Converter valve power module digital twinning construction method, system, equipment and medium

Abstract

The invention relates to the technical field of digital twinning, and discloses a converter valve power module digital twinning construction method, system, equipment and medium, and the method comprises the steps: obtaining a training sample set needed by a model through the physical field simulation data of a converter valve power module, and carrying out the training through a Transform-based neural network model, the simulation digital twinning prediction model obtained through training of a Tranformer neural network structure can output physical field data of the converter valve power module according to input operation condition parameters of the converter valve power module, so that the purpose of digital twinning is achieved, online monitoring of the converter valve power module in engineering equipment is achieved, and the safety of the engineering equipment is improved. And the model calculation efficiency is improved, and the method can adapt to different working condition scenes.

Description

Technical Field

[0001] The present invention relates to the field of digital twin technology, and in particular to a method, system, device and medium for constructing a digital twin of a converter valve power module. Background Art

[0002] Digital twins play a vital role in engineering design and equipment monitoring tasks, and are mainly divided into physical model-based methods and data-driven methods.

[0003] Among them, the physical model-based method is based on the finite element and finite volume methods. Its essence is to solve partial differential equations. Faced with complex working conditions and structures, the dimensions of the equations are high, and it is generally necessary to reduce the degrees of freedom of solving partial differential equations through model reduction methods. In the numerical calculation of the reduced-order model, the physical model-based method often requires multiple iterations over a long period of time, resulting in low model calculation efficiency. In practical applications, the reduced-order model has low accuracy and poor generalization ability, making it difficult to cope with complex and changing working conditions. Summary of the invention

[0004] In view of this, in order to solve the above-mentioned technical problems, the present invention provides a method, system, device and medium for constructing a digital twin of a converter valve power module.

[0005] A first aspect of the present invention provides a method for constructing a digital twin of a converter valve power module, comprising:

[0006] According to the physical field simulation data of the converter valve power module, the simulation sample data of the converter valve power module is obtained; wherein the simulation sample data includes operating condition parameters;

[0007] Determine a training sample set according to the simulation sample data;

[0008] The training sample set is input into a Transformer-based neural network model for training to obtain a simulation digital twin prediction model of the converter valve power module. The simulation digital twin prediction model is used to output the physical field data of the converter valve power module according to the input operating condition parameters of the converter valve power module.

[0009] Preferably, obtaining simulation sample data of the converter valve power module according to the physical field simulation data of the converter valve power module includes:

[0010] Performing finite element modeling on the converter valve power module to construct a finite element model of the converter valve power module;

[0011] Inputting a plurality of operating condition parameters of the converter valve power module into a finite element model of the converter valve power module in sequence to perform finite element analysis, and obtaining physical field data corresponding to the plurality of operating condition parameters respectively;

[0012] Simulation sample data of the converter valve power module are constructed according to the plurality of operating condition parameters and the physical field data respectively corresponding to the plurality of operating condition parameters.

[0013] Preferably, the training sample set is input into a Transformer-based neural network model for training to obtain a simulation digital twin prediction model of the converter valve power module, and the simulation digital twin prediction model is used to output the physical field data of the converter valve power module according to the input operating condition parameters of the converter valve power module, and before that, the simulation digital twin prediction model further includes:

[0014] The training sample set is subjected to standardization preprocessing.

[0015] Preferably, the Transformer-based neural network model includes an embedding layer, an encoder layer, an average pooling layer, a hidden linear layer and an output linear layer connected in sequence.

[0016] Preferably, the step of inputting the training sample set into a Transformer-based neural network model for training to obtain a simulation digital twin prediction model of the converter valve power module comprises:

[0017] Performing an embedding operation on a plurality of input features of the training sample set through the embedding layer to obtain a feature embedding vector matrix, wherein the embedding operation is used to convert the plurality of input features into vectors of equal length;

[0018] Determine a position encoding of each input feature according to the parity feature of the position of each input feature in the feature embedding vector matrix;

[0019] Performing a serial operation on the position encoding of each of the input features to obtain a position vector matrix;

[0020] Performing a sum operation on the feature embedding vector matrix and the position vector matrix to obtain a feature embedding matrix;

[0021] Inputting the feature embedding matrix into the encoder layer, and performing feature extraction on the feature embedding matrix based on self-attention to obtain embedded features;

[0022] Performing pooling processing on the embedded features through the average pooling layer to obtain embedded pooling features;

[0023] Performing a hidden linear transformation on the embedded pooling features, and mapping the embedded pooling features to a new feature space to obtain embedded linear features;

[0024] The embedded linear features are output to obtain output features corresponding to each of the input features.

