Method and system for measuring and calculating leakage magnetic field of converter transformer based on GNN-TCN

Abstract

The invention discloses a method and a system for measuring and calculating a leakage magnetic field of a converter transformer based on GNN-TCN. The method comprises the following steps: establishing a graph neural network for extracting spatial distribution characteristics of the leakage magnetic field; the node attributes of the graph neural network comprise geometric structure attributes and electromagnetic state attributes; establishing a time convolution network for extracting the time sequence characteristics of the leakage magnetic field by adopting a dynamic expansion convolution structure; sequentially connecting the graph neural network with the time convolutional network to form a branch of a neural network layer; connecting the dual-branch neural network layer with the feature fusion layer, and establishing a GNN-TCN model; acquiring sample data, and training the model by adopting the sample data; and inputting real-time operation data of the converter transformer into the trained model, and measuring and calculating to obtain the leakage magnetic field of the converter transformer. According to the embodiment of the invention, by fusing the graph neural network and the time convolutional network, the space and time sequence characteristics of the leakage magnetic field can be efficiently and accurately captured, and the calculation efficiency and precision in a non-linear scene are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electrical engineering, and particularly relates to a GNN-TCN-based converter transformer leakage magnetic field calculation method and system. BACKGROUND

[0002] In the operation process of the converter transformer, the DC component generated by the valve side and the AC network side coil can cause DC bias. At the same time, the load current contains a large number of high harmonic components. These high harmonic currents can excite a distributed leakage magnetic field with uneven distribution in the converter transformer, and further generate harmonic loss in the winding and structural parts, and even may cause local overheating problems. The above-mentioned conditions can pose a serious threat to the load capacity and operation reliability of the converter transformer.

[0003] The traditional leakage magnetic field calculation method generally adopts a finite element simulation model. The finite element simulation model discretizes the continuous field problem into a linear equation set, and then solves it through an iterative algorithm. On the one hand, due to the insufficient expression ability of the value discretization method for physical nonlinear characteristics, the finite element method has defects in nonlinear processing. On the other hand, for large-scale problems, the number of iterations will increase exponentially, and the calculation efficiency of the finite element method will decrease significantly. SUMMARY

[0004] The purpose of the embodiment of the present application is to provide a GNN-TCN-based converter transformer leakage magnetic field calculation method and system. By fusing a graph neural network and a time convolution network, the spatial and time sequence characteristics of the leakage magnetic field can be efficiently and accurately captured, and the calculation efficiency and accuracy in a nonlinear scene can be improved.

[0005] The embodiment of the present application provides a GNN-TCN-based converter transformer leakage magnetic field calculation method, comprising: establishing a graph neural network for extracting the spatial distribution characteristics of the leakage magnetic field; the node attributes of the graph neural network include geometric structure attributes and electromagnetic state attributes; establishing a time convolution network for extracting the time sequence characteristics of the leakage magnetic field by using a dynamic expansion convolution structure; sequentially connecting the graph neural network and the time convolution network to form one branch of a neural network layer; connecting the double-branch neural network layer and a feature fusion layer to establish a GNN-TCN model; the double branches of the neural network layer are respectively a current branch and a voltage branch; obtaining sample data, and training the GNN-TCN model by using the sample data; inputting the real-time operation data of the converter transformer into the trained GNN-TCN model to calculate the leakage magnetic field of the converter transformer.

[0006] As an improvement of the above scheme, the graph neural network for extracting the spatial distribution characteristics of the magnetic leakage field comprises: According to the geometric structure and electromagnetic information of the physical entity in the converter transformer, the node and the node attribute are obtained; the node attribute includes the geometric structure attribute and the electromagnetic state attribute; According to the connection relationship and electromagnetic interaction relationship of the physical entity, the edge connecting each node is constructed; the edge includes the physical connection edge and the electromagnetic interaction edge; According to the law of electromagnetic induction, the edge weight control function is established, and the weight of the edge is adaptively updated by using the edge weight control function; According to the node, the node attribute and the edge, a pyramid multi-scale graph neural network is established to extract the spatial distribution characteristics of the magnetic leakage field.

[0007] As an improvement of the above scheme, the node and the node attribute are obtained according to the geometric structure and electromagnetic information of the physical entity in the converter transformer, which comprises: The physical entity in the converter transformer is mapped to the node in the graph structure; According to the geometric structure of the physical entity in the converter transformer, the geometric structure attribute of the node is obtained; According to the electromagnetic information of the physical entity in the converter transformer, the electromagnetic state attribute of the node is obtained; the electromagnetic information includes current excitation, voltage state and magnetic flux density.

[0008] As an improvement of the above scheme, according to the law of electromagnetic induction, the edge weight control function is established, and the weight of the edge is adaptively updated by using the edge weight control function, which comprises: According to the law of electromagnetic induction, the influence intensity of the magnetic flux change on the neighbor node is calculated, and the edge weight control function is established; According to the edge weight control function, the edge weight control factor is established, and in the process of graph convolution, the information aggregation weight of the neighbor node is dynamically adjusted through the edge weight control factor, so as to adaptively update the weight of the edge.

