Converter transformer modal prediction method, system, equipment and medium

Abstract

The invention discloses a converter transformer modal prediction method, system, equipment and medium, which are applied to the field of vibration modal prediction, and comprises the following steps: establishing a theoretical frequency response relationship between an excitation vector and a response vector based on an oscillatory differential equation of a target converter transformer; training the neural network model through the excitation data and response data of the target converter transformer under a plurality of operation conditions, constructing a first converter transformer modal prediction model, and performing optimization to obtain a second converter transformer modal prediction model, and collecting real-time operation parameters of the target converter transformer and inputting the real-time operation parameters into the second converter transformer modal prediction model to obtain corresponding modal prediction data. According to the method, the modal prediction efficiency is remarkably improved on the basis of ensuring the physical rationality by constructing the hybrid framework combining the vibration analysis model and the deep learning, and the prediction precision and generalization ability of the model among transformers with different sizes are effectively enhanced by introducing geometric and structural mechanics similarity constraints.

Description

Technical Field

[0001] This invention relates to the field of vibration mode prediction technology, and in particular to a method, system, device and medium for predicting the modes of a converter transformer. Background Technology

[0002] With the continuous expansion of power system scale and the rapid development of ultra-high voltage direct current transmission technology, the requirements for vibration and noise control and operational reliability of converter transformers, as core equipment, are increasing. Modal analysis, as an important means of studying transformer vibration characteristics, can provide key basis for vibration mechanism explanation, fault diagnosis, and vibration reduction design.

[0003] Currently, transformer modal analysis primarily relies on the finite element method (FEM). This method calculates natural frequencies and mode shapes by establishing a three-dimensional physical model and combining multi-physics coupling of electromagnetic fields, structural mechanics, and acoustics. However, the FEM suffers from high computational resource consumption, complex modeling, and long processing time. Furthermore, it requires remodeling for transformers of different models and sizes, making it difficult to meet the demand for rapid prediction. The limitations of traditional methods become even more pronounced in practical engineering applications where rapid modal parameter evaluation of transformers under various operating conditions and with different structural dimensions is crucial.

[0004] Therefore, how to effectively predict the modal parameters of converter transformers has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] This invention provides a method, system, device, and medium for predicting the modal parameters of a converter transformer, enabling rapid and accurate prediction of transformer vibration characteristics.

[0006] To address the aforementioned technical problems, embodiments of the present invention provide a converter transformer mode prediction method, comprising: Based on the vibration differential equation of the target converter transformer, a vibration analytical model of modal superposition is constructed by Laplace transform to establish the theoretical frequency response relationship between the excitation vector and the response vector of the target converter transformer. Using the theoretical frequency response relationship as the learning objective, the neural network model is trained by the excitation and response data of the target converter transformer under several operating conditions, thereby constructing the first converter transformer modal prediction model. The first converter transformer modal prediction model is jointly optimized based on geometric and structural mechanics similarity constraints to obtain the second converter transformer modal prediction model. The real-time operating parameters of the target converter transformer are collected and input into the second converter transformer modal prediction model to obtain the corresponding modal prediction data.

[0007] Furthermore, based on the vibration differential equation of the target converter transformer, a modal superposition vibration analytical model is constructed through Laplace transform to establish the theoretical frequency response relationship between the excitation vector and the response vector of the target converter transformer, including: A multi-degree-of-freedom vibration differential equation for the target converter transformer is established, wherein the vibration differential equation includes a mass matrix, a stiffness matrix, and a damping matrix; Perform a Laplace transform on the vibration differential equation to obtain the generalized impedance matrix of the power system where the target converter transformer is located in the frequency domain, and invert the generalized impedance matrix to obtain the generalized admittance matrix. Based on the generalized admittance matrix, a frequency response function matrix is ​​constructed to characterize the quantitative mapping relationship between the excitation vector and the response vector of the target converter transformer, which serves as the frequency response relationship.

[0008] Furthermore, the step of using the theoretical frequency response relationship as the learning objective, and training the neural network model with excitation and response data of the target converter transformer under several operating conditions to construct a first converter transformer modal prediction model includes: Acquire the excitation and response data of the target converter transformer under various operating conditions; The excitation data and response data are preprocessed, and the excitation matrix and response matrix are constructed based on the preprocessed data. Using the excitation matrix as input and the response matrix as output, a pre-defined deep neural network is trained under supervision. The supervised training process is designed as follows: the network output is calculated by forward propagation, the error gradient between the predicted output and the true response matrix is ​​calculated by backpropagation algorithm, and the network weight parameters are iteratively updated by gradient descent method until the model prediction error converges to a preset threshold, thereby obtaining the first converter transformer modal prediction model.

[0009] Further, the joint optimization of the first converter transformer modal prediction model based on geometric and structural mechanics similarity constraints to obtain the second converter transformer modal prediction model includes: Calculate the geometric scaling factor between the original model and the scaled-down model of the target converter transformer, and introduce the scaling factor as an additional input feature into the modal prediction model of the first converter transformer; Based on the structural mechanics similarity criterion, the scaling relationship between the original model and the scaled-down model in terms of material density, elastic modulus and stiffness parameters is established. Based on the scaling relationship, a composite loss function is constructed that includes a geometric similarity error term, a structural mechanical similarity error term, and a prediction output error term; The gradient descent algorithm is used to optimize and iterate the composite loss function, update the network parameters of the first converter transformer modal prediction model, and obtain the second converter transformer modal prediction model that satisfies the constraints.