[0025] Preferably, the encoder layer comprises a plurality of stacked encoding layers, each encoding layer comprising a multi-head attention layer, a first residual link and a layer normalization layer, a feedforward neural network layer and a second residual link and a layer normalization layer connected in sequence;

[0026] The step of inputting the feature embedding matrix into the encoder layer and extracting features from the feature embedding matrix based on self-attention to obtain embedded features includes:

[0027] Inputting the feature embedding matrix into a first encoding layer of the encoder layer, and in the first encoding layer, performing a multi-head attention operation on the feature embedding matrix based on self-attention through the multi-head attention layer;

[0028] Performing residual link and layer normalization operations on the feature embedding matrix after the multi-head attention operation through the first residual link and layer normalization layer;

[0029] Inputting the feature embedding matrix after the residual link and layer normalization operations into the feedforward neural network layer, and extracting the initial embedding features of the feature embedding matrix after the residual link and layer normalization operations;

[0030] Performing residual link and layer normalization operations on the initial embedded features through the second residual link and layer normalization layer, and outputting the normalized embedded features as input features of a next encoding layer adjacent to the first encoding layer;

[0031] Based on the next encoding layer, the step of inputting the feature embedding matrix into the first encoding layer of the encoder layer is performed, in the first encoding layer, through the multi-head attention layer, based on self-attention, on the feature embedding matrix, until all encoding layers of the encoder layer complete the multi-head attention operation on the feature embedding matrix, and then output the embedded features.

[0032] Preferably, the method further comprises:

[0033] The learning rate of the Transformer-based neural network model is optimized according to the cosine annealing algorithm.

[0034] In a second aspect, the present invention further provides a system for constructing a digital twin of a converter valve power module, comprising:

[0035] A data acquisition module, used to obtain simulation sample data of the converter valve power module according to the physical field simulation data of the converter valve power module; wherein the simulation sample data includes operating condition parameters;

[0036] A sample determination module, used to determine a training sample set according to the simulation sample data;

[0037] A model building module is used to input the training sample set into a Transformer-based neural network model for training to obtain a simulated digital twin prediction model of the converter valve power module. The simulated digital twin prediction model is used to output the physical field data of the converter valve power module based on the input operating condition parameters of the converter valve power module.

[0038] In a third aspect, the present invention further provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the method for constructing a digital twin of a converter valve power module as described in the first aspect.

[0039] In a fourth aspect, the present invention further provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed, the steps of the method for constructing a digital twin of a converter valve power module as described in the first aspect are implemented.

[0040] It can be seen from the above technical solutions that the present invention obtains the training sample set required by the model through the physical field simulation data of the converter valve power module, and uses the Transformer-based neural network model for training. The simulated digital twin prediction model obtained by training with the Transformer neural network structure can output the physical field data of the converter valve power module according to the input operating condition parameters of the converter valve power module. The operating condition parameters and physical field data of the converter valve power module can be mapped by the simulated digital twin prediction model, thereby achieving the purpose of digital twin, realizing online monitoring of the converter valve power module in the engineering equipment, improving the model calculation efficiency, and being able to adapt to different operating scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0042] Figure 1An application environment for a method for constructing a digital twin of a converter valve power module provided in an embodiment of the present invention;

[0043] Figure 2 A flowchart of a method for constructing a digital twin of a converter valve power module provided in an embodiment of the present invention;

[0044] Figure 3 An architecture diagram of a Transformer-based neural network model provided in an embodiment of the present invention;

[0045] Figure 4 A structural diagram of a converter valve power module provided in an embodiment of the present invention;

[0046] Figure 5 A schematic diagram of the structure of a digital twin construction system for a converter valve power module provided by an embodiment of the present invention;

