Power distribution network power quality disturbance deduction method based on graph neural network

Abstract

The invention discloses a power distribution network power quality disturbance deduction method based on a graph neural network, and the method comprises the steps: obtaining a topological structure of a power distribution network power system, and converting the topological structure into a graph network structure; selecting feature vectors of nodes in the graph network structure; based on the feature vectors, training a graph neural network model through node historical data; and based on the source load data, performing power quality situation deduction through a graph neural network model. According to the method provided by the invention, the electric energy quality information of each measuring point in the power grid can be quickly and accurately deduced through the distributed power supply and load information in the power distribution network.

Description

Technical Field

[0001] The present invention relates to the technical field of distribution network, and in particular to a method for deducing power quality disturbances in a distribution network based on a graph neural network. Background Art

[0002] Driven by the "dual carbon" goal, a high proportion of distributed renewable energy is connected to the power grid. Due to the randomness and volatility of photovoltaic and wind power generation, many types of complex power quality problems such as voltage deviation, harmonic content and three-phase imbalance have emerged in the distribution network. Among them, the specific mechanism logic of the impact of distributed power generation access on the power quality problems of each node in the distribution network is relatively complex. The main mechanism includes changing the power flow distribution of the power grid, injecting harmonic currents, and affecting the load distribution mode of the power grid. When managing and warning power quality problems in the distribution network, if it is expected to adopt a mechanism derivation method, through the active and reactive output, harmonic emission level and grid-connected position of the distributed power source, based on the power system mechanism equations such as Ohm's law, the propagation law of multiple types of power quality problems in the system will be deduced, which will result in problems such as long time consumption, narrow area and large information demand, which is not conducive to the efficient management and lightweight warning of multiple types of power quality problems in the distribution network. Therefore, how to quickly and accurately complete the deduction of the propagation of multiple types of power quality disturbances in the distribution network based on a small amount of data is an urgent problem to be solved by the regional power grid. Summary of the invention

[0003] In view of this, the purpose of the present invention is to provide a distribution network power quality disturbance deduction method based on graph neural network, which can quickly and accurately deduce the power quality information of each measuring point in the power grid through the distributed power source and load information in the distribution network, so as to better support the problem management and early warning analysis in areas with frequent power quality problems.

[0004] In a first aspect, an embodiment of the present invention provides a method for deducing power quality disturbances in a distribution network based on a graph neural network, characterized in that it includes:

[0005] Acquire a topological structure of a power distribution network system, and convert the topological structure into a graph network structure;

[0006] Selecting feature vectors of nodes in the graph network structure;

[0007] Based on the feature vector, the graph neural network model is trained through the node historical data;

[0008] Based on the source-load data, the power quality situation is deduced through the graph neural network model.

[0009] Preferably, converting the topological structure into the graph network structure comprises:

[0010] The distributed power sources, power loads, and power quality observation points in the topological structure are used as various types of nodes in the graph network structure, namely, type I nodes, type II nodes, and type III nodes; the physical link relationships among the distributed power sources, the power loads, and the power quality observation points are used as edges in the graph network structure;

[0011] Among them, the characteristic dimensions of the first type of nodes and the second type of nodes are 3 dimensions, and the characteristic dimension of the third type of nodes is 4 dimensions.

[0012] Preferably, the link directions between the various types of nodes are determined based on the types of the physical link relationships among the distributed power sources, the power loads and the power quality observation points.

[0013] Preferably, the characteristic vector of the first type of node includes active output, reactive output and harmonic current emission level; the characteristic vector of the second type of node includes active power, reactive power and harmonic current emission level; the characteristic vector of the third type of node includes voltage deviation, three-phase imbalance, total harmonic distortion and power factor.

[0014] Preferably, before training the graph neural network model using the node history data, the feature vector is structured, including:

[0015] A time-scale alignment algorithm is used to reconstruct the time-scales of the first-category nodes, the second-category nodes, and the third-category nodes, so that the feature vectors of the above-mentioned various types of nodes can maintain a corresponding relationship at the same time.

