Power distribution network harmonic source positioning method capable of adapting to topology uncertainty

Abstract

The invention provides a power distribution network harmonic source positioning method adaptable to topology uncertainty, and the method comprises the steps: constructing a power distribution network model based on a topological structure of a power distribution network; determining a time characteristic matrix and an electrical distance connection matrix based on the power distribution network model; and inputting the time characteristic matrix and the electrical distance connection matrix into a power distribution network harmonic source positioning model, and outputting and determining a node where a harmonic source is located. According to the power distribution network harmonic source positioning model provided by the invention, the static topology constructed based on the electrical distance is combined with the dynamic topology reflecting the dynamic correlation characteristics of the nodes, and compared with a method only using the static connection relation, the problem that the positioning precision is reduced due to the topology uncertainty is effectively solved; and the accuracy and robustness of harmonic source positioning are obviously improved.

Description

Technical Field

[0001] The present application relates to the technical field of distribution networks, and in particular to a method for locating harmonic sources in distribution networks that can adapt to topology uncertainty. Background Art

[0002] With the large-scale integration of renewable energy, the number of power electronic devices in distribution networks has rapidly increased, introducing greater uncertainty and randomness. This has made system operation increasingly complex and has led to a growing problem of harmonic pollution in distribution networks. Harmonics in distribution networks can stimulate resonance, increase line losses, and compromise power quality. Therefore, accurate harmonic source location is crucial for effective harmonic pollution control.

[0003] Traditional harmonic source location methods include the equivalent circuit method, the time domain extraction method, and the harmonic state estimation method. The equivalent circuit method can only determine whether the harmonic source is located on the system side or the user side, but cannot accurately locate the node. The time domain extraction method models non-harmonic sources from a time domain perspective and uses model correlation to locate them, but it is only applicable to a single harmonic source. The harmonic state estimation method determines the harmonic source location by solving equations using harmonic phasor data. However, these methods rely on the harmonic phasor data of the PMU, and even a small phase error can seriously affect the positioning results. In addition, the above methods rely on accurate network topology and line impedance parameters for effective modeling.

[0004] In contrast, deep learning methods can autonomously extract harmonic features from monitoring data and achieve high-precision harmonic source location without relying on complex harmonic propagation models. Under the deep learning framework, harmonic source location is usually modeled as a classification problem, aiming to accurately classify nodes in the distribution network under various operating conditions to determine whether they are harmonic sources and achieve precise positioning. However, the existing mainstream graph convolutional networks (GCNs) and spatiotemporal graph convolutional networks (STGCNs) are highly dependent on predefined network topologies during training, making them extremely sensitive to topological changes. With the adjustment of distribution network operation modes and the large-scale access of new energy sources, the network topology and line parameters show significant dynamic changes, making it difficult for these methods to maintain stable and accurate positioning performance in scenarios with uncertain topology. Summary of the Invention

[0005] In order to overcome the above technical defects, the present application provides a method for locating harmonic sources in a distribution network that can adapt to topological uncertainty. To achieve the above purpose, the present application is implemented according to the following technical solutions:

[0006] This application provides a method for locating harmonic sources in a distribution network that can adapt to topology uncertainty, including:

[0007] Based on the topological structure of the distribution network, a distribution network model is constructed;

[0008] Based on the distribution network model, determining a time characteristic matrix and an electrical distance connection matrix;

[0009] The time characteristic matrix and the electrical distance connection matrix are input into the distribution network harmonic source positioning model, and the node where the harmonic source is located is determined by output.

[0010] Optionally, determining a time characteristic matrix based on the distribution network includes:

[0011] Determine the effective value of harmonic voltage and the effective value of harmonic current based on the distribution network model;

[0012] A time characteristic matrix is ​​determined based on the harmonic voltage effective value and the harmonic current effective value.

[0013] Optionally, determining an electrical distance connection matrix based on the distribution network model includes:

[0014] Based on the distribution network model, determining an impedance matrix of the distribution network;

[0015] determining an electrical distance between any two nodes of the distribution network based on the impedance matrix;

[0016] After normalizing the electrical distance between any two nodes, an electrical distance connection matrix is ​​obtained.

[0017] Optionally, the distribution network harmonic source positioning model includes a normalization module, a static-dynamic topology fusion module and a harmonic source positioning module.