[0047] Figure 6 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0048] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0049] The method for constructing a digital twin of a converter valve power module provided in the embodiment of the present application can be applied to Figure 1 In the application environment shown. Among them, the finite element simulation software of the converter valve power module communicates with the server 102 through the network. The data storage system can store the data that the server 102 needs to process. The data storage system can be integrated on the server 102, or it can be placed on the cloud or other network servers. The server 102 obtains the simulation sample data of the converter valve power module based on the physical field simulation data of the converter valve power module; wherein the simulation sample data includes operating condition parameters; a training sample set is determined according to the simulation sample data; the training sample set is input into the Transformer-based neural network model for training to obtain a simulation digital twin prediction model of the converter valve power module, and the simulation digital twin prediction model is used to output the physical field data of the converter valve power module according to the input operating condition parameters of the converter valve power module. Server 102 can be an independent physical server, or it can be a server cluster or distributed system composed of multiple physical servers, or it can be a cloud server that provides cloud computing services.

[0050] like Figure 2 As shown, the embodiment of the present application provides a method for constructing a digital twin of a converter valve power module, and the method is applied to Figure 1 The server 102 in the example is used as an example to illustrate the method, which includes the following steps S1 to S3. Among them:

[0051] Step S1, obtaining simulation sample data of the converter valve power module according to the physical field simulation data of the converter valve power module; wherein the simulation sample data includes operating condition parameters.

[0052] The physical field simulation data of the converter valve power module may be obtained according to the physical field simulation process of the converter valve power module.

[0053] Specifically, according to the physical field simulation data of the converter valve power module, simulation sample data of the converter valve power module is obtained, including:

[0054] Step S101: Perform finite element modeling on the converter valve power module to construct a finite element model of the converter valve power module.

[0055] Among them, the present invention can perform finite element modeling on the physical field information of the converter valve power module according to known information such as material parameters, grid division, load boundary conditions, etc. of the converter valve power module.

[0056] Step S102: inputting a plurality of operating condition parameters of the converter valve power module into a finite element model of the converter valve power module in sequence, performing finite element analysis, and obtaining physical field data corresponding to the plurality of operating condition parameters.

[0057] Among them, the operating condition parameters include but are not limited to system reactive power, bridge arm current, capacitor current, IGBT1 loss, diode loss (specifically two different diode losses), loss (specifically three different IGBT losses), balancing resistor loss and DC capacitor loss.

[0058] The physical field data may be temperature field data, humidity field data, force field data, etc.

[0059] Step S103: construct simulation sample data of the converter valve power module according to the multiple operating condition parameters and the physical field data corresponding to the multiple operating condition parameters.

[0060] Step S2: determine a training sample set based on the simulation sample data.

[0061] The simulation sample data may be divided into a training sample set and a test sample set according to a certain ratio.

[0062] Step S3: input the training sample set into the Transformer-based neural network model for training to obtain a simulation digital twin prediction model of the converter valve power module. The simulation digital twin prediction model is used to output the physical field data of the converter valve power module according to the input operating condition parameters of the converter valve power module.

[0063] Among them, Transformer is a model based on the self-attention mechanism, which uses the self-attention layer to process the input and extract the comprehensive features of the input. Compared with feedforward neural networks, it does not require too deep hidden layers, thus solving the problems of gradient vanishing and overfitting; unlike recurrent neural networks that cannot be calculated in parallel, the self-attention mechanism solves the problem of inability to parallelize, giving Transformer the ability to parallelize, thereby greatly accelerating the speed of model training; compared with convolutional neural networks, Transformer using the self-attention mechanism has stronger feature extraction capabilities and generalization performance, without the need to stack a deeper model depth. When the number of features is large, the feature extraction capability of a single-layer Transformer is comparable to that of a stacked multi-layer convolutional neural network.

[0064] In some embodiments, in order to improve the accuracy of the model, the embodiments of the present application also test the simulation digital twin prediction model of the converter valve power module through a test sample set, and use the test results to optimize the simulation digital twin prediction model of the converter valve power module until the accuracy requirements are met.

[0065] At the same time, the embodiment of the present application can also visualize the physical field data output by the simulation digital twin prediction model of the converter valve power module.

[0066] Exemplarily, VTK technology can be used to perform three-dimensional graphics rendering of the physical field data output by the simulation digital twin prediction model of the converter valve power module based on the grid file information (including unit composition and node coordinate description) to form a 3D result distribution cloud map for user viewing.