[0009] As an improvement of the above scheme, according to the node, the node attribute and the edge, a pyramid multi-scale graph neural network is established to extract the spatial distribution characteristics of the magnetic leakage field, which comprises: According to the node, the node attribute and the edge, the graph structure is obtained; Using graph pooling algorithm, the bottom layer graph structure is gradually compressed to generate the upper layer graph structure, and the top layer graph structure is generated; Interlayer cross-layer jump connection is established to realize multi-scale feature transmission; After obtaining the global feature of the top layer graph structure, the global feature is reconstructed by using the graph sampling algorithm to obtain the spatial distribution characteristics.

[0010] As an improvement of the above scheme, the time convolution network for extracting the time sequence characteristics of the magnetic leakage field is established by using the dynamic dilated convolution structure, comprising: A plurality of channels are established in parallel, each channel comprising a multi-layer convolution network for extracting the time sequence characteristics of the magnetic leakage field, and a cross-layer residual connection is used between the network layers of the multi-layer convolution network; The convolution kernel of the channel and the network layer is adaptively set by using the dynamic dilated convolution structure; The top layer output of the multi-layer convolution network of the multi-channel is spliced for feature, and the time convolution network for extracting the time sequence characteristics of the magnetic leakage field.

[0011] As an improvement of the above scheme, the convolution kernel of the channel and the network layer is adaptively set by using the dynamic dilated convolution structure, comprising: According to the preset dilated rate, a cavity is injected in the convolution kernel of the network layer to obtain a traditional dilated convolution structure; An adaptive weight module is added at each time step of the traditional dilated convolution structure to form a dynamic dilated convolution structure; the adaptive weight module adaptively adjusts the dilated rate according to the feature complexity of the input data; The dynamic dilated convolution structure is introduced in each channel and each network layer of the multi-layer convolution network to adaptively adjust the convolution kernel.

[0012] As an improvement of the above scheme, the graph neural network and the time convolution network are sequentially connected to form one branch of the neural network layer, comprising: A space-time cross attention mechanism is introduced into the graph neural network to embed the response of the historical time step in the node to form a time sequence enhanced graph neural network; The time sequence enhanced graph neural network and the time convolution network are sequentially connected to form one branch of the neural network layer.

[0013] As an improvement of the above scheme, the sample data is obtained, and the GNN-TCN model is trained by using the sample data, comprising: A space synchronization mechanism based on geometric mapping is constructed to convert the heterogeneous space coordinate systems of different sampling nodes into a unified coordinate system; The sampling interval error of the sampling nodes in different time periods is eliminated by resampling; The time sequence electrical data and the corresponding magnetic flux density data of the sampling nodes are obtained as sample data; the time sequence electrical data includes time sequence current data and time sequence voltage data; The time sequence electrical data is input into the GNN-TCN model, and the space-time features of the time sequence current data and the time sequence voltage data are extracted by the double branches of the GNN-TCN model respectively, so as to output the magnetic flux density data as a target, and the GNN-TCN model is trained.

[0014] The application also provides a GNN-TCN-based converter transformer leakage magnetic field measuring and calculating system, comprising: A graph neural network construction module is configured to construct a graph neural network for extracting spatial distribution features of the leakage magnetic field, wherein the node attributes of the graph neural network comprise geometric structure attributes and electromagnetic state attributes. A time convolution network construction module is configured to adopt a dynamic dilated convolution structure to construct a time convolution network for extracting time sequence features of the leakage magnetic field. A branch generation module is configured to sequentially connect the graph neural network and the time convolution network to form a branch of a neural network layer. A model construction module is configured to connect the double-branch neural network layer and a feature fusion layer to construct a GNN-TCN model, wherein the double branches of the neural network layer are a current branch and a voltage branch. A model training module is configured to obtain sample data and train the GNN-TCN model by using the sample data. A model application module is configured to input real-time operation data of the converter transformer into the trained GNN-TCN model to measure and calculate the leakage magnetic field of the converter transformer.

[0015] Compared with the prior art, the GNN-TCN-based converter transformer leakage magnetic field measuring and calculating method and system can efficiently and accurately capture the spatial and time sequence features of the leakage magnetic field by fusing the graph neural network and the time convolution network, and improve the calculation efficiency and accuracy in a nonlinear scenario. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1is a step flow schematic diagram of a GNN-TCN-based converter transformer leakage magnetic field measurement method provided by the embodiment of the application. Figure 2 is a structural schematic diagram of a graph neural network provided by the embodiment of the application. Figure 3 is a schematic diagram of self-feature updating before and after neighborhood interaction in a graph neural network provided by the embodiment of the application. Figure 4 is a structural schematic diagram of a time convolution network provided by the embodiment of the application. Figure 5 is a structural schematic diagram of a GNN-TCN-based converter transformer leakage magnetic field measurement system provided by the embodiment of the application. DETAILED DESCRIPTION

[0017] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.

[0018] In the description and claims of the specification, it is to be understood that the terms first, second, etc. are used only for the purpose of description and are not to be construed as indicating or implying relative importance or an indicated number of technical features. They are not necessarily used in the order or time sequence described. Where appropriate, terms are interchangeable. Therefore, the features defined with "first", "second" can explicitly or implicitly include at least one of the features.

[0019] The embodiment of the application provides a GNN-TCN-based converter transformer leakage magnetic field measurement method. Figure 1 In the embodiment, the GNN-TCN-based converter transformer leakage magnetic field measurement method is specifically executed through steps S1 to S6. S1, a graph neural network for extracting spatial distribution characteristics of the leakage magnetic field is established; the node attributes of the graph neural network include geometric structure attributes and electromagnetic state attributes.