[0010] Further, the step of collecting the real-time operating parameters of the target converter transformer and inputting them into the second converter transformer modal prediction model to obtain the corresponding modal prediction data includes: Obtain the operating status data and corresponding structural dimension data of the target converter transformer; The operating status data and structural dimension data are input into the second converter transformer modal prediction model; The natural frequency, modal damping ratio, and mode shape distribution data of the target converter transformer are obtained through forward calculation of the second converter transformer modal prediction model.

[0011] Another embodiment of the present invention provides a converter transformer mode prediction system, comprising: The response relationship construction module is used to construct a vibration analytical model of modal superposition based on the vibration differential equation of the target converter transformer through Laplace transform, so as to establish the theoretical frequency response relationship between the excitation vector and the response vector of the target converter transformer. The model building module is used to train the neural network model with the theoretical frequency response relationship as the learning target and the excitation data and response data of the target converter transformer under several operating conditions, so as to build the first converter transformer modal prediction model. The model optimization module is used to perform joint optimization of the first converter transformer modal prediction model based on geometric and structural mechanics similarity constraints to obtain the second converter transformer modal prediction model. The modal prediction module is used to collect the real-time operating parameters of the target converter transformer and input them into the second converter transformer modal prediction model to obtain the corresponding modal prediction data.

[0012] Furthermore, the response relationship construction module is specifically used for: A multi-degree-of-freedom vibration differential equation for the target converter transformer is established, wherein the vibration differential equation includes a mass matrix, a stiffness matrix, and a damping matrix; Perform a Laplace transform on the vibration differential equation to obtain the generalized impedance matrix of the power system where the target converter transformer is located in the frequency domain, and invert the generalized impedance matrix to obtain the generalized admittance matrix. Based on the generalized admittance matrix, a frequency response function matrix is ​​constructed to characterize the quantitative mapping relationship between the excitation vector and the response vector of the target converter transformer, which serves as the frequency response relationship.

[0013] Furthermore, the model building module is specifically used for: Acquire the excitation and response data of the target converter transformer under various operating conditions; The excitation data and response data are preprocessed, and the excitation matrix and response matrix are constructed based on the preprocessed data. Using the excitation matrix as input and the response matrix as output, a pre-defined deep neural network is trained under supervision. The supervised training process is designed as follows: the network output is calculated by forward propagation, the error gradient between the predicted output and the true response matrix is ​​calculated by backpropagation algorithm, and the network weight parameters are iteratively updated by gradient descent method until the model prediction error converges to a preset threshold, thereby obtaining the first converter transformer modal prediction model.

[0014] Another embodiment of the present invention provides a computer device including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the converter transformer mode prediction method as described above.

[0015] In another embodiment of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program, wherein when the device containing the computer-readable storage medium executes the computer program, the converter transformer mode prediction method described above is implemented.

[0016] Compared with the prior art, the beneficial effects of the embodiments of the present invention are at least one of the following: By establishing a hybrid framework of vibration analytical model and deep learning, combining physical laws with data-driven methods, the model avoids the complex modeling process and huge computational resource consumption of traditional finite element methods while ensuring the physical rationality of the prediction results. This significantly improves the efficiency of modal prediction. By introducing geometric similarity constraints and structural mechanics similarity constraints, the model can accurately capture the dynamic similarity relationship between transformers of different sizes, effectively improving the prediction accuracy and generalization ability of the model under cross-scale conditions. The final optimized prediction model can quickly and accurately output complete modal parameters, including natural frequency, modal damping ratio and mode shape, based on the transformer's operating parameters and structural characteristics. Attached Figure Description

[0017] Figure 1 This is a flowchart of the steps of the converter transformer mode prediction method in one embodiment of the present invention; Figure 2 This is a flowchart of converter transformer mode prediction in one embodiment of the present invention; Figure 3 This is a simplified model diagram of a converter transformer in one embodiment of the present invention; Figure 4 This is a BH curve diagram of silicon steel sheet corresponding to a reference converter transformer in one embodiment of the present invention; Figure 5 This is a modal distribution diagram of the core and winding in one embodiment of the present invention, wherein: Figure 5 (a) is the modal distribution diagram of the first order of the initial model iron core; Figure 5 (b) is the modal distribution diagram of the first order of the iron core of the 1 / 2 model; Figure 5 (c) is the modal distribution diagram of the first order of the iron core of the 1 / 5 model; Figure 5 (d) is the modal distribution diagram of the third order of the initial model iron core; Figure 5 (e) is the modal distribution diagram of the third order of the iron core of the 1 / 2 model; Figure 5 (f) is the modal distribution diagram of the third order of the iron core of the 1 / 5 model; Figure 5 (g) is the modal distribution diagram of the first order winding of the initial model; Figure 5 (h) is the modal distribution diagram of the first order of the 1 / 2 model winding; Figure 5 (i) is the modal distribution diagram of the first order of the 1 / 5 model winding; Figure 6 This is a structural block diagram of a converter transformer mode prediction system according to one embodiment of the present invention; Figure 7 A structural diagram of a computer device provided in an embodiment of the present invention; Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] In the description of this application, the terms "first," "second," "third," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first," "second," "third," etc., may explicitly or implicitly include one or more of that feature. In the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0020] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to the internal communication between two components. The terms "vertical," "horizontal," "left," "right," "upper," "lower," and similar expressions used herein are for illustrative purposes only and do not indicate or imply that the device or component referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as limiting the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0021] In the description of this application, it should be noted that, unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing specific embodiments only and is not intended to limit the invention. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0022] One embodiment of the present invention provides a method for predicting the modal characteristics of a converter transformer. For details, please refer to [link to relevant documentation]. Figure 1 , Figure 1 The flowchart shown is a step-by-step diagram of the converter transformer mode prediction method in one embodiment of the present invention, including steps S11-S14: S11. Based on the vibration differential equation of the target converter transformer, a vibration analytical model of modal superposition is constructed through Laplace transform to establish the theoretical frequency response relationship between the excitation vector and the response vector of the target converter transformer.