[0067] It should be noted that the embodiment of the present application obtains the training sample set required by the model through the physical field simulation data of the converter valve power module, and uses the Transformer-based neural network model for training. The simulated digital twin prediction model obtained by training with the Transformer neural network structure can output the physical field data of the converter valve power module according to the input operating condition parameters of the converter valve power module. The operating condition parameters and physical field data of the converter valve power module can be mapped through the simulated digital twin prediction model, thereby achieving the purpose of digital twin, realizing online monitoring of the converter valve power module in engineering equipment, improving the model calculation efficiency, and adapting to different operating scenarios.

[0068] In some embodiments, in order to eliminate the differences between features, the embodiment of the present application further includes, after step S2 and before step S3:

[0069] The training sample set is preprocessed by standardization.

[0070] Among them, Z-score standardization can be used to convert the training sample set into dimensionless data, eliminate the differences between features, and enable the model to converge quickly, so as to ensure the comparability between different types of data. The specific formula of Z-score standardization is as follows:

[0071]

[0072] In the formula, X represents the original data, and represent the mean and standard deviation of the data respectively, and Z represents the standardized data.

[0073] It should be noted that in the prior art, the data-driven method uses multiple boundary conditions and loads as the input of the proxy model to train the model and achieve the purpose of predicting the physical field. It is a black box model method for learning mapping relationships. For example, a real-time simulation model is constructed using random forests, feedforward neural networks, recurrent neural networks, and convolutional neural networks. Based on the data-driven method, in the process of building a real-time simulation model, the feedforward neural network is prone to gradient disappearance and overfitting problems, resulting in low accuracy in practical applications. The recurrent neural network is a time series model that iteratively processes each time step in the sequence and has serious defects in parallel computing. The convolutional neural network needs to stack the depth and width of the neural network model to improve the comprehensive extraction capability of the features to improve the accuracy of the model, but it leads to the problem of increased model complexity and slower learning efficiency.

[0074] To this end, the present application embodiment proposes the following Figure 3 The architecture of the Transformer-based neural network model shown in the embodiment of the present application is based on Transformer, and uses the Z-score standardized data as input, passes through the Transformer-based network model, and integrates position encoding to finally obtain the physical field simulation result.

[0075] like Figure 3As shown in the figure, the neural network model based on Transformer includes an embedding layer (Embedding), an encoder layer (TransformerEncoderLayer), an average pooling layer (Mean-Pooling), a hidden linear layer (Hidden linear layer) and an output linear layer (Output linear layer) connected in sequence. Among them, in the encoder layer (TransformerEncoderLayer), positional encoding (Positional Encoding) can be performed.

[0076] Specifically, in the embodiment of the present application, the training sample set is input into the Transformer-based neural network model for training in step S3 to obtain a simulation digital twin prediction model of the converter valve power module, including:

[0077] Step S301: embedding a plurality of input features of a training sample set through an embedding layer to obtain a feature embedding vector matrix. The embedding operation is used to convert the plurality of input features into vectors of equal length.

[0078] The embedding operation is:

[0079]

[0080] In the formula, represents the feature embedding vector matrix, represents the input features, and Both represent parameter matrices that can be learned.

[0081] Step S302: Determine the position encoding of each input feature according to the parity feature of the position of each input feature in the feature embedding vector matrix.

[0082] Among them, the feature embedding vector matrix obtained after the input data is embedded, and the parity feature of the position is the position information of each input feature in the feature embedding vector matrix.

[0083] Among them, the feature embedding vector matrix can be 3-dimensional, and its dimension can be (batch_size, input_size, dim) where batch_size is the batch size, input_size is the input dimension size, and dim is the dimension of the embedding vector.

[0084] The calculation process of the position encoding for each input feature is:

[0085]

[0086] In the formula, Represents a d-dimensional vector The first elements, i is a natural number, , represents the position of the input feature in the input sequence, is the feature embedding vector matrix Dimension.

[0087] Step S303: perform a serial operation on the position encoding of each input feature to obtain a position vector matrix.

[0088] Among them, the position vector matrix after the series operation can be expressed as:

[0089]

[0090] In the formula, is the position vector matrix, is the positional encoding of the input feature at position t.

[0091] Step S304: Add the feature embedding vector matrix and the position vector matrix to obtain a feature embedding matrix.

[0092] Among them, the feature embedding matrix E can be expressed as:

[0093]

[0094] In the formula, is a constant that can be set to prevent the embedding vector from being overwhelmed by the positional encoding.