[0016] Preferably, the graph neural network model is constructed according to the graph network structure, and the graph neural network model adopts two layers of graph convolutional neural network and one layer of fully connected network;

[0017] Based on the feature vector, training the graph neural network model through the node historical data includes:

[0018] Using the graph convolutional neural network of the first layer, the feature dimension of the feature vectors of the first type of nodes and the second type of nodes is expanded from 3 dimensions to 64 dimensions to achieve feature extraction;

[0019] Using the graph convolutional neural network of the second layer, the feature dimension of the feature vectors of the first type of nodes and the second type of nodes is reduced from 64 dimensions to 32 dimensions, thereby achieving feature dimensionality reduction;

[0020] Mapping the feature dimensions of the feature vectors of the first-category nodes and the second-category nodes to the feature dimensions of the feature vectors of the third-category nodes using the fully connected network;

[0021] Divide the training set and the validation set based on the historical data of the first type of nodes, the second type of nodes, and the third type of nodes;

[0022] Training the graph network model based on the training set, repeatedly calculating the loss function of the model deduction value and the label value, and gradually reducing the loss value until the loss function converges;

[0023] The historical data of the first type of nodes and the second type of nodes are used as input features, and the historical data of the third type of nodes are used as verification labels.

[0024] Preferably, in the graph convolutional neural network, the node representation of each convolutional layer is updated by a propagation rule, and the expression of the propagation rule is:

[0025]

[0026] In the formula, H k is the node representation of the convolutional layer of the kth layer; is the self-connected adjacency matrix of the undirected graph, allowing the node features themselves to be included when updating the node representation, A is the adjacency matrix, I∈R N×N is the identity matrix; is a diagonal matrix; σ is the activation function; is a hierarchical linear transformation matrix that can be trained during the training of the graph neural network model, where F k and F k+1 They are respectively the dimensions of the node representations of the k-th convolutional layer and the k+1-th convolutional layer.

[0027] Preferably, PyQt5 is used to construct a visual interactive interface, and the user can input the source and load data in the terminal through the visual interactive interface, and perform power quality situation deduction through the graph neural network model packaged to the terminal.

[0028] The beneficial effects of the present invention are as follows:

[0029] The data-driven graph neural network model is used to deduce the disturbance and propagation of various types of power quality problems in the distribution network, overcoming the shortcomings of the traditional mechanism deduction, such as slow speed, large data requirements and regional restrictions. The power quality information of each measuring point in the power grid can be quickly and accurately deduced through the distributed power source and load information in the distribution network, so as to better support the problem management and early warning analysis of power quality frequent areas.

[0030] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description, claims and drawings.

[0031] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0033] Figure 1 A schematic diagram of a flow chart of a method for deducing power quality disturbances in a distribution network based on a graph neural network provided in an embodiment of the present invention;

[0034] Figure 2 A conversion relationship diagram of the topological structure of an actual low-voltage distribution network power system and the graph network structure of the distribution network power quality disturbance deduction method based on the graph neural network provided in an embodiment of the present invention;

[0035] Figure 3 A schematic diagram of the topological structure of measured data of a distribution network power system according to a distribution network power quality disturbance deduction method based on a graph neural network provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0036] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are 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 ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0037] To facilitate understanding of this embodiment, Figure 1 The method for deducing power quality disturbances in a distribution network based on a graph neural network disclosed in an embodiment of the present invention is introduced in detail.

[0038] Embodiment 1:

[0039] This embodiment provides a distribution network power quality disturbance deduction method based on a graph neural network, including:

[0040] Step 1: Obtain the topological structure of the distribution network power system and convert the topological structure into a graph network structure.

[0041] Among them, the topological structure of the distribution network power system is composed of all key power equipment in the distribution network power system and the connection relationship between these key power equipment.

[0042] Preferably, converting the topological structure into a graph network structure comprises:

[0043] The distributed power sources, power loads, and power quality observation points in the topological structure are regarded as various types of nodes in the graph network structure, namely, type I nodes, type II nodes, and type III nodes; the physical link relationships among the distributed power sources, power loads, and power quality observation points are regarded as edges in the graph network structure;

[0044] Among them, the feature dimension of type I nodes and type II nodes is 3 dimensions, and the feature dimension of type III nodes is 4 dimensions.

[0045] Furthermore, based on the type of physical link relationship between the distributed power sources, the power loads, and the power quality observation points, the link directions between the various types of nodes are determined.

[0046] In this embodiment, since there is a mutual influence rule between various types of nodes in the graph network structure, all edges are set as bidirectional links.

[0047] Step 2: Select the feature vectors of nodes in the graph network structure.

[0048] In this embodiment, the actual internal measurement data are power quality indicators to be evaluated, including: voltage deviation, three-phase unbalance, total harmonic distortion and power factor.

[0049] Furthermore, the characteristic vectors of type I nodes include active output, reactive output and harmonic current emission level; the characteristic vectors of type II nodes include active power, reactive power and harmonic current emission level; and the characteristic vectors of type III nodes include voltage deviation, three-phase imbalance, total harmonic distortion and power factor.

[0050] Step 3: Based on the feature vector, the graph neural network model is trained using the node historical data.

