Graph generation adversarial learning estimation method for active distribution network topology under incomplete measurement

Abstract

The invention relates to the technical field of power systems, and discloses a graph generation adversarial learning estimation method for active distribution network topology under incomplete measurement, which comprises the following steps of: carrying out topology transformation on a distribution network; converting the original power distribution network topology to obtain a line topological graph; image convolution feature extraction; based on the line topological graph and the graph convolutional neural network, spreading a dependency relationship between node features through a graph structure, extracting spatial features of graph data, and constructing a generative adversarial topology estimation model; a generative adversarial network is constructed, the extracted spatial features are input into the generative adversarial network for training, a generator function is trained through a minimum logarithmic probability, and a discriminator function is trained through a maximum logarithmic probability; constructing an observability layout optimization design model; variational dropout regularization is added to an input layer of the GCN-GAN, and observability layout optimization design model learning reduces the discard rate of important features of correct prediction topology of the GCN-GAN model.

Description

Technical Field

[0001] The present invention relates to the technical field of power systems, and in particular to a graph generation adversarial learning estimation method for active distribution network topology under incomplete measurement. Background Art

[0002] With the continuous advancement of the construction of new power systems, the scale of power grids continues to expand, and the access to distributed energy sources continues to increase. Due to the volatility, randomness and intermittency of distributed energy output, its power generation and stability are difficult to predict and control, which will threaten the quality of power supply and the stability of distribution network operation. In addition, the operating characteristics of the distribution network will also undergo profound changes. The bidirectional flow distribution, the gradually blurred source-load boundary, the more diverse operating modes, the higher demand for flexible resources, the power balance from determinism to probability, and the mutual coupling of multiple energy sources have brought new challenges to the safe operation of the distribution network. At the same time, the access of new loads such as electric vehicles and energy storage systems and the comprehensive development of demand-side response have caused the uncertainty factors faced by the distribution network to rise sharply, and the operation and management of the distribution network will therefore face a more complex and changeable environment. The correct identification and interpretation of the topological structure is a key task of the power system, because accurate network topology knowledge is a prerequisite for system planning, operation and control, and lays a necessary foundation for its operation monitoring and management.

[0003] Topology identification is a key building block that can provide valuable information to serve other applications such as state estimation, power flow calculation, fault analysis, network reconstruction, economic dispatch, reactive voltage regulation, resilience improvement, demand response, etc. The topology is determined by the status of the section switches and tie switches on the line. In actual operation, the system model that dispatchers rely on usually comes from planning documents. When performing maintenance, overhaul, responding to faults, or implementing optimal control strategies, the status of section switches and tie switches will change, resulting in dynamic adjustment of the system's topology. This adjustment may cause significant deviations between the actual topology and the recorded structure. In order to accurately capture these changes, topology identification is particularly critical. Summary of the invention

[0004] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a graph generation adversarial learning estimation method for active distribution network topology under incomplete measurement, aiming to solve the technical problems of poor adaptability and low learning efficiency of traditional topology identification methods.

[0005] The object of the present invention is achieved by the following technical solution: A graph generation adversarial learning estimation method for active distribution network topology under incomplete measurement, comprising the following steps: Step S1, distribution network topology transformation; The original distribution network topology is transformed to obtain a line topology diagram; Step S2, graph convolution feature extraction; Based on the line topology graph, the graph convolutional neural network (GCN) propagates the dependencies between node features through the graph structure, extracts the spatial features of the graph data, obtains hidden information, and thus updates its own line status; Step S3, constructing a generative adversarial topology estimation model; Construct a generative adversarial network and input the extracted spatial features into the generative adversarial network for training. The generator function is trained by minimizing the logarithmic probability, and the discriminator function is trained by maximizing the logarithmic probability. In order to generate a topology corresponding to a specific operating scenario, additional information needs to be added to the model input, including the line status, total active power, and total reactive power after the graph convolutional neural network is updated. Step S4, constructing an observable layout optimization design model; By adding variational Dropout regularization to the input layer of GCN-GAN, the observable layout optimization design model learns to reduce the discard rate of important topological features that are correctly predicted by the GCN-GAN model, while increasing the discard rate of other unimportant features, thereby effectively screening out important current features.