[0095] Step S305: input the feature embedding matrix into the encoder layer, and perform feature extraction on the feature embedding matrix based on self-attention to obtain embedded features.

[0096] like Figure 3 As shown, the encoder layer includes multiple stacked encoding layers, each encoding layer includes a multi-head attention layer (Multi-HeadAttention), a first residual link and a layer normalization layer (Add & Norm), a feedforward neural network layer (Feed Forward) and a second residual link and a layer normalization layer (Add & Norm) connected in sequence.

[0097] Specifically, the feature embedding matrix is ​​input into the encoder layer, and the feature embedding matrix is ​​extracted based on self-attention to obtain embedded features, including:

[0098] Step S3051: input the feature embedding matrix to the first encoding layer of the encoder layer. In the first encoding layer, a multi-head attention operation is performed on the feature embedding matrix based on self-attention through a multi-head attention layer.

[0099] Among them, multi-head attention is based on self-attention, and the formula of self-attention is as follows:

[0100]

[0101] In the formula, is the self-attention score, Q, K and V are the query vector (Query), key vector (Key) and value vector (Value), respectively.

[0102] The self-attention scores are normalized by the softmax function to obtain the attention weights, which are divided by the square root of the dimension of each head. This operation is called "scaling", which helps keep the gradient of the softmax function within a reasonable range, thereby improving the training stability of the model.

[0103] The calculation formula of multi-head attention is as follows:

[0104]

[0105]

[0106] In the formula, , represents the number of attention heads, , , , are all learnable parameter matrices, , is the weight matrix of the linear transformation, which is used to restore the concatenated output to the original dimension.

[0107] Step S3052: Perform residual link and layer normalization operations on the feature embedding matrix after the multi-head attention operation through the first residual link and layer normalization layer.

[0108] Among them, the residual connection and layer normalization operations are:

[0109]

[0110] Step S3053: input the feature embedding matrix after residual link and layer normalization operations into the feedforward neural network layer, and extract the initial embedding features of the feature embedding matrix after residual link and layer normalization operations.

[0111] Among them, it is assumed that the input of the feedforward neural network is , then the operation of extracting features in the feedforward neural network layer is:

[0112]

[0113] In the formula, It is the output after one encoding layer.

[0114] Step S3054: Perform residual linking and layer normalization operations on the initial embedded features through the second residual linking and layer normalization layer, and output the normalized embedded features as input features of the next coding layer adjacent to the first coding layer.

[0115] Among them, the second residual link and layer normalization layer are consistent with the residual link and layer normalization operations of the first residual link and layer normalization layer.

[0116] Step S3055, based on the next coding layer, jump to the first coding layer that inputs the feature embedding matrix into the encoder layer. In the first coding layer, a multi-head attention operation is performed on the feature embedding matrix based on self-attention through a multi-head attention layer, until all coding layers of the encoder layer complete the multi-head attention operation on the feature embedding matrix, and then output the embedded features.

[0117] Step S306: pooling the embedded features through an average pooling layer to obtain embedded pooled features.

[0118] Step S307: perform hidden linear transformation on the embedded pooling features, and map the embedded pooling features to a new feature space to obtain embedded linear features.

[0119] Among them, the hidden linear layer performs a linear transformation on the embedded pooling features, embedding the pooling features into a new feature space, thereby extracting higher-level features and obtaining embedded linear features.

[0120] Step S308: output the embedded linear features to obtain output features corresponding to each input feature.

[0121] Among them, the output linear layer outputs the embedded linear features to obtain the output features corresponding to each input feature, that is, the physical field data.

[0122] In the training process of convolutional neural networks, the size of the learning rate plays a vital role, controlling the degree to which the neural network adjusts the network weights according to the loss gradient. Too large a learning rate will cause the gradient to disappear, while too small a learning rate will lead to a slow learning speed and fall into a local optimum. Therefore, it is particularly important to choose a suitable learning rate. In some embodiments, the embodiments of the present application optimize the learning rate of the Transformer-based neural network model according to the cosine annealing algorithm.

[0123] Specifically, the process of optimizing the learning rate of the Transformer-based neural network model according to the cosine annealing algorithm includes:

[0124]

[0125] In the formula, Represents the ratio of the learning rate after the tth iteration to the basic learning rate, Represents the total number of iterations of model training.