[0023] A multi-degree-of-freedom vibration differential equation for the target converter transformer is established, wherein the vibration differential equation includes a mass matrix, a stiffness matrix, and a damping matrix; Perform a Laplace transform on the vibration differential equation to obtain the generalized impedance matrix of the power system where the target converter transformer is located in the frequency domain, and invert the generalized impedance matrix to obtain the generalized admittance matrix. Based on the generalized admittance matrix, a frequency response function matrix is ​​constructed to characterize the quantitative mapping relationship between the excitation vector and the response vector of the target converter transformer, which serves as the frequency response relationship.

[0024] In specific implementations, a mechanical-dynamic model of the converter transformer can be established, and differential equations can be obtained using the common finite element method. Boundary constraints can then be set under experimental conditions. In engineering practice, a vibration test can be conducted on a specific type of converter transformer on an experimental platform to measure vibration signals at different locations. These data can then be substituted into the established dynamic equations for analysis. A Laplace transform is performed using computer software to obtain the model in the frequency domain, thereby establishing the frequency response function matrix. In practical applications, researchers often choose to place sensors at key locations in the windings and core to improve the reliability of the frequency response function.

[0025] In specific embodiments, the physical structure of the converter transformer can be idealized and simplified, treating windings, core, etc., as elastic elements and discretizing them using the finite element method. Researchers often use commercial finite element software to construct a three-dimensional model during the design phase, inputting material density, elastic modulus, and damping coefficients into this model. By geometrically modeling different components, they are transformed into nodes and elements in multi-degree-of-freedom dynamic equations, thus obtaining the mass matrix, stiffness matrix, and damping matrix. This discretized model can be rapidly iterated in a computer environment, facilitating subsequent solutions and predictions.

[0026] Perform a Laplace transform on the differential equation to obtain the generalized impedance matrix of the system in the frequency domain, and then obtain the generalized admittance matrix.

[0027] During implementation, the aforementioned discretized differential equations are subjected to a Laplace transform using mathematical software or numerical algorithm libraries, transforming the time-domain problem into a frequency-domain problem. Specifically, the collected nodal displacement, velocity, and acceleration signals are input into numerical tools to automatically generate the frequency domain response. By comparing simulation results with experimental data, a relatively accurate generalized impedance matrix can be obtained. The further obtained generalized admittance matrix can be used for subsequent modal feature extraction. For example, when analyzing a 220kV converter transformer, excitation force and response signals can be collected under different operating conditions, and the admittance matrix can be obtained through numerical calculations, laying the foundation for constructing the frequency response function.

[0028] In a specific implementation, the excitation signal can be used as the input and the response signal as the output, establishing a mapping relationship between the two. In practical engineering, researchers typically place accelerometers or displacement sensors on converter transformers to acquire vibration responses and apply external excitation using a hammer or electromagnetic excitation device. The data collected experimentally can form an excitation vector and a response vector. Subsequently, the ratio or transfer function between the two is calculated in numerical analysis software, thus obtaining the frequency response function matrix. This matrix clearly reflects the dynamic relationship between different excitation points and measurement points, providing labeled data for subsequent neural network-based training.

[0029] S12. Using the theoretical frequency response relationship as the learning target, the neural network model is trained by the excitation data and response data of the target converter transformer under several operating conditions to construct the first converter transformer modal prediction model.

[0030] Acquire the excitation and response data of the target converter transformer under various operating conditions; The excitation data and response data are preprocessed, and the excitation matrix and response matrix are constructed based on the preprocessed data. Using the excitation matrix as input and the response matrix as output, a pre-defined deep neural network is trained under supervision. The supervised training process is designed as follows: the network output is calculated by forward propagation, the error gradient between the predicted output and the true response matrix is ​​calculated by backpropagation algorithm, and the network weight parameters are iteratively updated by gradient descent method until the model prediction error converges to a preset threshold, thereby obtaining the first converter transformer modal prediction model.

[0031] By changing operating conditions, such as load current magnitude, external short-circuit impact, or excitation at different voltage levels, excitation and response data under multiple operating conditions can be obtained. In this embodiment, researchers applied current excitation of different amplitudes to the converter transformer on an experimental platform and simultaneously collected response data from the windings and core to form the dataset required for training. Subsequently, the excitation matrix was used as the network input, and the response matrix as the network output, to train the initial model.