[0020] In the embodiment of the present application, physical perception with nonlinear correlation is fused in the node attribute of a graph neural network (GNN), and the complex geometric structure and electromagnetic relationship of the converter transformer are abstracted into a graph structure, so that the graph convolution is more in line with the actual physical mechanism. Through the powerful representation learning ability of the graph neural network, the distribution law and interaction relationship of the magnetic leakage field in the spatial dimension can be comprehensively and accurately described, and the extraction accuracy of the spatial features of the magnetic leakage field is effectively improved.

[0021] S2, a dynamic dilated convolution structure is used to establish a time convolution network for extracting time sequence features of the magnetic leakage field.

[0022] It should be noted that the time convolution network (TCN) established by using the dynamic dilated convolution structure can adaptively adjust the dilated rate of the convolution kernel according to the data characteristics, expand the receptive field without increasing too much calculation amount, automatically identify the short-term and long-term dependence relationship, and more accurately capture the dynamic change trend of the voltage and current data under different operating conditions of the converter transformer, and mine the hidden features in the time sequence data.

[0023] S3, the graph neural network and the time convolution network are sequentially connected to form a branch of the neural network layer.

[0024] The high-dimensional spatial features are extracted by the graph neural network, and to avoid the loss of time sequence information, the output of the graph neural network is directly input into the time convolution network to capture the forward and backward dependence relationship in the time dimension. By establishing a unified feature extraction channel, the preliminary fusion of the nonlinear relationship in the spatial and time dimensions can be realized, and the isolation of information in different dimensions can be avoided.

[0025] S4, the dual-branch neural network layer is connected with a feature fusion layer to establish a GNN-TCN model; the two branches of the neural network layer are a current branch and a voltage branch.

[0026] The neural network layer processes the current and voltage through the current branch and the voltage branch respectively, fully considers the influence of the voltage and current on the magnetic leakage field, and then decodes and maps the distribution of the magnetic leakage field through the feature fusion layer, so as to comprehensively reflect the real nonlinear relationship among the converter transformer, the current and the voltage during the operation of the converter transformer.

[0027] S5, sample data is obtained, and the GNN-TCN model is trained by using the sample data.

[0028] In some preferred embodiments, the sample data is obtained from historical operation data; in other preferred embodiments, the sample data is obtained from the results of simulation experiments.

[0029] Exemplarily, a finite element simulation method is adopted, a high-performance computing platform is used for systematic parameterization scanning of external circuit input parameters, simulation data covering multiple nonlinear working conditions are obtained, and leakage magnetic field information at different spatial coordinates and different time points is obtained through simulation as sample data.

[0030] S6, inputting real-time operation data of the converter transformer into the trained GNN-TCN model to measure the leakage magnetic field of the converter transformer.

[0031] In the above scheme, the graph neural network is applied to leakage magnetic field spatial feature extraction through geometric properties and electromagnetic state properties, and the time convolution network with dynamic dilated convolution structure is combined to capture time features, which can deeply mine the leakage magnetic field distribution law under the complex spatial structure and time sequence change of the converter transformer. By constructing a double-branch GNN-TCN model, the leakage magnetic field features are efficiently extracted from multiple dimensions and multiple physical quantities. The GNN-TCN model trained and put into use in sequence can accurately capture the nonlinear relationship between the transformer port voltage, current and internal leakage magnetic field, and thus significantly improve the accuracy and efficiency of leakage magnetic field measurement in a complex electromagnetic environment.

[0032] As a preferred embodiment, step S1, establishing a graph neural network for extracting spatial distribution features of the leakage magnetic field, is specifically performed through steps S11-S14: S11, obtaining nodes and node attributes according to the geometric structure and electromagnetic information of physical entities in the converter transformer; the node attributes include geometric structure attributes and electromagnetic state attributes; S12, constructing edges connecting nodes according to the connection relationship and electromagnetic interaction relationship of the physical entities; the edges include physical connection edges and electromagnetic interaction edges; S13, establishing an edge weight control function according to the electromagnetic induction law, and adaptively updating the weight of the edge using the edge weight control function; S14, establishing a pyramid multi-scale graph neural network according to the nodes, the node attributes and the edges, to extract spatial distribution features of the leakage magnetic field.

[0033] The graph neural network is a kind of neural network specially used for processing structural data, which can effectively capture the features of nodes, edges and their local neighborhood structure, and generate embedded representations of nodes, edges or entire graphs. The graph neural network can process structural data, capture complex relationships between nodes, and flexibly adapt to multiple tasks such as node classification, link prediction and graph classification. The structure of the graph neural network GNN in the preferred embodiment of the present application is as shown in Figure 2 The structure includes an input layer, a hidden layer and an output layer.

[0034] The core of the graph neural network is a message passing mechanism, that is, each node updates its feature representation by aggregating the features of itself and its neighbor nodes. The process of neighborhood interaction in the message passing mechanism is shown in FIG. 1. Figure 3 The feature update is realized by performing multiple rounds of graph convolution.