[0051] Furthermore: Before training the graph neural network model through the node history data, the feature vector is structured, including:

[0052] The time-scale alignment algorithm is used to reconstruct the time-scales of the first-class nodes, the second-class nodes, and the third-class nodes, so that the feature vectors of the above-mentioned various types of nodes can maintain a corresponding relationship at the same time.

[0053] In this embodiment, the node historical data includes: source-load disturbance data at historical time, and power quality coupling data at historical time.

[0054] Furthermore, the graph neural network model is constructed according to the graph network structure. The graph neural network model uses two layers of graph convolutional neural network (GCN) and one layer of fully connected network;

[0055] Based on the feature vector, the graph neural network model is trained through the node historical data including:

[0056] Step 301: using the first layer of graph convolutional neural network, the feature dimension of the feature vectors of the first-class nodes and the second-class nodes is expanded from 3 dimensions to 64 dimensions to achieve feature extraction;

[0057] Step 302: using the second-layer graph convolutional neural network, the feature dimension of the feature vectors of the first-class nodes and the second-class nodes is reduced from 64 dimensions to 32 dimensions, thereby achieving feature dimensionality reduction;

[0058] Step 303: Mapping the feature dimensions of the feature vectors of the first-class nodes and the second-class nodes to the feature dimensions of the feature vectors of the third-class nodes using a fully connected network;

[0059] Step 304: Divide the training set and the validation set based on the historical data of the first-class nodes, the second-class nodes, and the third-class nodes;

[0060] Step 305: Train the graph network model based on the training set, repeatedly calculate the loss function of the model deduction value and the label value, and gradually reduce the loss value until the loss function converges;

[0061] Among them, the historical data of type I nodes and type II nodes are used as input features, and the historical data of type III nodes are used as verification labels.

[0062] In this embodiment, the division of training sets and validation sets based on historical data of type I nodes, type II nodes, and type III nodes includes: randomly sampling the historical data of type I nodes, type II nodes, and type III nodes, and randomly extracting 80% of the historical data as a training set and the remaining 20% ​​of the historical data as a test set while ensuring time scale alignment.

[0063] In this embodiment, when the loss function of the validation set converges, the trained graph neural network model is exported.

[0064] Furthermore, in the graph convolutional neural network, the node representation of each convolutional layer is updated by the propagation rule, and the expression of the propagation rule includes:

[0065]

[0066] In the formula, H kis the node representation of the kth convolutional layer; is the self-connected adjacency matrix of the undirected graph, allowing the node features themselves to be included when updating the node representation, A is the adjacency matrix, I∈R N×N is the identity matrix; is a diagonal matrix; σ is the activation function; is a hierarchical linear transformation matrix that can be trained during the training of the graph neural network model, where F k and F k+1 are the dimensions of the node representations of the k-th convolutional layer and the k+1-th convolutional layer, respectively.

[0067] In this embodiment, the activation function is a ReLU function or a Tanh function.

[0068] Step 4: Based on the source and load data, the power quality situation is deduced through the graph neural network model.

[0069] Among them, the power quality situation deduction is carried out through the graph neural network model, which can obtain the power quality data of each power quality observation point.

[0070] In this embodiment, the source load data is the source load disturbance data of the area to be deduced at a certain moment.

[0071] Preferably, PyQt5 is used to build a visual interactive interface, and the user can input source and load data in the terminal through the visual interactive interface, and perform power quality situation deduction through the graph neural network model packaged to the terminal.

[0072] Figure 2 It is a conversion relationship diagram between the topological structure of the actual low-voltage distribution network power system and the graph network structure.

[0073] Figure 3 This is a schematic diagram of the topological structure of the measured data of the distribution network power system from 23 measurement points in a certain area of ​​Henan Province. Figure 3 In the embodiment, the power distribution network power system includes 23 power quality observation points, 6 groups of distributed power sources, and 6 groups of power loads. Among them, nodes 23 to 28 are six groups of renewable energy power generation facilities; nodes 29 to 34 are six groups of power loads; nodes 0 to 22 are 23 bus nodes. Based on the method provided in this embodiment, the distributed power sources, power loads, and power quality observation points in the topological structure are respectively regarded as Class I nodes, Class II nodes, and Class III nodes. Among them, the characteristics of Class I nodes and the characteristic vectors of Class II nodes are active output and active power, respectively, and the characteristic vector of Class III nodes is voltage deviation.

[0074] The above-mentioned feature vector is used as measurement data. Based on the method provided in this embodiment, the 15-minute measurement data measured by the terminal within one day is selected for training and verification. PyQt5 is used to build a visual interactive interface. The user can input source and load data in the terminal through the visual interactive interface, and the power quality situation is deduced through the graph neural network model.