[0032] Researchers can use sensor arrays to place measuring points on key components such as windings and cores, and collect signals under different operating conditions. In this example, the converter transformer operates under high and low load conditions, and multiple sets of excitation and response signals are acquired using accelerometers and current sensors to provide real-world operating data for model training.

[0033] The original acquired signal can be filtered to remove interference noise, the data amplitude can be normalized to maintain dimensional consistency, and phase alignment can ensure the consistency of input and output timing. In the embodiment, engineers use signal processing software to process the acquired excitation and response signals, eliminate electromagnetic interference, and unify the amplitude range, thereby forming a high-quality training sample set.

[0034] Researchers can construct multi-layer feedforward neural networks, where the number of nodes in the input and output layers corresponds to the dimensions of the excitation and response data, respectively, and the hidden layers extract features through a fully connected structure. In this embodiment, a three-layer hidden layer network is used, with the number of nodes in each layer set according to empirical values ​​to ensure effective learning of the nonlinear mapping relationship between the input and output.

[0035] Researchers can define the objective function as the average of the sum of squared differences between the elements of the predicted matrix and the true matrix, thereby measuring the accuracy of the model's predictions. In this example, the mean squared error is calculated using validation data from multiple operating conditions, and this error is used as the criterion for determining training convergence, ensuring that the output results are close to the true results.

[0036] Researchers can choose the adaptive gradient descent optimization algorithm and set the initial learning rate and adjustment rules to make the training process more stable. In this example, the Adam optimizer is used for parameter updates, and the learning rate is dynamically adjusted in different training epochs to accelerate model convergence and avoid overfitting.

[0037] The convergence of training can be determined by monitoring the trend of the validation set error in real time. If the error decrease is insufficient for an extended period, training will automatically stop. In this example, researchers set a threshold of 0.1%. When the validation set error fails to decrease for ten consecutive rounds, training stops, and the final preliminary model is output.

[0038] The predicted output is calculated via forward propagation, and the parameters are updated using backpropagation and gradient descent algorithms until the training error converges, resulting in a preliminary converter transformer modal prediction model. In a specific implementation, forward and backpropagation can be performed using batch training, continuously updating the parameters with the training set until the set convergence condition is met. In this example, researchers trained with 1000 samples per batch, achieving convergence after hundreds of iterations, thus obtaining a preliminary model capable of accurately predicting the frequency response function matrix.

[0039] S13. Perform joint optimization on the first converter transformer modal prediction model based on geometric and structural mechanics similarity constraints to obtain the second converter transformer modal prediction model.

[0040] Calculate the geometric scaling factor between the original model and the scaled-down model of the target converter transformer, and introduce the scaling factor as an additional input feature into the modal prediction model of the first converter transformer; Based on the structural mechanics similarity criterion, the scaling relationship between the original model and the scaled-down model in terms of material density, elastic modulus and stiffness parameters is established. Based on the scaling relationship, a composite loss function is constructed that includes a geometric similarity error term, a structural mechanical similarity error term, and a prediction output error term; The gradient descent algorithm is used to optimize and iterate the composite loss function, update the network parameters of the first converter transformer modal prediction model, and obtain the second converter transformer modal prediction model that satisfies the constraints.

[0041] In a specific implementation, physical constraints can be added during the training phase to ensure that the model conforms to the theoretical principles of geometric and mechanical similarity based on data-driven principles. In this example, researchers compare the geometric parameters of the original model and the scaled-down model, and introduce scaling factors for density and elastic modulus to correct the prediction results, thereby obtaining the final model.

[0042] Geometric similarity constraints are introduced into the preliminary converter transformer modal prediction model, and the length, area and volume scaling factors of the scaled-down model and the original model are used as constraint variables input. In a specific implementation, the dimensions of different scaled-down models compared to the original model can be calculated, and the scaling factor can be input into the network as an additional feature. In an example, researchers established 1 / 2 and 1 / 5 scaled-down models, calculated the ratio of their length to volume, and used this as additional input for training.

[0043] In the geometric similarity constraint, the scaling factor is calculated for the length, area and volume parameters of the original model and the scaled-down model of the converter transformer, and the normalized scaling factor is used as an additional input variable to the prediction network. In a specific implementation, researchers can first perform dimensionless processing on the scaling coefficients of each parameter to eliminate differences between different dimensions and ensure the uniformity of the network input data. In this example, the volume scaling coefficients are uniformly normalized to between 0 and 1 and then input into the prediction network to improve the model learning effect.

[0044] In the structural mechanics similarity constraint, based on the scaling law of material density, elastic modulus and moment of inertia, the modal frequencies predicted by the preliminary model are proportionally corrected so that the natural frequencies of the scaled-down model and the original model are inversely proportional. In specific implementations, researchers can combine scaling relationships to proportionally correct the frequency prediction results. In this example, for a 1 / 5 scaled-down model, the frequency prediction values ​​are proportionally converted to obtain the modal frequencies corresponding to the original model, ensuring consistency in predictions across different scales.

