by layer, and improve the matching degree between the features and the actual topology of the power system, and then takes the minimum total system output as the screening target, and based on regularization, screens the node features in the first node feature and fuses them to obtain the second node feature, which can screen out unnecessary node feature types, and by reducing the node feature types, reduce the computational complexity, and improve the computational efficiency and solution accuracy of the optimal power flow solution, and then iterates the initial extreme learning machine model for multiple times according to the second node feature, and optimizes the model parameters based on the mean square error loss to obtain the first extreme learning machine model, which can improve the accuracy of the first extreme learning machine model, thereby improving the accuracy of the optimal power flow solution when the optimal power flow solution result is obtained according to the real-time operation data.

[0122] It should be understood that the device provided in the embodiment of the present application corresponds to the aforementioned method. The optimal power flow solving device based on graph neural network and extreme learning machine provided in the embodiment of the present application can realize the optimal power flow solving method based on graph neural network and extreme learning machine provided in any embodiment of the present application.

[0123] Adaptively, the embodiments of the present application also provide a computer device and a computer-readable storage medium.

[0124] The computer device comprises: a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor;

[0125] Among them, when the processor executes the computer program, it implements an optimal power flow solution method based on graph neural network and extreme learning machine of the present application.

[0126] The computer-readable storage medium stores a plurality of instructions, and the instructions are suitable for a processor to load to execute an optimal power flow solution method based on a graph neural network and an extreme learning machine of the present application.

[0127] The above are some implementation methods of the present application, and the purpose, technical scheme and beneficial effects of the present application are further described in detail. It should be clear that the above-mentioned partial implementation methods of the present application cannot be understood as limiting the present application. It is particularly pointed out that for those skilled in the art, any changes, modifications, equivalent substitutions and variations made within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A method for solving optimal power flow based on graph neural network and extreme learning machine, characterized in that: include: Obtaining the node topology of the power system to be solved and the historical operation data of each power node, and constructing a system topology diagram according to the node topology and the historical operation data; wherein the historical operation data includes node operation characteristics and line electrical characteristics; Based on a preset multi-layer graph convolutional neural network, feature extraction is performed on the system topology graph layer by layer to obtain a first node feature; Taking the minimum total output of the system as the screening target, each node feature in the first node feature is screened and merged based on regularization to obtain a second node feature; According to the second node feature, a preset initial extreme learning machine model is iterated for multiple times, and model parameters of the initial extreme learning machine model are optimized based on mean square error loss during iteration, until the model parameters meet preset values, the training is completed, and a first extreme learning machine model is obtained; Real-time operation data of the power system to be solved is obtained, and based on the real-time operation data and the first extreme learning machine model, an optimal power flow solution result of the power system to be solved is obtained.

2. The optimal power flow solving method based on graph neural network and extreme learning machine according to claim 1, characterized in that: The constructing a system topology diagram according to the node topology and the historical operation data specifically includes: According to the node topology, a system topology graph is constructed with the node operation characteristics as nodes of the system topology graph and the line electrical characteristics as edges of the system topology graph.

3. The optimal power flow solving method based on graph neural network and extreme learning machine according to claim 1, characterized in that: The method based on the preset multi-layer graph convolutional neural network extracts features of the system topology graph layer by layer to obtain the first node feature, specifically including: Based on the multi-layer graph convolutional neural network, feature extraction is performed on the system topology map layer by layer to obtain first node features; wherein, during feature extraction, each layer of the graph convolutional neural network updates the node features of the previous layer based on graph convolution to obtain the node features of the current layer; each current layer node feature is connected with the initial features of the system topology map before being output to the next layer to obtain the node features of the updated layer.