[0075] When the user inputs the active output of 6 groups of distributed power sources and the active power of 6 groups of power loads, the terminal can immediately deduce the voltage deviation of 23 measuring points through the program, and the calculation speed is less than 0.5s. 2 The test accuracy is greater than 99%, which can replace the traditional mechanism deduction model to carry out the deduction of power quality problems in the distribution network.

[0076] Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-described embodiments within the technical scope disclosed by the present invention, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A method for deducing power quality disturbances in distribution networks based on graph neural networks, characterized in that: include: Acquire a topological structure of a power distribution network system, and convert the topological structure into a graph network structure; Selecting feature vectors of nodes in the graph network structure; Based on the feature vector, the graph neural network model is trained through the node historical data; Based on the source-load data, the power quality situation is deduced through the graph neural network model.

2. The method for deducing power quality disturbance in distribution network based on graph neural network according to claim 1 is characterized in that: Converting the topological structure into the graph network structure includes: The distributed power sources, power loads, and power quality observation points in the topological structure are used as various types of nodes in the graph network structure, namely, type I nodes, type II nodes, and type III nodes; the physical link relationships among the distributed power sources, the power loads, and the power quality observation points are used as edges in the graph network structure; Among them, the characteristic dimensions of the first type of nodes and the second type of nodes are 3 dimensions, and the characteristic dimension of the third type of nodes is 4 dimensions.

3. The method for deducing power quality disturbance in distribution network based on graph neural network according to claim 2 is characterized in that: Based on the type of the physical link relationship between the distributed power source, the power load, and the power quality observation point, the link direction between each type of the nodes is determined.

4. The method for deducing power quality disturbance of distribution network based on graph neural network according to claim 2 is characterized in that: The characteristic vectors of the first type of nodes include active output, reactive output and harmonic current emission level; the characteristic vectors of the second type of nodes include active power, reactive power and harmonic current emission level; the characteristic vectors of the third type of nodes include voltage deviation, three-phase imbalance, total harmonic distortion and power factor.

5. The method for deducing power quality disturbance in distribution network based on graph neural network according to claim 4 is characterized in that: Before training the graph neural network model using the node history data, the feature vector is structured, including: A time-scale alignment algorithm is used to reconstruct the time-scales of the first-category nodes, the second-category nodes, and the third-category nodes, so that the feature vectors of the above-mentioned various types of nodes can maintain a corresponding relationship at the same time.

6. The method for deducing power quality disturbance in distribution network based on graph neural network according to claim 4 is characterized in that: The graph neural network model is constructed according to the graph network structure, and the graph neural network model adopts two layers of graph convolutional neural network and one layer of fully connected network; Based on the feature vector, training the graph neural network model through the node historical data includes: Using the graph convolutional neural network of the first layer, the feature dimension of the feature vectors of the first type of nodes and the second type of nodes is expanded from 3 dimensions to 64 dimensions to achieve feature extraction; Using the graph convolutional neural network of the second layer, the feature dimension of the feature vectors of the first type of nodes and the second type of nodes is reduced from 64 dimensions to 32 dimensions, thereby achieving feature dimensionality reduction; Mapping the feature dimensions of the feature vectors of the first-category nodes and the second-category nodes to the feature dimensions of the feature vectors of the third-category nodes using the fully connected network; Divide the training set and the validation set based on the historical data of the first type of nodes, the second type of nodes, and the third type of nodes; Training the graph network model based on the training set, repeatedly calculating the loss function of the model deduction value and the label value, and gradually reducing the loss value until the loss function converges; The historical data of the first type of nodes and the second type of nodes are used as input features, and the historical data of the third type of nodes are used as verification labels.

7. The method for deducing power quality disturbance in distribution network based on graph neural network according to claim 6 is characterized in that: In the graph convolutional neural network, the node representation of each convolutional layer is updated by a propagation rule, and the expression of the propagation rule is: In the formula, H k is the node representation of the convolutional layer of the kth layer; is the self-connected adjacency matrix of the undirected graph, allowing the node features themselves to be included when updating the node representation, A is the adjacency matrix, I∈R N×N is the identity matrix; is a diagonal matrix; σ is the activation function; is a hierarchical linear transformation matrix that can be trained during the training of the graph neural network model, where F k and F k+1 They are respectively the dimensions of the node representations of the k-th convolutional layer and the k+1-th convolutional layer.

8. The method for deducing power quality disturbance in distribution network based on graph neural network according to claim 1 is characterized in that: PyQt5 is used to construct a visual interactive interface, and the user can input the source and load data in the terminal through the visual interactive interface, and perform power quality situation deduction through the graph neural network model packaged to the terminal.

Images