[0018] Optionally, inputting the time characteristic matrix and the electrical distance connection matrix into a distribution network harmonic source positioning model and outputting a node where the harmonic source is located includes:

[0019] Input the time feature matrix and the electrical distance connection matrix into the normalization module, and output a first feature matrix and an electrical distance adjacency matrix;

[0020] Inputting the first characteristic matrix and the electrical distance adjacency matrix into the static-dynamic topology fusion module, and outputting an output node state matrix;

[0021] The output node state matrix is ​​input into the harmonic source positioning module, and the node where the harmonic source is located is output.

[0022] Optionally, the static-dynamic topology fusion module includes a dynamic topology generation module and a dynamic graph convolution loop operation module.

[0023] Optionally, inputting the first feature matrix and the electrical distance adjacency matrix into the static-dynamic topology fusion module and outputting an output node state matrix includes:

[0024] Inputting the first characteristic matrix into the dynamic topology generation module to obtain a first dynamic characteristic matrix, wherein the first dynamic characteristic includes dynamic spatiotemporal characteristics of the entire distribution network;

[0025] The first dynamic feature matrix and the electrical adjacency distance matrix are input into the dynamic graph convolution loop operation module, and output to obtain an output node state matrix.

[0026] Optionally, inputting the first feature matrix into the dynamic topology generation module to obtain a first dynamic feature matrix specifically includes:

[0027] Based on the first characteristic matrix, a node dynamic basic characteristic matrix is ​​obtained;

[0028] Based on the node dynamic basic feature matrix, a first dynamic feature matrix is ​​obtained.

[0029] Optionally, the step of inputting the first dynamic feature matrix and the electrical adjacency distance matrix into the dynamic graph convolution loop operation module and outputting an output node state matrix includes:

[0030] Performing a graph convolution operation on the first dynamic feature matrix and the electrical adjacency distance matrix to obtain a dynamic tensor;

[0031] generating a dynamic graph embedding based on the dynamic tensor;

[0032] constructing a dynamic electrical adjacency distance matrix based on the dynamic graph embedding;

[0033] determining input node characteristics based on the dynamic electrical adjacency distance matrix and the electrical adjacency distance matrix;

[0034] An output node state matrix is ​​determined based on the dynamic electrical adjacency distance matrix, the electrical adjacency distance matrix, and the input node features.

[0035] Optionally, inputting the output node state matrix into the harmonic source positioning module and outputting a node where the harmonic source is located includes:

[0036] Determining a harmonic source prediction value matrix based on the output node state matrix;

[0037] Determining a prediction label vector matrix based on the harmonic source prediction value matrix;

[0038] Determining a predicted binary variable matrix based on the predicted label vector matrix;

[0039] Determining whether a node in the predicted binary variable matrix exceeds a preset threshold;

[0040] If it exceeds, the node that exceeds the preset threshold will be determined as the node where the harmonic source is located.

[0041] This application has the following beneficial effects:

[0042] The distribution network harmonic source positioning model proposed in this application combines the static topology constructed based on electrical distance with the dynamic topology that reflects the dynamic association characteristics of nodes. Compared with the method that only uses static connection relationships, it effectively overcomes the problem of reduced positioning accuracy caused by topological uncertainty and significantly improves the accuracy and robustness of harmonic source positioning.

[0043] In addition to the above-described purposes, features and advantages, the present application has other purposes, features and advantages. The present application will be further described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of this application. The exemplary embodiments and descriptions of this application are intended to explain this application and do not constitute an improper limitation on this application. In the accompanying drawings:

[0045] Figure 1 This is a flow chart of a method for locating harmonic sources in a distribution network that is adaptable to topology uncertainty, provided in an embodiment of the present application;

[0046] Figure 2 It is a structural diagram of the improved IEEE 13-node test system for experimental simulation provided in an embodiment of the present application. DETAILED DESCRIPTION

[0047] The embodiments of the present application are described in detail below with reference to the accompanying drawings, but the present application can be implemented in many different ways as defined and covered by the claims.

[0048] In order to solve the problems raised by the background technology, such as Figure 1 As shown, the present application provides a method for locating harmonic sources in a distribution network that can adapt to topology uncertainty, including:

[0049] Step S101: constructing a distribution network model based on the topological structure of the distribution network;

[0050] In order to effectively model the temporal dynamic characteristics and spatial dependencies of harmonic voltages and currents, this application regards the power sources, loads, etc. in the distribution network as nodes in the graph, and the power lines as edges connecting the nodes, thereby abstracting the topological structure of the distribution network into an undirected graph model. In order to fully explore its spatial information, instead of treating the voltages of each node and the currents of each branch as independent discrete variables, a generalized graph structure is introduced to model the distribution network, thereby obtaining a distribution network model. Since the operating status of each node is not only affected by its own current moment, but also has a complex spatiotemporal coupling relationship with the status of its adjacent nodes in adjacent time segments, constructing a reasonable graph structure is crucial to achieving accurate harmonic source positioning.