[0035] In the embodiments of the present application, the task of the graph neural network is to extract the spatial distribution features of the magnetic leakage field. Naturally, the nodes, attributes and edges of the graph neural network are closely related to the network task.

[0036] In some preferred embodiments, the physical entity includes a coil winding, a core; the geometric structure includes the size, shape and relative position of the physical entity; the electromagnetic information includes the current density and magnetic permeability of the physical entity.

[0037] Further, preferably, in step S11, the nodes and node attributes are obtained according to the geometric structure and electromagnetic information of the physical entity in the converter transformer, including: mapping the physical entity in the converter transformer to a node in the graph structure; obtaining the geometric structure attribute of the node according to the geometric structure of the physical entity in the converter transformer; obtaining the electromagnetic state attribute of the node according to the electromagnetic information of the physical entity in the converter transformer; the electromagnetic information includes current excitation, voltage state and magnetic flux density.

[0038] Node setting is a digital mapping of the physical entity of the converter transformer. By converting the actual physical objects such as windings and cores into nodes in the graph structure, the model can focus on the influence of these entities on the magnetic leakage field; the node attributes further enrich the connotation of the nodes. The geometric structure attribute describes the shape, size and relative position of the physical entity from the spatial dimension, and the electromagnetic state attribute describes the operating characteristics of the entity from the electromagnetic perspective. These attributes provide specific information for the model to analyze the correlation between the physical entity and the magnetic leakage field.

[0039] In the embodiments of the present application, the setting of the edge also serves the task of calculating the magnetic leakage field. In step S12, the correlation between nodes is reflected by the physical connection edge and the electromagnetic interaction edge, wherein the physical connection edge is the actual physical connection relationship, and the electromagnetic interaction edge emphasizes the electromagnetic action relationship between entities.

[0040] For example, the current passes through the winding to generate a magnetic field, which further electromagnetically couples with other windings, so that there is an electromagnetic interaction edge between the two windings.

[0041] Preferably, in step S13, an edge weight control function is established according to the electromagnetic induction law, and the edge weight control function is used to adaptively update the weight of the edge, including: According to the electromagnetic induction law, the influence intensity of the magnetic flux change on the neighbor node is calculated, and an edge weight control function is established; According to the edge weight control function, an edge weight control factor is established, and in the graph convolution process, the information aggregation weight of the neighbor node is dynamically adjusted through the edge weight control factor to adaptively update the weight of the edge.

[0042] Considering that the material permeability of different physical entities in the converter transformer may be different, and the size and position of the physical entities in different converter transformers may also be different, resulting in differences in magnetic coupling. In the embodiment of the present application, the weight of the electromagnetic interaction edge is mainly updated.

[0043] Exemplarily, the edge weight control function is input into the sigmoid activation function to establish the edge weight control factor. Then, in the process of neighborhood interaction as shown in Figure 3 , the information aggregation weight of the neighbor node is optimized.

[0044] Preferably, in step S14, a pyramid multi-scale graph neural network is established according to the node, the node attribute and the edge to extract the spatial distribution feature of the magnetic leakage field, comprising: According to the node, the node attribute and the edge, a graph structure is obtained; A graph pooling algorithm is used to start from the bottom graph structure and gradually compress upwards to generate the upper graph structure until the top graph structure is generated; Cross-layer jump connections are established between layers to perform multi-scale feature transmission; After obtaining the global feature of the top graph structure, the global feature is reconstructed by a graph sampling algorithm to obtain the spatial distribution feature.

[0045] The pyramid multi-scale graph neural network architecture can efficiently extract features of different granularities by gradually compressing the hierarchy of the graph structure through the graph pooling operation. By establishing cross-layer jump connections, the information barrier between layers is broken, so that features of different scales can be transmitted and fused with each other, realizing the complement and integration of multi-scale features. In the decoding stage, through graph sampling and feature reconstruction, the complete spatial distribution information can be recovered.

[0046] Please refer to Figure 4 , the preferred embodiment of the present application gives a structure of a time convolution network, including convolution layers and residual blocks (represented by white squares in Figure 4 ). Convolution operation can be used for efficient parallel computing, making full use of modern hardware acceleration capabilities, and is suitable for large-scale data processing. By stacking convolution layers, features of different scales can be extracted, which can effectively capture local dependencies in sequence data.

[0047] As a preferred embodiment, step S2 adopts a dynamic dilated convolution structure to establish a time convolution network for extracting time sequence characteristics of the magnetic leakage field, which is specifically implemented through steps S21-S23. S21, a plurality of parallel channels are established, each channel comprising a multi-layer convolution network for extracting time sequence characteristics of the magnetic leakage field, and a cross-layer residual connection is used between network layers of the multi-layer convolution network.

[0048] By setting a parallel multi-channel structure, parallel extraction of different time scale characteristics can be achieved, improving the adaptability of the model to non-stationary signals; the cross-layer residual connection enables the trained model to maintain good convergence performance even when the number of layers of the multi-layer convolution network exceeds 10, solving the gradient vanishing problem in deep network training.

[0049] S22, the convolution kernel of the channel and the network layer is adaptively set by using a dynamic dilated convolution structure.

[0050] In the embodiments of the present application, the time sequence extraction part of the multi-layer convolution network is improved. By using a dynamic dilated convolution structure, the weight of each time step can be dynamically adjusted to automatically identify short-term and long-term dependencies and more accurately capture voltage and current data under different operating conditions of the converter transformer.