[0126] Based on Adaptive Moment Estimation (Adam), the embodiment of the present application selects cosine annealing to perform periodic changes of the cosine function, so as to further approach the global optimal solution and improve the overall accuracy of the model. It shows a very excellent display effect in the physical field-level digital twin task of the converter valve power module.

[0127] The following is a calculation example of a method for constructing a digital twin of a converter valve power module provided in combination with an embodiment of the present application.

[0128] This example is used for the digital twin of the converter valve power module. The converter valve power module structure includes silicon stacks, conductive copper bars, power module supports and shells, insulating pads and other components. The silicon stack is mainly made of switch devices, water cooling plates, and thyristors. The converter valve power module model is meshed with an average mesh size of 5mm. The model has a total of 1.19 million nodes and 769,000 units. Figure 4 shown.

[0129] During the operation of the converter valve power module, the current passing through it generates a strong alternating electromagnetic field inside it, which causes eddy currents and electromagnetic forces to be generated in the chassis, copper busbars, silicon stacks and other structural parts. Eddy current heating and device loss heating cause the temperature of the power module to rise during operation. At the same time, the water cooling system cools the power module through the flow of cooling medium to ensure its working reliability. The current, water cooling system flow, inlet temperature and ambient temperature are used as input variables, and the output variable is the temperature data of the converter valve power module.

[0130] The input variables and output results are divided into training sets and test sets according to a certain ratio, and the training sets are processed with standardized features before being input into the network based on the Transformer neural network model for training. After the training, the model can perform calculations in seconds based on the input variables, thus achieving the purpose of digital twins.

[0131] Based on the same inventive concept, an embodiment of the present application also provides a digital twin construction system for a converter valve power module for implementing the above-mentioned digital twin construction method for a converter valve power module.

[0132] The implementation solution for solving the problem provided by the system is similar to the implementation solution recorded in the above method. Therefore, the specific limitations in one or more embodiments of the digital twin construction system for the converter valve power module provided below can be referred to the limitations on the digital twin construction method for the converter valve power module in the above text, and will not be repeated here.

[0133] like Figure 5 As shown, the embodiment of the present application also provides a digital twin construction system for a converter valve power module, including:

[0134] The data acquisition module 100 is used to obtain simulation sample data of the converter valve power module according to the physical field simulation data of the converter valve power module; wherein the simulation sample data includes operating condition parameters.

[0135] The sample determination module 200 is used to determine a training sample set according to the simulation sample data.

[0136] The model building module 300 is used to input the training sample set into the Transformer-based neural network model for training to obtain a simulated digital twin prediction model of the converter valve power module. The simulated digital twin prediction model is used to output the physical field data of the converter valve power module based on the input operating condition parameters of the converter valve power module.

[0137] In some embodiments, the data acquisition module 100 is used to perform finite element modeling on the converter valve power module and construct a finite element model of the converter valve power module; input multiple operating condition parameters of the converter valve power module into the finite element model of the converter valve power module in sequence for finite element analysis to obtain physical field data corresponding to the multiple operating condition parameters; and construct simulation sample data of the converter valve power module based on the multiple operating condition parameters and the physical field data corresponding to the multiple operating condition parameters.

[0138] In some embodiments, the system further comprises:

[0139] The preprocessing module is used to perform standardized preprocessing on the training sample set.

[0140] In some embodiments, the Transformer-based neural network model includes an embedding layer, an encoder layer, an average pooling layer, a hidden linear layer, and an output linear layer connected in sequence.

[0141] In some embodiments, the model building module 300 includes:

[0142] The embedding operation module is used to perform an embedding operation on multiple input features of the training sample set through an embedding layer to obtain a feature embedding vector matrix. The embedding operation is used to convert multiple input features into vectors of equal length.

[0143] The position encoding module is used to determine the position encoding of each input feature according to the parity feature of the position of each input feature in the feature embedding vector matrix.

[0144] A serial operation module is used to perform serial operation on the position encoding of each input feature to obtain a position vector matrix;

[0145] An addition operation module, used for performing an addition operation on the feature embedding vector matrix and the position vector matrix to obtain a feature embedding matrix;

[0146] The feature extraction module is used to input the feature embedding matrix into the encoder layer and extract features from the feature embedding matrix based on self-attention to obtain embedded features.