[0051] While introducing more harmonic features can improve location accuracy in distribution network harmonic source location algorithms, the massive amount of data transmitted from edge power quality monitoring equipment to the grid company's monitoring and control system inevitably creates significant communication and storage pressures, a problem that is generally overlooked in current research. Therefore, to achieve efficient and lightweight deployment of harmonic source location algorithms in practical power quality monitoring platforms, the simplicity of input features should be a key consideration. Prioritizing the selection of a small number of representative features while ensuring location accuracy while minimizing the use of data transmission and storage resources, thereby improving the algorithm's engineering practicality and scalability.

[0052] Step S102: determining a time characteristic matrix and an electrical adjacency distance matrix based on the distribution network model;

[0053] After the distribution network model is constructed, the time characteristic matrix and electrical adjacency distance matrix can be determined based on the distribution network model. The specific process is as follows:

[0054] From a temporal perspective, given the asynchrony of harmonic monitoring data, this application selects the effective value of harmonic voltage and the effective value of harmonic current as core indicators based on the distribution network model, and the calculation method is the average effective value per minute. Since harmonic voltage and harmonic current are the basis of harmonic-related quantities such as total harmonic voltage distortion, total harmonic current distortion, and harmonic power, using their effective values ​​as input can reconstruct harmonic-related quantity parameters such as total harmonic voltage distortion, while reducing the dimensional redundancy brought about by multi-feature extraction in traditional methods. The features used retain the physical essence of harmonic propagation and can effectively represent multidimensional harmonic information.

[0055] Instantaneous values ​​of harmonic voltage and current v h (t) and i h (t) will change with time, but the effective value takes into account the energy distribution in the entire cycle and can stably describe the influence of the hth harmonic. It is more stable and reliable to use it as an input feature. At time t, the time feature matrix X of the time dimension is t It can be expressed as

[0056]

[0057] Where N represents the number of nodes in the distribution network, V ht N , I ht N It represents the average value of the effective value of harmonic voltage and harmonic current at a certain node.

[0058] From a spatial perspective, the distribution network can be viewed as a graph structure. The state of each node in the distribution network is constantly affected by adjacent nodes and more distant nodes. The connection relationship between nodes reflects its spatial characteristics. This application uses the electrical adjacency distance matrix instead of the adjacency matrix of the traditional graph neural network as the input feature of the spatial dimension. The calculation method of the electrical distance connection matrix A0 is as follows:

[0059] First, according to the distribution network model, the impedance matrix Z of the distribution network is calculated D

[0060]

[0061] Among them, Z ii and Z jj is the self-impedance of nodes i and j, Z ij and Z ji is the mutual impedance between nodes i and j.

[0062] Then, by calculating the modulus of the complex number based on the above impedance matrix, the electrical distance d between any two nodes i and j in the distribution network is obtained. ij :

[0063] d ij =|Z eqij |=|Z ii +Z jj -Z ij -Z ji |(1≤i,j≤N) (3)

[0064] Among them, || represents the modular operation, Z eqij It is plural.

[0065] Right ij Perform normalization and get a ij .

[0066] a ij =d ij / max(d ij ) (4)

[0067] Finally, the electrical distance connection matrix A0 reflecting the electrical distance in the spatial dimension can be obtained:

[0068]

[0069] Step S103: input the time characteristic matrix and the electrical distance connection matrix into the distribution network harmonic source positioning model, and output the node where the harmonic source is located.

[0070] The distribution network harmonic source location model builds on the original DGCRN by introducing a static-dynamic topology fusion module, making it more suitable for harmonic source location tasks. This model can adapt to the uncertainty of distribution network topology and achieve higher-precision harmonic source location.

[0071] The distribution network harmonic source location model includes a normalization module, a static-dynamic topology fusion module, and a harmonic source location module. During the neural network model training, when a node has a harmonic source and transmits harmonic current to the distribution network, the node's label is marked as 1, and the remaining nodes are marked as 0. At this point, the harmonic source location problem is converted into a graph classification problem based on DGCRN, that is, the time feature matrix X is converted into a graph classification problem based on DGCRN. t The electrical distance connection matrix A0 is input into the trained DGCRN harmonic source location model to output the nodes where harmonic sources exist.