[0051] For example, under the dynamic dilated convolution structure, when a signal mutation is detected, the dilated rate is automatically reduced to capture short-term high-frequency details; and in the signal stable stage, the dilated rate is increased to obtain long-term trends.

[0052] S23, the top layer outputs of the multi-layer convolution network of the multi-channel are spliced to form a time convolution network for extracting time sequence characteristics of the magnetic leakage field.

[0053] The features extracted by different channels may reflect different aspects of the time sequence changes of the magnetic leakage field, such as short-term fluctuations and long-term trends. Through the feature splicing operation, the network can make full use of the complementary information extracted by each channel to form a more comprehensive and rich time sequence feature representation.

[0054] Further, preferably, step S22 adopts a dynamic dilated convolution structure to adaptively set the convolution kernel of the channel and the network layer, comprising: According to a preset dilated rate, a hole is injected into the convolution kernel of the network layer to obtain a traditional dilated convolution structure; An adaptive weight module is added to each time step of the traditional dilated convolution structure to form a dynamic dilated convolution structure; the adaptive weight module adaptively adjusts the dilated rate according to the feature complexity of the input data; The dynamic dilated convolution structure is introduced into each channel and each network layer of the multi-layer convolution network to adaptively adjust the convolution kernel.

[0055] It should be noted that the traditional dilated convolution structure expands the receptive field by a preset dilation rate, and the network can more effectively extract the long-term dependencies in the magnetic leakage field time series data, avoiding the problem of information loss caused by insufficient receptive field. At the same time, the injection of the cavity reduces the parameter amount and reduces the computational complexity, improves the computational efficiency on the premise of ensuring the performance of the model, and provides an efficient basic structure for subsequent dynamic adjustment.

[0056] In the embodiment of the application, by increasing the adaptive weight module, the network can flexibly capture multi-scale features at different time steps, avoiding the problem of information redundancy or omission caused by fixed dilation rate.

[0057] It should be further pointed out that in the adaptive adjustment process, there are two levels of adaptive adjustment process. The first level is to adaptively adjust the convolution kernel and the dilation rate of the channel, and the second level is to adaptively adjust the convolution kernel and the dilation rate of the network layer.

[0058] As a preferred embodiment, step S3, the graph neural network and the time convolution network are sequentially connected to form a branch of the neural network layer, comprising: Introducing a spatio-temporal cross-attention mechanism in the graph neural network to embed the response of the historical time step in the node to form a time-enhanced graph neural network; The time-enhanced graph neural network and the time convolution network are sequentially connected to form a branch of the neural network layer.

[0059] By introducing the spatio-temporal cross-attention mechanism, the response information of the historical time step is integrated into the node of the graph neural network, breaking the limitation of the graph neural network that only focuses on spatial features, enabling the graph neural network to process time series information, providing a more suitable feature expression form for subsequent connection with the time convolution network, and enhancing the relevance of the graph neural network and the time dimension.

[0060] The branch of the neural network layer formed after sequentially connecting the time-enhanced graph neural network and the time convolution network can focus on spatio-temporal feature extraction in one dimension.

[0061] In the embodiment of the application, as described in step S4, the double-branch neural network layer is connected with the feature fusion layer to establish a GNN-TCN model, and the double branches are current branch and voltage branch. The neural network layer of the current branch focuses on magnetic leakage field feature extraction in the current dimension, and the neural network layer of the voltage branch focuses on magnetic leakage field feature extraction in the voltage dimension.

[0062] Compared with a traditional space-time modeling method, the two branches respectively extract space-time features, and then realize deep fusion by means of an attention guiding mechanism, so that the coupling relationship between a space node and a time node thereof is effectively established, and therefore, the GNN-TCN model provided in the embodiment of the application can more accurately and synchronously depict complex space-time interaction characteristics and has stronger reconstruction expression.

[0063] As a preferred embodiment, the step S5 of acquiring sample data and training the GNN-TCN model by using the sample data comprises: A space synchronization mechanism based on geometric mapping is constructed to convert heterogeneous space coordinate systems of different sampling nodes into a unified coordinate system; Sampling interval errors of the sampling nodes in different time periods are eliminated by resampling; Time sequence electrical data and corresponding magnetic flux density data of the sampling nodes are acquired as sample data; the time sequence electrical data comprises time sequence current data and time sequence voltage data; The time sequence electrical data is input into the GNN-TCN model, and space-time features of the time sequence current data and the time sequence voltage data are respectively extracted by the double branches of the GNN-TCN model to output the magnetic flux density data as a target, so as to train the GNN-TCN model.

[0064] In the embodiment of the application, by increasing the space synchronization mechanism based on geometric mapping and the time correction module, the space coordinate systems of different sampling nodes can be unified, and the sampling interval errors can be corrected, so as to ensure the consistency of the input data in the space and time dimensions.

[0065] In a preferred embodiment, the sample data comprises converter transformer port voltage and current time sequence data , a space coordinate matrix and magnetic field time sequence data . The magnetic field time sequence data is any direction component or modulus value of x and z.

[0066] The converter transformer port voltage and current time sequence data is expressed as: ; wherein, and respectively represent winding voltage and winding current; represents the number of windings; is the length of the sequence data.