[0006] Specifically, in step S1, according to the Whitney isomorphism principle, for each branch in the original distribution network topology, a node is created in the line graph topology; for every two branches in the original distribution network topology that share a node, a branch is created between their corresponding line graph topology nodes to obtain a line topology graph.

[0007] Specifically, in step S2, the line topology graph is represented as an undirected connected graph ,in, Represents a node, Represents the edge, adjacency matrix The topological structure is represented in the form of a binary matrix, where each element indicates whether the nodes are connected. The node feature matrix Contains the available measurements for each node; matrix Defined as:

[0008] in, represents the magnitude of the current, Represents the phase angle of the current.

[0009] Specifically, the graph convolutional neural network uses the adjacency matrix and the node feature matrix The binary group As input, its expression is:

[0010] in, Indicates The hidden output feature matrix of the layer; In the In the layer, the forward propagation rule of GCN is defined as:

[0011] in, Indicates The learned weight matrix of the layer, express Activation function.

[0012] Specifically, the unit matrix is ​​introduced in step S2:

[0013] in, is the adjacency matrix with self-connected line graph topology, is the identity matrix; Adjacency Matrix Normalize as follows:

[0014] in, represents the degree-absolute matrix of the self-loop adjacency matrix; No. The layer-by-layer propagation rule is reformulated as:

[0015] when When: .

[0016] Specifically, in step S3, the generative adversarial network simultaneously trains the generator function and the discriminator function; The training objective of the generator function is to minimize the following log probability:

[0017] in, represents the mathematical expectation, and represents the generator function and the discriminator function, represents the generator function network parameters, represents random noise, and represents its probability distribution; The discriminator function is trained by maximizing the following log probability:

[0018] in, represents the discriminator function network parameters, Represents the actual topology The probability distribution of Specifically, in the optimization process of the generative adversarial network, the generator function and the discriminator function compete with each other to form a zero-sum game. The learning estimation model of the graph generative adversarial network is expressed as a minimax game in the following form: .

[0019] Specifically, in order to generate a topology suitable for a specific state in step S3, the following additional information is introduced:

[0020] in, is the current feature extracted by step 2, represents the total active power of the system, represents the total reactive power of the system, and the minimax game is reformulated as: .

[0021] Specifically, the step S3 uses Wasserstein distance to define the loss function of the generator, and the Wasserstein distance is defined as follows:

[0022] Among them, the discriminator function needs to satisfy the 1-Lipschitz condition, and a loss term is set in the topology estimation model as follows:

[0023]

[0024] in, represents the penalty parameter, is a random number uniformly distributed in the interval [0,1]. represents the pseudo-topology generated by the generator function, For the real topology, For the general The second norm of the gradient after entering the discriminator, To pass The expected value obtained by sampling.

[0025] Specifically, in step S4, the variational Dropout randomly discards some neurons with a certain probability, thereby improving learning efficiency and reducing the risk of overfitting. The Dropout ratio of the randomly discarded neurons is obtained by minimizing the following logarithmic probability:

[0026] in, represents a mask that follows the Bernoulli distribution and is obtained by sampling according to the above formula, represents the discard rate, Represents the Sigmoid function; In this setting, the loss functions of the generator function and the discriminator function are respectively expressed as:

[0027]

[0028] The weight parameters of the generator function and the discriminator function are updated by applying the stochastic gradient descent method in the following ways:

[0029]

[0030] in, The representation is either noise or a mini-batch of topology samples.

[0031] The present invention has the following advantages: 1. The adversarial learning estimation method of the present invention can effectively seek the minimum measurement configuration set that supports high-accuracy identification of the global topology, solve the layout problem of a small number of line current sensors, ensure the practicality and economy of the above-mentioned topology identification method, and provide an overall solution from measurement layout to topology identification for engineering applications.