[0045] Structural mechanics similarity constraints are introduced into the preliminary converter transformer modal prediction model, and the prediction results are corrected based on the scaling rules of mass density, elastic modulus and stiffness parameters. In specific implementations, researchers can use known material mechanical parameters to modify the model output according to a scaling relationship, making the prediction results more consistent with actual physical laws. In this example, for different core materials, the frequency prediction accuracy of the model is optimized by adjusting the ratio of the elastic modulus parameter.

[0046] By integrating the aforementioned geometric and mechanical constraints into the loss function, the preliminary model is optimized and iterated to output the final converter transformer modal prediction model; In a specific implementation, researchers can combine geometric constraint errors and mechanical constraint errors with conventional prediction errors to form a new optimization objective function. In this example, different weights are assigned to the three types of errors, allowing training to optimize while maintaining both prediction accuracy and physical consistency, ultimately outputting a reliable prediction model.

[0047] In the process of constructing the loss function, the geometric similarity error, structural mechanical similarity error and prediction output error are weighted and combined. The geometric similarity error is obtained by calculating the deviation between the scaling factor and the prediction output, and the structural mechanical similarity error is obtained by calculating the difference between the frequency scaling formula and the prediction frequency. Finally, together with the prediction output error, they are used as components of the comprehensive optimization objective function to drive the model to converge iteratively into the final converter transformer modal prediction model. In specific implementations, researchers can continuously adjust the three types of error weights to ensure the model maintains high accuracy across different scales. In one example, for scaled-down models with larger errors, physical constraint error weights are increased to ensure the final model remains reliable in cross-scale applications.

[0048] S14. Collect the real-time operating parameters of the target converter transformer and input them into the second converter transformer modal prediction model to obtain the corresponding modal prediction data.

[0049] Obtain the operating status data and corresponding structural dimension data of the target converter transformer; The operating status data and structural dimension data are input into the second converter transformer modal prediction model; The natural frequency, modal damping ratio, and mode shape distribution data of the target converter transformer are obtained through forward calculation of the second converter transformer modal prediction model.

[0050] In a specific implementation, researchers can input the operating voltage, load conditions, and geometric parameters of the target converter transformer into the final model to obtain complete modal parameter prediction results. In this example, for an ultra-high voltage converter transformer, its operating data is input, and the model outputs the natural frequency, damping ratio, and mode shape for reference by design and maintenance departments.

[0051] The method further includes: testing the final converter transformer modal prediction model obtained through the aforementioned steps, the testing including: After inputting the target converter transformer's operating parameters and structural characteristics, the natural frequency, damping ratio, and mode shape of the model output are collected. The predicted output is compared with the finite element simulation results or measured data to calculate the average absolute percentage error of the prediction results, and based on this, it is determined whether the prediction accuracy meets the preset threshold. If the prediction accuracy does not reach the threshold, the similarity constraint weight coefficients are adjusted and the training steps are re-executed to correct the final converter transformer modal prediction model.

[0052] In specific implementations, researchers can compare the model's predicted values ​​with different scaled-down models or actual experimental data to verify the model's reliability through error analysis. In one example, when the prediction error of some scaled-down models exceeds a threshold, researchers adjust the constraint coefficients and retrain the model to ensure that the final model meets the accuracy requirements.

[0053] In elastic systems with continuously distributed mass, such as converter transformers, their inherent dynamic behavior theoretically needs to be described with infinite degrees of freedom. When using the finite element method for modal analysis, it is necessary to first establish its geometric and physical model, and then discretize it into a finite number of elements through mesh generation, thus transforming it into a discretized system with n degrees of freedom.

[0054] According to classical mechanics, the vibration behavior of each node in this discrete system can be described by the following differential equation: In the formula, It is an N-dimensional excitation force vector; , , These are the N-dimensional displacement, velocity, and acceleration response vectors, respectively; , , These are the mass, stiffness, and damping matrices of the structure, respectively, and are usually real symmetric N-order matrices.

[0055] Assuming the initial state of the system is zero, applying the Laplace transform to both sides of equation (1) yields a matrix algebraic equation with the complex number s as the variable: (2) The method used in this embodiment is to predict the matrix using the frequency response function of all measurement points. , , Let the matrix in equation (2) be: (3) The above equation reflects the dynamic characteristics of the system and is called the system dynamic matrix or generalized impedance matrix. Its inverse is: (4) Equation (4) is the generalized admittance matrix, which is also the frequency response function matrix. From equation (2), we know that: (5) By letting s = jω in equation (5), we can obtain the relationship between the system's output (response vector X(ω)) and input (excitation vector F(ω)) in the frequency domain: (6) In the formula, H(ω) is the frequency response function matrix. The element in the i-th row and j-th column of the H(ω) matrix is ​​equal to the ratio of the response at coordinate i to the excitation force when the excitation force is zero only at coordinate j (the excitation force at other coordinates is zero).

[0056] (7) Letting s = jω in equation (3), we can obtain the impedance matrix: (8) Using the weighted orthogonality of real symmetric matrices, we have: , Where the matrix This is called the mode shape matrix. Let the damping matrix C also satisfy the mode shape orthogonality relation: Substituting into equation (8), we get: (9) In the formula, .

[0057] (10) (11) In equation (11), , m r k r These are referred to as the r-th modal mass and modal stiffness, respectively. ω r ξ r and φ r These are respectively called the r-th modal frequency, modal damping ratio, and mode shape.