[0067] Further, according to the setting of the winding area and the core area, the space area is divided into and Two-part, structured field point matrix : ; Each row of the field point matrix represents the coordinates of a field point in three-dimensional space, exemplarily, represents the mth field point coordinates of the part. In embodiments of the present application, by constructing the field point matrix, the position information of multiple field points in space is orderly organized in matrix form.

[0068] On the basis of the field point matrix , the magnetic field time series data can be represented as: ; wherein, and are the corresponding coordinate point quantities derived from the leakage flux density.

[0069] When training the model, the is taken as the time series electrical data , and the is taken as the magnetic flux density data , and is represented as: ; ; wherein, is all time series electrical data at time t, including winding voltage and winding current of each winding; includes the leakage flux density at all coordinate points at time t.

[0070] It is considered that in actual measurement, the data collected by the probe or sensor is easily affected by random interference of the field environment, resulting in noise in the measurement results. In order to reduce the influence of noise on data, median filtering is used to process the time series electrical data and magnetic flux density data in the model training process.

[0071] It can be understood that when real-time running data is input into the trained GNN-TCN model, the same method can also be used to reduce the influence of noise on data.

[0072] Further, in order to eliminate the training weight bias caused by the order of magnitude between each dimension, the mean normalization method is used to preprocess the original data and , so that the model is easier to learn the relationship between features, and improve the training effect and generalization performance of the model.

[0073] It should be noted that, since the two branches have been enabled to focus on different electrical quantity extraction of leakage magnetic field characteristics in the training process of the model, when the real-time operation data of the converter transformer is input into the trained GNN-TCN model in step S6, the real-time operation data is directly input into the model as a whole, without the need to separately input the current data and the voltage data, and the model can automatically process the real-time operation data.

[0074] The GNN-TCN-based converter transformer leakage magnetic field measurement method provided by the embodiment of the application applies a graph neural network to leakage magnetic field spatial feature extraction through geometric properties and electromagnetic state properties, and combines a time convolution network with a dynamic expansion convolution structure to capture time features, so as to deeply mine the leakage magnetic field distribution law of the converter transformer under complex spatial structures and time sequence changes; the GNN-TCN model with two branches is constructed to realize efficient extraction of leakage magnetic field characteristics from multiple dimensions and multiple physical quantities; and the GNN-TCN model trained and put into use in sequence can accurately capture the nonlinear relationship between the transformer port voltage, current and internal leakage magnetic field, and further significantly improve the accuracy and efficiency of leakage magnetic field measurement in a complex electromagnetic environment.

[0075] The embodiment of the application provides a GNN-TCN-based converter transformer leakage magnetic field measurement system. Figure 5 The GNN-TCN-based converter transformer leakage magnetic field measurement system comprises a graph neural network construction module 11, a time convolution network construction module 12, a branch generation module 13, a model construction module 14, a model training module 15 and a model application module 16, wherein: The graph neural network construction module 11 is configured to establish a graph neural network for extracting spatial distribution characteristics of the leakage magnetic field; the node properties of the graph neural network comprise geometric structure properties and electromagnetic state properties; The time convolution network construction module 12 is configured to establish a time convolution network for extracting time sequence characteristics of the leakage magnetic field by using a dynamic expansion convolution structure; The branch generation module 13 is configured to sequentially connect the graph neural network and the time convolution network to form one branch of a neural network layer; The model construction module 14 is configured to connect the neural network layer with two branches and a feature fusion layer to establish a GNN-TCN model; the two branches of the neural network layer are a current branch and a voltage branch; The model training module 15 is configured to obtain sample data and train the GNN-TCN model by using the sample data; The model application module 16 is configured to input real-time operation data of the converter transformer into the trained GNN-TCN model to measure the leakage magnetic field of the converter transformer.

[0076] As a preferred implementation, the graph neural network construction module 11 comprises: a node and attribute construction unit configured to obtain nodes and node attributes according to geometric structures and electromagnetic information of physical entities in the converter transformer; the node attributes comprise geometric structure attributes and electromagnetic state attributes; an edge construction unit configured to construct edges connecting the nodes according to connection relationships and electromagnetic interaction relationships of the physical entities; the edges comprise physical connection edges and electromagnetic interaction edges; an edge weight updating unit configured to establish an edge weight control function according to electromagnetic induction law, and to adaptively update weights of the edges by using the edge weight control function; a graph neural network construction unit configured to establish a pyramid multi-scale graph neural network according to the nodes, the node attributes and the edges, so as to extract spatial distribution features of the magnetic flux leakage field.

[0077] Further, preferably, the node and attribute construction unit is specifically configured to: map the physical entities in the converter transformer to nodes in a graph structure; obtain geometric structure attributes of the nodes according to geometric structures of the physical entities in the converter transformer; obtain electromagnetic state attributes of the nodes according to electromagnetic information of the physical entities in the converter transformer; the electromagnetic information comprises current excitation, voltage state and magnetic flux density.

[0078] Further, preferably, the edge construction unit is specifically configured to: calculate an influence intensity of magnetic flux variation on neighbor nodes according to electromagnetic induction law, and establish an edge weight control function; establish an edge weight control factor according to the edge weight control function, and dynamically adjust information aggregation weights of the neighbor nodes by using the edge weight control factor in a graph convolution process, so as to adaptively update the weights of the edges.