[0032] 2. It can identify topology by learning the distribution characteristics of topological data closely related to current and load information, and shows better identification effect than traditional supervised learning methods on unknown topology test sets. At the same time, it significantly enhances the algorithm's anti-noise ability by increasing the redundancy of topology identification related data. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 A flowchart of a graph generation adversarial learning estimation method for active distribution network topology under incomplete measurement of the present invention; Figure 2 It is a schematic diagram of topological transformation of the present invention; Figure 3 A model architecture diagram of the graph generation adversarial learning estimation method for active distribution network topology under incomplete measurement of the present invention. DETAILED DESCRIPTION

[0034] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention, that is, the embodiments described are only part of the embodiments of the present invention, rather than all of the embodiments. The components of the embodiments of the present invention described and shown in the drawings herein can be arranged and designed in various different configurations.

[0035] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present invention.

[0036] It should be noted that relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "including a..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0037] The present invention is further described below in conjunction with the accompanying drawings, but the protection scope of the present invention is not limited to the following description.

[0038] like Figures 1 to 3 As shown, the graph generation adversarial learning estimation method of active distribution network topology under incomplete measurement includes the following steps: Step S1, distribution network topology transformation; According to the Whitney isomorphism principle, the distribution network topology transformation is performed to obtain the topological graph after the node edges are interchanged. According to the Whitney isomorphism principle, if the graph G is connected, then except for one special case, the graph L(G) after the topological transformation can always restore the graph G, so such transformation is equivalent. For each branch in the original distribution network topology, a node is created in the line graph topology; for every two branches in the original distribution network topology that share a node, a branch is created between their corresponding line graph topological nodes to obtain a line topological graph; Step S2, graph convolution feature extraction; Based on the line topology graph, the graph convolutional neural network (GCN) propagates the dependencies between node features through the graph structure, extracts the spatial features of the graph data, obtains hidden information, and thus updates its own line status; Represent the converted line topology graph as an undirected connected graph ,in, Represents a node, Represents the edge, adjacency matrix The topological structure is represented in the form of a binary matrix, where each element indicates whether the nodes are connected. The node feature matrix Contains the available measurements for each node; matrix Defined as:

[0039] in, represents the magnitude of the current, Indicates the phase angle of the current; the current measurement data is collected by the current transformer; The graph convolutional neural network uses an adjacency matrix and the node feature matrix The binary group As input, each graph node representing a busbar aggregates the current information of itself and its neighboring nodes by performing a convolution operation to capture the topological or spatial connectivity in the current measurement data. For multi-layer GCN, its graph convolution layer is usually approximated as a nonlinear function, which is expressed as:

[0040] in, Indicates The hidden output feature matrix of the layer; In the In the layer, the forward propagation rule of GCN is defined as:

[0041] in, Indicates The learned weight matrix of the layer, express Activation function; However, this propagation rule has two problems. First, there are no self-connected nodes, which leads to the adjacency matrix The diagonal elements of are all zero. In the feature extraction process, the current measurement data of the node itself is ignored. In order to solve this problem, the unit matrix is ​​introduced:

[0042] in, is the adjacency matrix with self-connected line graph topology, is the identity matrix; Another problem is the adjacency matrix Not normalized. Due to the different node degrees, multiple convolution operations will lead to asymmetric distribution of current amplitude and phase angle, which may cause gradient vanishing or explosion problems during back propagation. To solve this problem, the adjacency matrix Normalize as follows:

[0043] in, represents the degree matrix of the self-loop adjacency matrix; then, The layer-by-layer propagation rule is reformulated as:

[0044] when When:

[0045] Hidden features extracted by propagation rules It contains information about current flow and its topology or spatial connectivity; Step S3, constructing a generative adversarial topology estimation model; A generative adversarial network is constructed, and the extracted current characteristics and operating status information of the load demand are input into the generative adversarial network to conditionally generate topology samples, and additional information is added to the model input to generate a topology corresponding to a specific operating scenario, wherein the generator function is trained by minimizing the logarithmic probability, and the discriminator function is trained by maximizing the logarithmic probability. In order to generate a topology corresponding to a specific operating scenario, additional information needs to be added to the model input, including the line status, total active power, and total reactive power after the graph convolutional neural network is updated; In order to overcome the timeliness problem of the traditional physical model and the generalization problem of supervised learning, the Generative Adversarial Network (GAN) is considered to be used for distribution network topology identification. The distribution network topology identification driven by the generative adversarial network is similar to supervised learning. Relying on machine learning technology, its identification speed is better than the distribution network identification model driven by the traditional physical model. The difference is that the generative adversarial network does not directly learn the mapping relationship between measurement information and topology, but trains two networks with opposing optimization objectives, indirectly mining the distribution of topology, thereby enhancing the generalization ability of the algorithm. Specifically, GAN trains the generator function and the discriminator function at the same time, and the two are trained in a competitive manner. Both the generator function and the discriminator function are modeled by neural networks with strong expressive power. The generator function takes random noise as input and tries to approximate the real topological distribution, thereby "cheating" the discriminator function. The training objective of the generator function is to minimize the following log probability:

[0046] in, represents the mathematical expectation, and represents the generator function and the discriminator function, represents the generator function network parameters, represents random noise, and represents its probability distribution; The discriminator function receives the true topology encoding and the pseudo topology distribution generated by the generator function to output a probability to judge the authenticity of the input topology. The discriminator function is trained by maximizing the following log probability:

[0047] in, represents the discriminator function network parameters, Represents the actual topology The probability distribution of In the above-mentioned generative adversarial network optimization process, the generator function and the discriminator function compete with each other, forming a zero-sum game. The learning estimation model of the graph generative adversarial network is expressed as a minimax game in the following form:

[0048] However, this topology estimation model is difficult to generate a topology corresponding to a specific scenario. In fact, the current measurement data will also change when the load demand and topology conditions change over time. In order to generate a topology suitable for a specific state, the following additional information is introduced:

[0049] in, is the current feature extracted by step S2, represents the total active power of the system, represents the total reactive power of the system, and the minimax game is reformulated as: .

[0050] The training method of the generative adversarial topology estimation model is specifically as follows: On the one hand, when the discriminator is trained to a certain extent, GAN can be regarded as using JS divergence to measure the distance between the real data distribution and the data distribution generated by the generator, and the JS divergence between any two non-overlapping distributions is always log2. It should be noted that it is very likely that the real topological distribution and the estimated topological distribution of the generator do not overlap. At this time, the loss function of the generator will remain constant, causing the gradient to disappear and unable to continue training, making the model lose its recognition ability; on the other hand, during the GAN training process, it is difficult for the discriminator and the generator to converge at the same time, and mode collapse is prone to occur, resulting in the generator being able to generate only a few topologies, and unable to achieve the correct correspondence between current measurement information and topology.

[0051] Unlike JS divergence, Wasserstein distance can still accurately reflect the distance between two non-overlapping distributions, thereby providing an accurate training direction for the generator to fit the topological data distribution. Therefore, using Wasserstein distance to define the loss function of the generator can effectively solve the above problem. Wasserstein distance is defined as follows:

[0052] Among them, the discriminator function needs to satisfy the 1-Lipschitz condition, but this may sacrifice the convergence speed to a certain extent. Therefore, a loss term is set in the topology estimation model as follows:

[0053]

[0054] in, represents the penalty parameter, is a random number uniformly distributed in the interval [0,1]. represents the pseudo-topology generated by the generator function, For the real topology, For the general The second norm of the gradient after entering the discriminator, for pass The expected value obtained by sampling.

[0055] Step S4, constructing an observable layout optimization design model; Variational Dropout regularization is added to the input layer of GCN-GAN, and the observable layout optimization design model learns to reduce the discard rate of important topological features that the GCN-GAN model correctly predicts, while increasing the discard rate of other unimportant features, thereby effectively screening out important current features and providing theoretical support for the selection of current sensor locations.