[0147] The pooling module is used to perform pooling processing on the embedded features through the average pooling layer to obtain the embedded pooling features;

[0148] The hidden transformation module is used to perform hidden linear transformation on the embedded pooling features and map the embedded pooling features to a new feature space to obtain embedded linear features.

[0149] The output module is used to output the embedded linear features to obtain the output features corresponding to each input feature.

[0150] Among them, the encoder layer includes multiple layers of stacked encoding layers, each encoding layer includes a multi-head attention layer, a first residual link and a layer normalization layer, a feedforward neural network layer, and a second residual link and a layer normalization layer connected in sequence.

[0151] The feature extraction module includes:

[0152] The attention module is used to input the feature embedding matrix into the first encoding layer of the encoder layer. In the first encoding layer, a multi-head attention operation is performed on the feature embedding matrix based on self-attention through a multi-head attention layer;

[0153] A first residual module, used for performing residual link and layer normalization operations on the feature embedding matrix after the multi-head attention operation through a first residual link and a layer normalization layer;

[0154] A feed-forward neural module, used for inputting the feature embedding matrix after residual link and layer normalization operation into the feed-forward neural network layer, and extracting the initial embedding features of the feature embedding matrix after residual link and layer normalization operation;

[0155] a second residual module, used for performing residual link and layer normalization operations on the initial embedded features through a second residual link and layer normalization layer, and outputting the normalized embedded features as input features of a next encoding layer adjacent to the first encoding layer;

[0156] The output operation module is used to jump to the first encoding layer of the encoder layer to input the feature embedding matrix based on the next encoding layer. In the first encoding layer, the multi-head attention layer performs a multi-head attention operation on the feature embedding matrix based on self-attention until all encoding layers of the encoder layer complete the multi-head attention operation on the feature embedding matrix, and then output the embedded features.

[0157] In some embodiments, the system further comprises:

[0158] The learning rate optimization module is used to optimize the learning rate of the Transformer-based neural network model according to the cosine annealing algorithm.

[0159] like Figure 6 As shown, an embodiment of the present application also provides an electronic device, the electronic device 10 includes a memory 20 and a processor 30, the memory 20 stores a computer program, and when the computer program is executed by the processor 30, the processor 30 executes the steps of the method for constructing a digital twin of a converter valve power module as in any of the above embodiments.

[0160] An embodiment of the present application also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed, the steps of the method for constructing a digital twin of a converter valve power module as in any of the above embodiments are implemented.

[0161] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described system, electronic device and computer storage medium can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0162] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0163] In several embodiments provided by the present invention, it is understood that each box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and a part of a module, a program segment or a code includes one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from that marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved.

[0164] In the several embodiments provided by the present invention, it should be understood that the disclosed systems, electronic devices, computer storage media and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0165] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0166] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0167] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for executing all or part of the steps of the method described in each embodiment of the present invention through a computer device (which can be a personal computer, server, or network device, etc.). The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (full name in English: Read-Only Memory, English abbreviation: ROM), random access memory (full name in English: Random Access Memory, English abbreviation: RAM), disk or optical disk and other media that can store program codes.

[0168] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for constructing a digital twin of a converter valve power module, characterized in that: include: According to the physical field simulation data of the converter valve power module, the simulation sample data of the converter valve power module is obtained; wherein the simulation sample data includes operating condition parameters; Determine a training sample set according to the simulation sample data; The training sample set is input into a Transformer-based neural network model for training to obtain a simulation digital twin prediction model of the converter valve power module. The simulation digital twin prediction model is used to output the physical field data of the converter valve power module according to the input operating condition parameters of the converter valve power module.

2. The method for constructing a digital twin of a converter valve power module according to claim 1, characterized in that: The obtaining, according to the physical field simulation data of the converter valve power module, simulation sample data of the converter valve power module comprises: Performing finite element modeling on the converter valve power module to construct a finite element model of the converter valve power module; Inputting a plurality of operating condition parameters of the converter valve power module into a finite element model of the converter valve power module in sequence to perform finite element analysis, and obtaining physical field data corresponding to the plurality of operating condition parameters respectively; Simulation sample data of the converter valve power module are constructed based on the multiple operating condition parameters and the physical field data respectively corresponding to the multiple operating condition parameters.