[0079] Further, preferably, the graph neural network construction unit is specifically configured to: obtain a graph structure according to the nodes, the node attributes and the edges; start from a bottom graph structure and gradually compress the graph structure to generate an upper graph structure, until a top graph structure is generated, by using a graph pooling algorithm; establish cross-layer jump connections between layers to perform multi-scale feature transmission; reconstruct global features obtained from the top graph structure by using a graph sampling algorithm, so as to obtain spatial distribution features.

[0080] As a preferred implementation, the time convolution network construction module 12 comprises: The channel establishment unit is configured to establish multiple parallel channels, each of which comprises a multi-layer convolutional network for extracting time sequence features of the magnetic leakage field, and cross-layer residual connections are used between network layers of the multi-layer convolutional network. The convolution kernel setting unit is configured to adaptively set convolution kernels of the channels and the network layers by using a dynamic dilated convolution structure. The time convolution network construction unit is configured to splice features of top-layer outputs of the multi-layer convolutional networks of the multiple channels to construct a time convolution network for extracting time sequence features of the magnetic leakage field.

[0081] Further, preferably, the convolution kernel setting unit is specifically configured to: inject holes in the convolution kernels of the network layers according to a preset dilation rate to obtain a conventional dilated convolution structure; add an adaptive weight module to each time step of the conventional dilated convolution structure to form a dynamic dilated convolution structure, wherein the adaptive weight module adaptively adjusts the dilation rate according to the feature complexity of input data; introduce the dynamic dilated convolution structure into each channel and each network layer of the multi-layer convolutional network to adaptively adjust the convolution kernels.

[0082] As a preferred implementation, the branch generation module 13 is specifically configured to: introduce a spatio-temporal cross-attention mechanism into the graph neural network to embed responses of historical time steps in nodes to form a time sequence enhanced graph neural network; sequentially connect the time sequence enhanced graph neural network and the time convolution network to form one branch of the neural network layer.

[0083] As a preferred implementation, the model training module 15 is specifically configured to: construct a spatial synchronization mechanism based on geometric mapping to convert heterogeneous spatial coordinate systems of different sampling nodes into a unified coordinate system; eliminate sampling interval errors of the sampling nodes in different time periods by resampling; obtain time sequence electrical data and corresponding magnetic flux density data of the sampling nodes as sample data, wherein the time sequence electrical data includes time sequence current data and time sequence voltage data; input the time sequence electrical data into the GNN-TCN model, extract spatio-temporal features of the time sequence current data and the time sequence voltage data by the two branches of the GNN-TCN model respectively, and output the magnetic flux density data as a target to train the GNN-TCN model.

[0084] The leakage magnetic field measurement system based on the GNN-TCN provided by the embodiment of the present application can apply the graph neural network to leakage magnetic field space feature extraction through geometric properties and electromagnetic state properties, and can capture time features by combining a time convolution network with a dynamic expansion convolution structure, so as to deeply mine the leakage magnetic field distribution law under the complex spatial structure and time sequence change of the converter transformer; the GNN-TCN model with double branches is constructed, and leakage magnetic field features are efficiently extracted from multiple dimensions and multiple physical quantities; the GNN-TCN model trained and put into use in sequence can accurately capture the nonlinear relationship between the transformer port voltage, current and internal leakage magnetic field, and then significantly improve the accuracy and efficiency of leakage magnetic field measurement in a complex electromagnetic environment.

[0085] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by a computer program instructing related hardware, and the program can be stored in a computer readable storage medium. When the program is executed, the program can include the processes of the above-mentioned embodiments of each method. The storage medium can be a magnetic disc, an optical disc, a read-only memory (ROM) or a random access memory (RAM), etc.

[0086] The above is the preferred embodiment of the present application. It should be noted that those skilled in the art can make several improvements and refinements without departing from the principles of the present application. These improvements and refinements are also considered within the scope of protection of the present application.

Claims

1. A GNN-TCN-based converter transformer leakage magnetic field measurement method, characterized in that, The method comprises the following steps: a graph neural network for extracting spatial distribution characteristics of the magnetic leakage field is established; node attributes of the graph neural network include geometric structure attributes and electromagnetic state attributes; a time convolution network for extracting time sequence characteristics of the magnetic leakage field is established by using a dynamic dilated convolution structure; the graph neural network and the time convolution network are sequentially connected to form a branch of a neural network layer; the double-branch neural network layer is connected with a feature fusion layer to establish a GNN-TCN model; the double-branch neural network layer includes a current branch and a voltage branch; sample data is obtained, and the GNN-TCN model is trained by using the sample data; real-time operation data of the converter transformer is input into the trained GNN-TCN model to obtain the magnetic leakage field of the converter transformer.

2. The GNN-TCN-based converter transformer leakage magnetic field calculation method of claim 1, wherein, The graph neural network for extracting spatial distribution characteristics of the magnetic leakage field comprises the following steps: node and node attributes are obtained according to geometric structure and electromagnetic information of physical entities in the converter transformer; the node attributes include geometric structure attributes and electromagnetic state attributes; edges connected with the nodes are constructed according to connection relationships and electromagnetic interaction relationships of the physical entities; the edges include physical connection edges and electromagnetic interaction edges; an edge weight control function is established according to the electromagnetic induction law, and the edge weight control function is used to adaptively update weights of the edges; a pyramid multi-scale graph neural network is established according to the nodes, the node attributes and the edges to extract spatial distribution characteristics of the magnetic leakage field.