[0056] In order to identify which current features are sensitive to topology recognition accuracy, a Dropout feature sorting strategy is introduced to screen key current features and guide the deployment of current transformers. The idea of ​​this Dropout feature sorting is essentially a regularization method, which improves learning efficiency and reduces the risk of overfitting by randomly discarding some neurons with a certain probability. The Dropout ratio of randomly discarded neurons is obtained by minimizing the following logarithmic probability:

[0057] in, represents a mask that follows the Bernoulli distribution and is obtained by sampling according to the above formula, represents the discard rate, Represents the Sigmoid function; In this setting, the loss functions of the generator function and the discriminator function are respectively expressed as:

[0058]

[0059] The weight parameters of the generator function and the discriminator function are updated by applying the stochastic gradient descent method in the following ways:

[0060]

[0061] in, The representation is either noise or a mini-batch of topology samples.

[0062] The current feature importance ranking by Dropout is integrated into the proposed topology estimation model, which improves the quality of the generated topology and the recognition accuracy. During the training process, the discriminator function gradually reduces the Wasserstein distance between the true topology and the generated topology, and adjusts the Dropout ratio of the current features by Dropout ranking. At the same time, the generator function adaptively generates topology according to the feedback of the discriminator function to better match the true topology distribution and conditional information.

[0063] Obviously, the GCN-GAN embedded with Dropout FR incorporates the importance of features into the learning process of the model, so that feature selection and model optimization are carried out simultaneously, and the optimization goals of the two are consistent, which helps to better select features that contribute to the performance of the GCN-GAN model and improve the generation quality and discrimination accuracy of the model. During the training process, the discriminator will continuously reduce the Wasserstein distance between the true topology and the generator's estimated topology. At the same time, under the influence of the regularization term of Dropout FR, it will also continuously control the retention probability of the features to optimize its feature selection ability, and the generator will adjust its generated topology according to the feedback of the discriminator to better conform to the distribution and measurement characteristics of the true topology.

[0064] It should be noted that Dropout FR can be regarded as a feature selection method based on policy gradient, that is, optimizing the output of the best feature combination in the feature space corresponding to the topological structure of the training set. In other words, the layout optimization strategy of the proposed model is to select a layout scheme with the best identification effect from a variety of layout methods based on the topological structure in the training set. In practice, since the distribution network topology needs to remain radial, its common forms are limited. Therefore, it is entirely possible to build a common feasible topology library by collecting historical operating topological structures, and use its flow calculation results as the basic data of the training set, so as to achieve precise control of the layout scheme for the identification of common feasible topologies.

[0065] The above is only a preferred embodiment of the present invention and does not limit the present invention in any form. Any technician familiar with the art can make many possible changes and modifications to the technical solution of the present invention by using the above-mentioned technical content without departing from the scope of the technical solution of the present invention, or modify it into an equivalent embodiment of equivalent changes. Therefore, any changes, modifications, equivalent changes and modifications made to the above embodiments based on the technology of the present invention without departing from the content of the technical solution of the present invention belong to the protection scope of the present technical solution.

Claims

1. A graph generation adversarial learning estimation method for active distribution network topology under incomplete measurement, characterized by: The following steps are involved: Step S1, distribution network topology transformation; The original distribution network topology is transformed to obtain a line topology diagram; Step S2, graph convolution feature extraction; Based on the line topology graph, the graph convolutional neural network propagates the dependency between node features through the graph structure, extracts the spatial features of the graph data, obtains hidden information, and thus updates its own line status; Step S3, constructing a generative adversarial topology estimation model; Construct a generative adversarial network and input the extracted spatial features into the generative adversarial network for training, where the generator function is trained by minimizing the logarithmic probability and the discriminator function is trained by maximizing the logarithmic probability; Step S4, constructing an observable layout optimization design model; Variational Dropout regularization is added to the input layer of GCN-GAN. The observable layout optimization design model is learned to reduce the discard rate of important topological features correctly predicted by the GCN-GAN model, while increasing the discard rate of other unimportant features to screen out important current features.