[0058] The frequency response of an N-degree-of-freedom system is equal to the linear superposition of the frequency responses of N single-degree-of-freedom systems. To determine all modal parameters ω... r ξ r and φ r (r=1, 2, ..., N), only one column of the frequency response matrix needs to be obtained: H(ω) measured at each point corresponding to excitation at one point; or one row: H(ω) measured at each point corresponding to excitation at each point in sequence. H(ω) can be obtained through the response vector X(ω) and the input excitation vector F(ω).

[0059] H(ω) is related to X(ω) and F(ω), and the system has complex characteristics. Therefore, this embodiment uses a deep neural network to calculate the matrix H(ω). The network has a multi-layer structure, consisting of an input layer, multiple hidden layers, and an output layer. Setting multiple hidden layers is beneficial for discovering more complex nonlinear relationships between the input and output quantities.

[0060] Deep neural networks consist of multiple neural networks with fully connected layers, and use forward and backward propagation algorithms to update system parameters. In the constructed model, the matrix relationship between input and output variables is as follows: (12) In the formula, For the input variable matrix, For the first The output value of the layered neural network, For the first in a deep neural network Activation function of layer network For the first The weight matrix of the layer network, This is the deviation matrix.

[0061] During the training of the preliminary converter transformer modal prediction model, the network parameters are continuously updated iteratively. The iteration stops when the system objective function reaches convergence. The objective function for the training process is: (13) In the formula, The number of samples chosen for training The normalized input variable matrix, The normalized output variable matrix, In this context, L represents the number of layers in the neural network.

[0062] In this embodiment, the ReLU activation function is chosen as the activation function between deep neural network layers because the traditional Sigmoid activation function is saturated at both ends, and information is easily lost during propagation. The ReLU activation function has piecewise characteristics and stronger learning and optimization capabilities. Its mathematical expression is: (14) To study the vibration of converter transformers, it is necessary to start with parameters directly related to vibration and perform similarity analysis to ensure that the vibration characteristics of the model are consistent before and after similarity analysis. Based on this, a scaled-down model can be designed. When scaling down the model, the length, width, and radius are reduced proportionally according to the scaling factor. To ensure the vibration intensity and distribution of the equivalent core and windings, the transformer vibration modes need to remain similar.

[0063] Based on the principle of similarity, the overall dimensions of the converter transformer should be reduced proportionally; that is, the overall geometric length *l* of length, width, height, and radius should be changed to *l'=kl*. Therefore, the corresponding overall geometric area *S* becomes *S'=k*. 2 S, the overall volume V becomes V'=k 3 V.

[0064] When converter transformers are designed on a scaled-down basis, the model's mass and volume significantly affect its overall modal distribution, easily altering the original resonance effect and thus influencing the final vibration characteristic distribution. For the mass M, since the material density of the scaled model remains constant, the volume is reduced to the original k. 3 Therefore, M' becomes: (15) The elements of the stiffness matrix are proportional to the stiffness of the structure. The general form of the stiffness K is K = (E*S) / l, where E is the elastic modulus. Based on geometric similarity requirements, the following analysis can be performed: Tensile / compressive stiffness K=(E*S) / l: Bending stiffness Where the moment of inertia I scales k 4 .

[0065] The stiffness coefficient matrix K should satisfy the following requirements before and after similarity: (16) Therefore, the natural frequency ω' of the proportional model becomes: (17) From the above, the frequency response function matrix of the original model can be determined, and based on the matrix coefficient relationships, the r-th order modal frequency, modal damping ratio, and mode shape of the original model can be obtained. Based on this, the relevant parameters of the scaled-down prototype model are determined according to the proportional relationship between the scaled-down prototype and the original model. The overall framework of the converter transformer modal prediction method based on deep neural networks is as follows: Figure 2 As shown.

[0066] First, a multiphysics coupled simulation model is established based on the actual operating parameters and structural characteristics of the converter transformer. Applying different excitations yields different response conditions at various points. The excitation matrix and response matrix are used as inputs and outputs to train a preliminary converter transformer modal prediction model, obtaining the system's frequency response function matrix, and thus the transformer's modal frequencies. Combining the preliminary converter transformer modal prediction model with similarity theory, the frequencies of different modes in a scaled-down model of the converter transformer can be obtained.

[0067] Based on the finite element simulation platform and according to the structure and operating parameters of the converter transformer, a three-dimensional physical model of the converter transformer was constructed. The model effect is as follows: Figure 3 As shown. During the modeling process, the converter transformer structure was idealized, ignoring fixtures and other structures, and corresponding fixed constraints were applied to the windings and core as boundary conditions. The field-circuit coupling method was used to couple the magnetic field model with the external circuit. Material properties were assigned to each computational domain, the Lorentz force was applied as a body load to the windings, and the magnetostriction effect was added to the core domain to study the stress on the windings and core respectively. In the magnetic field physics module, voltage source excitation was added to each winding, and its input and output ports were set. The core material is the main contributor to the nonlinearity of the converter transformer. The B and H values ​​at each location need to be defined based on the BH curve of the silicon steel sheet to ensure the reliability of the model. The BH curve used for definition is shown below. Figure 4 As shown in Tables 1 and 2.