3. The method for constructing a digital twin of a converter valve power module according to claim 1, characterized in that: The method further comprises: inputting the training sample set into a Transformer-based neural network model for training to obtain a simulation digital twin prediction model of the converter valve power module, wherein the simulation digital twin prediction model is used to output the physical field data of the converter valve power module according to the input operating condition parameters of the converter valve power module; The training sample set is subjected to standardization preprocessing.

4. The method for constructing a digital twin of a converter valve power module according to claim 1, characterized in that: The Transformer-based neural network model includes an embedding layer, an encoder layer, an average pooling layer, a hidden linear layer and an output linear layer connected in sequence.

5. The method for constructing a digital twin of a converter valve power module according to claim 4, characterized in that: The step of inputting the training sample set into a Transformer-based neural network model for training to obtain a simulation digital twin prediction model of the converter valve power module comprises: Performing an embedding operation on a plurality of input features of the training sample set through the embedding layer to obtain a feature embedding vector matrix, wherein the embedding operation is used to convert the plurality of input features into vectors of equal length; Determine a position encoding of each input feature according to the parity feature of the position of each input feature in the feature embedding vector matrix; Performing a serial operation on the position encoding of each of the input features to obtain a position vector matrix; Performing a sum operation on the feature embedding vector matrix and the position vector matrix to obtain a feature embedding matrix; Inputting the feature embedding matrix into the encoder layer, and performing feature extraction on the feature embedding matrix based on self-attention to obtain embedded features; Performing pooling processing on the embedded features through the average pooling layer to obtain embedded pooling features; Performing a hidden linear transformation on the embedded pooling features, and mapping the embedded pooling features to a new feature space to obtain embedded linear features; The embedded linear features are output to obtain output features corresponding to each of the input features.

6. The method for constructing a digital twin of a converter valve power module according to claim 5, characterized in that: The encoder layer includes a plurality of stacked encoding layers, each encoding layer including a multi-head attention layer, a first residual link and a layer normalization layer, a feedforward neural network layer, and a second residual link and a layer normalization layer connected in sequence; The step of inputting the feature embedding matrix into the encoder layer and extracting features from the feature embedding matrix based on self-attention to obtain embedded features includes: Inputting the feature embedding matrix into a first encoding layer of the encoder layer, and in the first encoding layer, performing a multi-head attention operation on the feature embedding matrix based on self-attention through the multi-head attention layer; Performing residual link and layer normalization operations on the feature embedding matrix after the multi-head attention operation through the first residual link and layer normalization layer; Inputting the feature embedding matrix after the residual link and layer normalization operations into the feedforward neural network layer, and extracting the initial embedding features of the feature embedding matrix after the residual link and layer normalization operations; Performing residual link and layer normalization operations on the initial embedded features through the second residual link and layer normalization layer, and outputting the normalized embedded features as input features of a next encoding layer adjacent to the first encoding layer; Based on the next encoding layer, the step of inputting the feature embedding matrix into the first encoding layer of the encoder layer is performed, in the first encoding layer, through the multi-head attention layer, based on self-attention, on the feature embedding matrix, until all encoding layers of the encoder layer complete the multi-head attention operation on the feature embedding matrix, and then output the embedded features.

7. The method for constructing a digital twin of a converter valve power module according to claim 1, characterized in that: Also includes: The learning rate of the Transformer-based neural network model is optimized according to the cosine annealing algorithm.

8. A digital twin construction system for a converter valve power module, characterized in that: include: A data acquisition module, used to obtain simulation sample data of the converter valve power module according to the physical field simulation data of the converter valve power module; wherein the simulation sample data includes operating condition parameters; A sample determination module, used to determine a training sample set according to the simulation sample data; A model building module is used to input the training sample set into a Transformer-based neural network model for training to obtain a simulated digital twin prediction model of the converter valve power module. The simulated digital twin prediction model is used to output the physical field data of the converter valve power module based on the input operating condition parameters of the converter valve power module.

9. An electronic device, characterized in that: The electronic device includes a memory and a processor, wherein a computer program is stored in the memory. When the computer program is executed by the processor, the processor executes the steps of the method for constructing a digital twin of a converter valve power module as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed, the steps of the method for constructing a digital twin of a converter valve power module as described in any one of claims 1 to 7 are implemented.

Images