2. The graph generation adversarial learning estimation method for active distribution network topology under incomplete measurement according to claim 1 is characterized by: In step S1, according to the Whitney isomorphism principle, for each branch in the original distribution network topology, a node is created in the line graph topology; for every two branches in the original distribution network topology that share a node, a branch is created between their corresponding line graph topology nodes to obtain a line topology graph.

3. The graph generation adversarial learning estimation method for active distribution network topology under incomplete measurement according to claim 1 is characterized by: In step S2, the line topology graph is represented as an undirected connected graph. ,in, Represents a node, Represents the edge, adjacency matrix The topological structure is represented in the form of a binary matrix, where each element indicates whether the nodes are connected. The node feature matrix Contains the available measurements for each node; matrix Defined as: ； in, represents the magnitude of the current, Represents the phase angle of the current.

4. The graph generation adversarial learning estimation method for active distribution network topology under incomplete measurement according to claim 3 is characterized by: The graph convolutional neural network uses an adjacency matrix and the node feature matrix The binary group As input, its expression is: ； in, Indicates The hidden output feature matrix of the layer; In the In the layer, the forward propagation rule of GCN is defined as: ； in, Indicates The learned weight matrix of the layer, express Activation function.

5. The graph generation adversarial learning estimation method for active distribution network topology under incomplete measurement according to claim 4 is characterized by: The unit matrix is ​​introduced in step S2: ； in, is the adjacency matrix with self-connected line graph topology, is the identity matrix; Adjacency Matrix Normalize as follows: ； in, represents the degree-absolute matrix of the self-loop adjacency matrix; No. The layer-by-layer propagation rule is reformulated as: ； when When: 。 6. The graph generation adversarial learning estimation method for active distribution network topology under incomplete measurement according to claim 1, characterized in that: In step S3, the adversarial network is generated to train the generator function and the discriminator function simultaneously; The training objective of the generator function is to minimize the following log probability: ； in, represents the mathematical expectation, and represents the generator function and the discriminator function, represents the generator function network parameters, represents random noise, and represents its probability distribution; The discriminator function is trained by maximizing the following log probability: ； in, represents the discriminator function network parameters, Represents the actual topology The probability distribution of .

7. The graph generation adversarial learning estimation method for active distribution network topology under incomplete measurement according to claim 6 is characterized by: In the optimization process of the generative adversarial network, the generator function and the discriminator function compete with each other to form a zero-sum game. The learning estimation model of the graph generative adversarial network is expressed as a minimax game in the following form: 。 8. The graph generation adversarial learning estimation method for active distribution network topology under incomplete measurement according to claim 7 is characterized by: In order to generate a topology suitable for a specific state in step S3, the following additional information is introduced: ； in, is the current feature extracted by step S2, represents the total active power of the system, represents the total reactive power of the system, and the minimax game is reformulated as: 。 9. The graph generation adversarial learning estimation method for active distribution network topology under incomplete measurement according to claim 8, characterized in that: The step S3 uses Wasserstein distance to define the loss function of the generator, and the Wasserstein distance is defined as follows: ； Among them, the discriminator function needs to satisfy the 1-Lipschitz condition, and a loss term is set in the topology estimation model as follows: ； ； in, represents the penalty parameter, is a random number uniformly distributed in the interval [0,1]. represents the pseudo-topology generated by the generator function, For the real topology, For the general The second norm of the gradient after entering the discriminator, To pass The expected value obtained by sampling.

10. The graph generation adversarial learning estimation method for active distribution network topology under incomplete measurement according to claim 9, characterized in that: In step S4, the variational Dropout regularization randomly discards some neurons with a certain probability, thereby improving learning efficiency and reducing the risk of overfitting. The Dropout ratio of randomly discarded neurons is obtained by minimizing the following logarithmic probability: ； in, represents a mask that follows the Bernoulli distribution and is obtained by sampling according to the above formula, represents the discard rate, Represents the Sigmoid function; In this setting, the loss functions of the generator function and the discriminator function are respectively expressed as: ； ； The weight parameters of the generator function and the discriminator function are updated by applying the stochastic gradient descent method in the following ways: ； ； in, The representation is either noise or a mini-batch of topology samples.

Images