[0072] The following is a detailed description of the specific processing process of the time characteristic matrix and the electrical connection distance matrix in the distribution network harmonic source location model:

[0073] The time feature matrix X t Input the normalization module and output the first feature matrix. The normalization process of the time feature matrix can be expressed as:

[0074]

[0075] Where x′ i is the element in the normalized first characteristic matrix, x i is the time feature matrix X t The elements in x imax and x imin are the maximum and minimum values ​​of the i-th time feature.

[0076] The electrical distance connection matrix A0 is input into the normalization module, and the output is the electrical distance adjacency matrix A (also called the static electrical distance adjacency matrix). The process of normalizing the electrical distance connection matrix can be expressed as:

[0077] A=A0+I (7)

[0078] Where I is the identity matrix.

[0079] In order to capture the complex spatiotemporal relationships in distribution network harmonic data, the static-dynamic topology fusion module includes a dynamic topology generation module and a dynamic graph convolution loop operation module.

[0080] The first characteristic matrix and the electrical distance adjacency matrix are input into the static-dynamic topology fusion module, and the output node state matrix is ​​obtained. The specific processing process is as follows:

[0081] Input the first feature matrix into the dynamic topology generation module to obtain the first dynamic feature matrix I sa The first dynamic feature matrix contains the dynamic spatiotemporal features of the entire distribution network. The first feature is input into the dynamic topology generation module to first generate the node dynamic basic feature matrix I t ∈R B×N×D , D represents the feature dimension 2, B represents the batch size, and N is the number of nodes. Node dynamic basic feature matrix I t ∈R B×N×D It is composed of time characteristics, time T t and hidden state H t-1 To strengthen the modeling of node relationships, this application introduces a self-attention mechanism, which enables each node to pay attention to other nodes in the entire topology. We project the signal matrix into the Q, K, and V matrices through linear transformation. It can be expressed as Q = I t W Q , K=I t W K , V=I t W V , where W Q , W K , W V ∈R D×d is a learnable transformation matrix, and d is a hyperparameter. The formula for implementing the self-attention mechanism is as follows:

[0082]

[0083] Among them, I sa ∈R B×N×D Represents the output of the self-attention mechanism.

[0084] Through formula (8), the node dynamic basic feature matrix is ​​transformed into the first dynamic basic matrix I through the self-attention mechanism sa .

[0085] After obtaining the first dynamic basic matrix, the first dynamic basic matrix and the electrical adjacency distance matrix are input into the dynamic graph convolution loop operation module, and the output node state matrix is ​​obtained. The specific operations of the first dynamic basic matrix and the electrical adjacency distance matrix in the dynamic graph convolution loop operation module are as follows:

[0086] The electrical distance adjacency matrix A and the first dynamic characteristic matrix I sa Perform graph convolution operation to obtain dynamic tensor F t .

[0087] F t =A·I sa (9)

[0088] Then according to the dynamic tensor F t and two dynamic static node embeddings with learnable t 1 , e t 2 ∈R N×D Perform element-wise multiplication to generate dynamic graph embedding E t , the operation can be expressed as

[0089]

[0090] Then, based on the dynamic graph embedding, the dynamic electrical adjacency distance matrix DA is constructed t , when constructing the dynamic electrical adjacency distance matrix, we first embed the node dynamic topology into E t 1 ,E t 2 and their transpose E t 1T ,E t 2T Multiply to calculate the similarity between nodes. Then, to ensure that the values ​​in the matrix are non-negative, we apply the ReLU function for nonlinear transformation. This process can be expressed as

[0091]

[0092] Among them, α is a parameter to prevent overfitting in nonlinear transformation, DA t ∈R N×N represents the dynamic adjacency matrix at time t.

[0093] It should be noted that the static electrical distance adjacency matrix A and the dynamic electrical adjacency distance matrix DA t , which reflects the harmonic propagation characteristics between nodes from different angles.