3. The GNN-TCN-based converter transformer leakage magnetic field calculation method of claim 2, wherein, The node and node attributes are obtained according to geometric structure and electromagnetic information of physical entities in the converter transformer, and the method comprises the following steps: the physical entities in the converter transformer are mapped into nodes in a graph structure; geometric structure attributes of the nodes are obtained according to geometric structure of the physical entities in the converter transformer; electromagnetic state attributes of the nodes are obtained according to electromagnetic information of the physical entities in the converter transformer; the electromagnetic information includes current excitation, voltage state and magnetic flux density.

4. The GNN-TCN-based converter transformer leakage magnetic field calculation method of claim 2, wherein, The edge weight control function is established according to the electromagnetic induction law, and the edge weight control function is used to adaptively update weights of the edges, and the method comprises the following steps: the influence intensity of the magnetic flux change on neighbor nodes is calculated according to the electromagnetic induction law to establish the edge weight control function; an edge weight control factor is established according to the edge weight control function; in the graph convolution process, the information aggregation weight of the neighbor nodes is dynamically adjusted through the edge weight control factor to adaptively update the weights of the edges.

5. The GNN-TCN-based converter transformer leakage magnetic field calculation method of claim 2, wherein, The pyramid multi-scale graph neural network is established according to the nodes, the node attributes and the edges to extract spatial distribution characteristics of the magnetic leakage field, and the method comprises the following steps: a graph structure is obtained according to the nodes, the node attributes and the edges; a graph pooling algorithm is used to start from a bottom graph structure, gradually compresses, generates an upper graph structure, and generates a top graph structure until the top graph structure is generated; cross-layer jump connections are established between layers to perform multi-scale feature transmission; after global features are obtained from the top graph structure, the global features are reconstructed through a graph sampling algorithm to obtain spatial distribution characteristics.

6. The GNN-TCN-based converter transformer leakage magnetic field calculation method of claim 1, wherein, The time convolution network for extracting time sequence characteristics of the magnetic leakage field is established by using a dynamic dilated convolution structure, and the method comprises the following steps: Parallel multi-channels are established, each channel including a multi-layer convolutional network for extracting time sequence features of the magnetic leakage field, and cross-layer residual connections are used between network layers of the multi-layer convolutional network; A dynamic dilated convolution structure is used to adaptively set the convolution kernel of the channel and the network layer; The top layer output of the multi-layer convolutional network of the multi-channel is spliced to form a time convolutional network for extracting time sequence features of the magnetic leakage field.

7. The GNN-TCN-based converter transformer leakage magnetic field calculation method of claim 6, wherein, The adaptive weight module is added to each time step of the traditional dilated convolution structure to form a dynamic dilated convolution structure; the adaptive weight module adaptively adjusts the dilation rate according to the feature complexity of the input data; The dynamic dilated convolution structure is introduced into each channel and each network layer of the multi-layer convolutional network to adaptively adjust the convolution kernel. The spatio-temporal cross-attention mechanism is introduced into the graph neural network to embed the response of the historical time step in the node to form a time sequence enhanced graph neural network; The time sequence enhanced graph neural network and the time convolutional network are sequentially connected to form one branch of the neural network layer.

8. The GNN-TCN-based converter transformer leakage magnetic field calculation method of claim 1, wherein, The sample data is obtained, and the GNN-TCN model is trained using the sample data, including: A spatial synchronization mechanism based on geometric mapping is constructed to convert heterogeneous spatial coordinate systems of different sampling nodes into a unified coordinate system; The sampling interval error of the sampling nodes in different time periods is eliminated by resampling; 9. The GNN-TCN-based converter transformer leakage magnetic field measurement method of claim 1, wherein, The time sequence electrical data and the corresponding magnetic flux density data of the sampling nodes are obtained as sample data; the time sequence electrical data includes time sequence current data and time sequence voltage data; The time sequence electrical data is input into the GNN-TCN model, and the spatial and temporal features of the time sequence current data and the time sequence voltage data are extracted by the double branches of the GNN-TCN model to output the magnetic flux density data as the target, and the GNN-TCN model is trained. It includes: A graph neural network construction module is configured to establish a graph neural network for extracting spatial distribution features of the magnetic leakage field; the node attributes of the graph neural network include geometric structure attributes and electromagnetic state attributes; A time convolutional network construction module is configured to establish a time convolutional network for extracting time sequence features of the magnetic leakage field using a dynamic dilated convolution structure; 10. A GNN-TCN-based converter transformer leakage magnetic field measurement system, characterized in that, A branch generation module is configured to sequentially connect the graph neural network and the time convolutional network to form one branch of the neural network layer; A model construction module is configured to connect the neural network layer of the double branches and the feature fusion layer to establish a GNN-TCN model; the double branches of the neural network layer are current branches and voltage branches; A model training module is configured to obtain sample data and train the GNN-TCN model using the sample data. ​ ​ ​ A model application module is configured to input real-time operation data of the converter transformer into the trained GNN-TCN model to calculate leakage magnetic field of the converter transformer.

Images