[0068] Table 1 Model Parameter Configuration Table Table 2. Converter Transformer Model Parameter Table To save time and computational resources, the physical model of the converter transformer was divided into half models along the longitudinal section. The half-model was meshed, and the windings, core, and gaskets were dissected using ultra-fine free tetrahedral structures. Other regions were subjected to planar triangular and sweep operations. The tolerance was set to 0.1, and the number of degrees of freedom was 592,163. The established multiphysics coupling model was used to calculate the excitation and response matrices, providing data for training the deep neural network model.

[0069] The computational accuracy of the model is demonstrated by the Mean Absolute Percentage Error (MAPE) of the inherent frequencies of the core and windings. The mathematical expression for MAPE is: (18) In the formula, n is the number of samples selected. and These are the estimated and actual values ​​of the natural frequency of the core or winding.

[0070] Table 3. Distribution of natural frequencies of the core in the original model Table 4. Distribution of natural frequencies of the original model windings As can be seen from Tables 3 and 4, the errors in the core and winding natural frequencies calculated by the converter transformer modal prediction model based on deep neural networks proposed in this embodiment are both less than 0.5%, which meets the application requirements.

[0071] To verify the effectiveness of the converter transformer modal prediction method proposed in this embodiment, based on the similarity criterion, 1 / 2 and 1 / 5 scale models of the converter transformer were established respectively, and the geometric parameters are shown in Table 5.

[0072] Table 5 Operating Parameters To improve the versatility of modal analysis, this embodiment analyzes the initial model, the 1 / 2 scaled-down model, and the 1 / 5 scaled-down model, and obtains the first 6 natural frequency distributions of the model core and windings under the three geometric dimensions, as detailed in Tables 6 and 7. By comparing the modal shapes of the model under the three geometric dimensions, it is found that the natural frequencies are inversely proportional to the similarity coefficients, and this law follows the derivation of formula (17).

[0073] Table 6. Natural frequency distribution of the core in the three models Table 7. Natural Frequency Distribution of Windings in Three Models Based on the above research, this embodiment uses a deep neural network-based modal prediction method for converter transformer models to calculate the 1 / 2 scaled-down model and the 1 / 5 scaled-down model. The winding and core modes of the initial model, the 1 / 5 model, and the 1 / 2 model are as follows: Figure 5 As shown, as the model scale decreases proportionally, the modal distribution pattern of the model does not change significantly, and the natural frequencies of models of different sizes follow the load similarity criterion.

[0074] Table 8 Comparison of Natural Frequency Distribution of Core in Scaled-Down Model Table 9 Comparison of Natural Frequency Distribution of Windings in Scaled-Down Models The obtained natural frequencies are shown in Tables 8 and 9. The errors of both models are less than 5%, meeting the requirements for engineering applications.

[0075] The converter transformer modal prediction method of this invention combines physical laws with data-driven methods by establishing a hybrid framework of vibration analytical model and deep learning. While ensuring the physical rationality of the prediction results, it avoids the complex modeling process and huge computational resource consumption of traditional finite element methods, significantly improving the efficiency of modal prediction. By introducing geometric similarity constraints and structural mechanics similarity constraints, the model can accurately capture the dynamic similarity relationship between transformers of different sizes, effectively improving the prediction accuracy and generalization ability of the model under cross-scale operating conditions. The finally constructed optimized prediction model can quickly and accurately output complete modal parameters, including natural frequency, modal damping ratio and mode shape, based on transformer operating parameters and structural characteristics.

[0076] This invention also provides a converter transformer mode prediction system for performing the converter transformer mode prediction method described above. Figure 6 This is a structural block diagram of a converter transformer mode prediction system according to an embodiment of the present invention. The system includes: The response relationship construction module 21 is used to construct a vibration analytical model of modal superposition based on the vibration differential equation of the target converter transformer through Laplace transform, so as to establish the theoretical frequency response relationship between the excitation vector and the response vector of the target converter transformer. Model building module 22 is used to train the neural network model with the theoretical frequency response relationship as the learning target and the excitation data and response data of the target converter transformer under several operating conditions to build the first converter transformer modal prediction model. The model optimization module 23 is used to perform joint optimization of the first converter transformer modal prediction model based on geometric and structural mechanics similarity constraints to obtain the second converter transformer modal prediction model. The modal prediction module 24 is used to collect the real-time operating parameters of the target converter transformer and input them into the second converter transformer modal prediction model to obtain the corresponding modal prediction data.

[0077] Furthermore, the response relationship construction module is specifically used for: A multi-degree-of-freedom vibration differential equation for the target converter transformer is established, wherein the vibration differential equation includes a mass matrix, a stiffness matrix, and a damping matrix; Perform a Laplace transform on the vibration differential equation to obtain the generalized impedance matrix of the power system where the target converter transformer is located in the frequency domain, and invert the generalized impedance matrix to obtain the generalized admittance matrix. Based on the generalized admittance matrix, a frequency response function matrix is ​​constructed to characterize the quantitative mapping relationship between the excitation vector and the response vector of the target converter transformer, which serves as the frequency response relationship.