[0094] Then, the input node feature H is determined based on the dynamic electrical adjacency distance matrix and the electrical adjacency distance matrix. in In order to better simulate the spatiotemporal correlation of harmonic propagation in the distribution network, the dynamic graph convolution loop operation converts the input node feature H in , electrical distance adjacency matrix A and dynamic adjacency distance matrix DA tBy dynamically adjusting the weight parameters α, β and γ (where α+β+γ=1), a k-layer convolution feature is formed to improve the subsequent positioning performance. This process can be expressed as

[0095] H (k) =αH in +βH (k-1) D t (DA t +I)+γH (k-1) A (13)

[0096]

[0097] Among them, D t is the degree matrix, I is the identity matrix, W p (k) is the parameter matrix, K is the propagation depth, H out ∈R B×N×1 Represents the output node state matrix.

[0098] It should be noted that when only directed graph convolution is considered, in order to make full use of the directed graph, its representation method is as follows:

[0099] H out =Θ *G (H in ,DA t ,A) (15)

[0100] Among them, Θ *G represents graph convolution, and Θ represents learnable parameters.

[0101] In order to determine the possibility of harmonic sources at each node, a harmonic source localization module with a fully connected layer, an activation function and a binary classification layer is designed. out After that, the dimension is reduced through the fully connected layer and converted into the harmonic source prediction value matrix y∈R B×N Then, the activation function Sigmoid is applied to map the harmonic source prediction value matrix y to the range of 0 to 1, and the prediction label vector matrix Y0 is obtained.

[0102] Since harmonic source location is a binary classification problem, it is necessary to determine whether each node contains a harmonic source based on the monitoring data. Therefore, the binary classification layer converts the predicted label vector matrix Y0 into a predicted binary variable matrix: Then, it is judged whether the node output in the predicted binary variable matrix exceeds the preset threshold. If it exceeds, the node is identified as containing a harmonic source and is determined to be the node where the harmonic source is located, and the output is adjusted to 1; otherwise, the node is determined to not contain a harmonic source, and the output is adjusted to 0. The expression is

[0103]

[0104] It should be noted that the preset threshold is set according to actual needs, and this application does not limit the specific numerical range of the preset threshold.

[0105] Experimental simulation

[0106] This embodiment takes the improved IEEE 13-node test system as an example, performs corresponding numerical simulation on the MATLAB / Simulink platform, and uses MATLAB to process and analyze the simulation data. Figure 2 shown.

[0107] To more realistically simulate the distribution network, samples were generated by varying the number, location, magnitude of injected harmonic current, and load level of harmonic loads. Specifically, the number of harmonic loads varied from 0 to 3, with all harmonic loads randomly placed on all possible lines. For each harmonic load, the emitted harmonic current fluctuates with the load power and may contain the most common harmonic components found in real power grids, including the 5th, 7th, 11th, and 13th harmonics. The phase angles of these harmonic components were randomly selected within the range of 0° to 360° to simulate different operating conditions. Furthermore, the load power parameter error was set to ±20% to account for potential load power fluctuations and ensure a more realistic reflection of actual grid operation.

[0108] In the environment, harmonic currents are injected into the substation to simulate background harmonics. At the same time, only small harmonic currents are injected into the bus with linear loads. The average value of the harmonic voltage and current RMS values ​​is calculated every minute throughout the simulation.

[0109] The sample considers changes in the distribution network topology by introducing one or two disconnected lines, such as the disconnection between nodes 8 and 13, to simulate topological changes. This disconnected line condition effectively simulates topological uncertainties in real distribution networks caused by factors such as renewable energy islands, line failures, and distributed renewable energy scheduling.

[0110] When dividing the data, 60% of the data is used as the training data set, 20% of the data is used as the validation data set, and 20% of the data is used as the test data set. The hyperparameters of the model are set as follows: the learning rate is 0.001 and the batch size is 16.

[0111] The positioning accuracy is shown in Table 1.

[0112] Table 1 Comparison of model accuracy

[0113]

[0114]

[0115] As can be seen from Table 1, the method proposed in this application can accurately locate the multi-harmonic source when the distribution network topology changes, and the positioning accuracy exceeds 95%.

[0116] In summary, the method proposed in this application uses the effective values ​​of harmonic voltage and harmonic current as model input features. This type of low-dimensional feature not only has a good harmonic propagation characterization capability, but is also easy to obtain and process, which helps to improve the generalization ability and positioning accuracy of the model. At the same time, the method proposed in this application only relies on stable and collectible effective value data, which significantly reduces the dependence on harmonic phase information and alleviates the requirements for data synchronization and high-bandwidth communication. As well as the static-dynamic topology fusion module in the distribution network harmonic source positioning model proposed in this application, it combines the static topology constructed based on electrical distance with the dynamic topology that reflects the dynamic association characteristics of the nodes. Compared with the method that only uses static connection relationships, it effectively overcomes the problem of reduced positioning accuracy caused by topological uncertainty and significantly improves the accuracy and robustness of harmonic source positioning.