4. The optimal power flow solving method based on graph neural network and extreme learning machine according to claim 3, characterized in that: Each layer of the graph convolutional neural network updates the node features of the previous layer based on the graph convolution to obtain the node features of the current layer during feature extraction, specifically including: Among them, H (l) , H (l-1) are the node features of the lth layer and the node features of the (l-1)th layer, respectively, W (l) is the weight matrix of the lth layer, σ(·) is the activation function, A is the adjacency matrix corresponding to the system topology graph, D is the degree matrix composed of the degrees of each node in the system topology graph, and D ii =∑ j A ij , D ii is the element value of the i-th row and i-th column of the degree matrix, A ij is the element value in the i-th row and j-th column of the adjacency matrix.

5. The optimal power flow solving method based on graph neural network and extreme learning machine according to claim 1, characterized in that: The method takes the minimum total system output as the screening target, screens and fuses each node feature in the first node feature based on regularization to obtain the second node feature, specifically including: Taking the minimum total output of the system as the screening target, each node feature in the first node feature is screened based on regularization to obtain an intermediate node feature; Based on the preset node weights, the node features of the intermediate node features are fused to obtain the third node feature.

6. The optimal power flow solving method based on graph neural network and extreme learning machine according to claim 5, characterized in that: The screening target is to minimize the total system output, specifically: Among them, Y is the system target power flow, X is the first node feature matrix, β is the feature weight of each node feature of the first node feature, λ is the regularization parameter, ||·||2, ||·||1 are the 2-norm and 1-norm, respectively.

7. The optimal power flow solving method based on graph neural network and extreme learning machine according to claim 1, characterized in that: The step of performing multiple iterative training on a preset initial extreme learning machine model according to the second node feature, optimizing the model parameters of the initial extreme learning machine model based on mean square error loss during iteration until the model parameters meet preset values ​​to complete the training and obtain a first extreme learning machine model, specifically includes: According to the second node feature, the preset initial extreme learning machine model is iteratively trained for multiple times. During the iteration, the output layer weights of the output layer are updated according to the mean square error loss between the current output of the output layer of the initial extreme learning machine model and the actual output, until the training is completed when the mean square error loss meets the preset value, thereby obtaining the first extreme learning machine model; wherein the current output of the output layer is calculated based on the hidden layer.

8. The optimal power flow solution method based on graph neural network and extreme learning machine according to claim 7, characterized in that: The mean square error loss is specifically: min|Y (pred )AND (true) || 2 ; Among them, Y (pred) is the current output of the output layer, Y (true) is the actual output, and ||·|| is the matrix norm.

9. The optimal power flow solution method based on graph neural network and extreme learning machine according to claim 7, characterized in that: The current output of the output layer is calculated based on the hidden layer, specifically including: The output of the hidden layer is specifically: H hidden =σ(H fused W hidden ); Among them, H hidden is the output of the hidden layer, H fused is the third node feature matrix, W hidden is the weight of the hidden layer, σ(·) is the activation function; The current output of the output layer is specifically: Y (pred) =H hiden W output ; Among them, Y (pred) is the current output of the output layer, W output is the weight of the output layer.

10. An optimal power flow solving device based on graph neural network and extreme learning machine, characterized in that: It includes data acquisition construction module, node feature extraction module, node feature screening module, model iteration training module and optimal power flow solution module; The data acquisition construction module is used to obtain the node topology of the power system to be solved and the historical operation data of each power node, and construct a system topology diagram according to the node topology and the historical operation data; wherein the historical operation data includes node operation characteristics and line electrical characteristics; The node feature extraction module is used to extract features from the system topology graph layer by layer based on a preset multi-layer graph convolutional neural network to obtain a first node feature; The node feature screening module is used to screen and fuse each node feature in the first node feature based on regularization to obtain a second node feature with the minimum total system output as the screening target; The model iteration training module is used to perform multiple iteration training on the preset initial extreme learning machine model according to the second node feature, and optimize the model parameters of the initial extreme learning machine model based on the mean square error loss during iteration, until the training is completed when the model parameters meet the preset values, so as to obtain the first extreme learning machine model; The optimal power flow solving module is used to obtain real-time operating data of the power system to be solved, and obtain the optimal power flow solving result of the power system to be solved based on the real-time operating data and the first extreme learning machine model.

Images