[0117] The above description is merely a preferred embodiment of the present application and is not intended to limit the present application. Various modifications and variations are possible for those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present application shall be included within the scope of protection of the present application.

Claims

1. A method for locating harmonic sources in a distribution network that can adapt to topology uncertainty, characterized in that: include: Based on the topological structure of the distribution network, a distribution network model is constructed; Based on the distribution network model, determining a time characteristic matrix and an electrical distance connection matrix; The time characteristic matrix and the electrical distance connection matrix are input into the distribution network harmonic source positioning model, and the node where the harmonic source is located is determined by output.

2. The method according to claim 1, characterized in that The determining of a time characteristic matrix based on the distribution network includes: Determine the effective value of harmonic voltage and the effective value of harmonic current based on the distribution network model; A time characteristic matrix is ​​determined based on the harmonic voltage effective value and the harmonic current effective value.

3. The method according to claim 1, characterized in that The determining of the electrical distance connection matrix based on the distribution network model includes: Based on the distribution network model, determining an impedance matrix of the distribution network; determining an electrical distance between any two nodes of the distribution network based on the impedance matrix; After normalizing the electrical distance between any two nodes, an electrical distance connection matrix is ​​obtained.

4. The method according to claim 1, wherein The distribution network harmonic source positioning model includes a normalization module, a static-dynamic topology fusion module and a harmonic source positioning module.

5. The method according to claim 4, characterized in that The step of inputting the time characteristic matrix and the electrical distance connection matrix into a distribution network harmonic source positioning model and outputting a node where the harmonic source is located includes: Input the time feature matrix and the electrical distance connection matrix into the normalization module, and output a first feature matrix and an electrical distance adjacency matrix; Inputting the first characteristic matrix and the electrical distance adjacency matrix into the static-dynamic topology fusion module, and outputting an output node state matrix; The output node state matrix is ​​input into the harmonic source positioning module, and the node where the harmonic source is located is output.

6. The method according to claim 5, characterized in that The static-dynamic topology fusion module includes a dynamic topology generation module and a dynamic graph convolution cycle operation module.

7. The method according to claim 6, characterized in that The step of inputting the first characteristic matrix and the electrical distance adjacency matrix into the static-dynamic topology fusion module and outputting an output node state matrix includes: Inputting the first characteristic matrix into the dynamic topology generation module to obtain a first dynamic characteristic matrix, wherein the first dynamic characteristic includes dynamic spatiotemporal characteristics of the entire distribution network; The first dynamic feature matrix and the electrical adjacency distance matrix are input into the dynamic graph convolution loop operation module, and output to obtain an output node state matrix.

8. The method according to claim 7, characterized in that Inputting the first feature matrix into the dynamic topology generation module to obtain a first dynamic feature matrix specifically includes: Based on the first characteristic matrix, a node dynamic basic characteristic matrix is ​​obtained; Based on the node dynamic basic feature matrix, a first dynamic feature matrix is ​​obtained.

9. The method according to claim 7, characterized in that The step of inputting the first dynamic feature matrix and the electrical adjacency distance matrix into the dynamic graph convolution cycle operation module and outputting an output node state matrix includes: Performing a graph convolution operation on the first dynamic feature matrix and the electrical adjacency distance matrix to obtain a dynamic tensor; generating a dynamic graph embedding based on the dynamic tensor; constructing a dynamic electrical adjacency distance matrix based on the dynamic graph embedding; determining input node characteristics based on the dynamic electrical adjacency distance matrix and the electrical adjacency distance matrix; An output node state matrix is ​​determined based on the dynamic electrical adjacency distance matrix, the electrical adjacency distance matrix, and the input node features.

10. The method according to claim 5, characterized in that Inputting the output node state matrix into the harmonic source positioning module and outputting the node where the harmonic source is located, including: Determining a harmonic source prediction value matrix based on the output node state matrix; Determining a prediction label vector matrix based on the harmonic source prediction value matrix; Determining a predicted binary variable matrix based on the predicted label vector matrix; Determining whether a node in the predicted binary variable matrix exceeds a preset threshold; If it exceeds, the node that exceeds the preset threshold will be determined as the node where the harmonic source is located.

Images