[0078] Furthermore, the model building module is specifically used for: Acquire the excitation and response data of the target converter transformer under various operating conditions; The excitation data and response data are preprocessed, and the excitation matrix and response matrix are constructed based on the preprocessed data. Using the excitation matrix as input and the response matrix as output, a pre-defined deep neural network is trained under supervision. The supervised training process is designed as follows: the network output is calculated by forward propagation, the error gradient between the predicted output and the true response matrix is ​​calculated by backpropagation algorithm, and the network weight parameters are iteratively updated by gradient descent method until the model prediction error converges to a preset threshold, thereby obtaining the first converter transformer modal prediction model.

[0079] The technical features and effects of the system proposed in the embodiments of the present invention are the same as those of the method proposed in the embodiments of the present invention, and will not be repeated here. Each module in the above system can be implemented entirely or partially through software, hardware, or a combination thereof. Each module can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0080] See Figure 7 This is a structural block diagram of a computer device provided in an embodiment of the present invention. The computer device includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the steps described in the above embodiment of the converter transformer mode prediction method, for example... Figure 1 The steps S11 to S14 described above; or, when the processor executes the computer program, it implements the functions of each module in the above-described device embodiments, such as modules 21 to 24 of the converter transformer mode prediction system.

[0081] For example, the computer program may be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the computer device.

[0082] The computer device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the schematic diagram is merely an example of a computer device and does not constitute a limitation on the computer device. It may include more or fewer components than illustrated, or combine certain components, or different components. For example, the computer device may also include input / output devices, network access devices, buses, etc.

[0083] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the computer device, connecting various parts of the computer device via various interfaces and lines.

[0084] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the computer device by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0085] If the modules integrated into the computer device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0086] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0087] Accordingly, embodiments of the present invention provide a computer-readable storage medium, the computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform steps in the converter transformer mode prediction method of the above embodiments, for example... Figure 1 Steps S11 to S14 as described above.

[0088] In summary, compared with the prior art, the converter transformer mode prediction method, system, computer equipment, and computer-readable storage medium provided by the embodiments of the present invention have the following beneficial effects: By establishing a hybrid framework of vibration analytical model and deep learning, combining physical laws with data-driven methods, the model avoids the complex modeling process and huge computational resource consumption of traditional finite element methods while ensuring the physical rationality of the prediction results. This significantly improves the efficiency of modal prediction. By introducing geometric similarity constraints and structural mechanics similarity constraints, the model can accurately capture the dynamic similarity relationship between transformers of different sizes, effectively improving the prediction accuracy and generalization ability of the model under cross-scale conditions. The final optimized prediction model can quickly and accurately output complete modal parameters, including natural frequency, modal damping ratio and mode shape, based on the transformer's operating parameters and structural characteristics.

[0089] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method of modal prediction for a converter transformer, characterized by, The method comprises the steps of: Based on the vibration differential equation of the target converter transformer, the modal superposition vibration analytical model is constructed by Laplace transform to establish the theoretical frequency response relationship between the excitation vector and the response vector of the target converter transformer; Taking the theoretical frequency response relationship as the learning target, the neural network model is trained by the excitation data and response data of the target converter transformer under several operating conditions to construct a first converter transformer modal prediction model; The first converter transformer modal prediction model is optimized based on the geometric and structural mechanics similarity constraints to obtain a second converter transformer modal prediction model; Collect the real-time operating parameters of the target converter transformer and input them into the second converter transformer modal prediction model to obtain the corresponding modal prediction data.

2. The converter transformer modal prediction method of claim 1, wherein, The method comprises the steps of: A multi-degree-of-freedom vibration differential equation of the target converter transformer is established, and the vibration differential equation includes a mass matrix, a stiffness matrix and a damping matrix; Performing Laplace transform on the vibration differential equation, the generalized impedance matrix of the power system in which the target converter transformer is located in the frequency domain is obtained, and the generalized admittance matrix is obtained by inverting the generalized impedance matrix; Based on the generalized admittance matrix, a frequency response function matrix representing the quantitative mapping relationship between the excitation vector and the response vector of the target converter transformer is constructed as the frequency response relationship.

3. The converter transformer modal prediction method of claim 1, wherein, The method comprises the steps of: Obtaining excitation data and response data of the target converter transformer under multiple operating conditions; Pretreating the excitation data and response data, and constructing an excitation matrix and a response matrix according to the pretreated data; Supervised training of a preset deep neural network is performed by taking the excitation matrix as the input and the response matrix as the output, the supervised training process is designed as follows: calculating the network output by forward propagation, calculating the error gradient between the predicted output and the true response matrix by back propagation algorithm, and updating the network weight parameters by gradient descent method until the model prediction error converges to a preset threshold, thereby obtaining the first converter transformer modal prediction model.

4. The converter transformer modal prediction method of claim 1, wherein, The method comprises the steps of: Calculating the geometric scaling coefficient between the original model and the scaled model of the target converter transformer, and introducing the scaling coefficient as an additional input feature into the first converter transformer modal prediction model; Based on the structural mechanics similarity criterion, the scaling relationship of the original model and the scaled model in material density, elastic modulus and stiffness parameters is established; construct a composite loss function comprising a geometric similarity error term, a structural mechanics similarity error term and a predicted output error term based on the scaling relationship; adopt a gradient descent algorithm to optimize and iterate the composite loss function, update network parameters of the first converter transformer modal prediction model, and obtain a second converter transformer modal prediction model satisfying a constraint condition.

5. The converter transformer modal prediction method of claim 1, wherein, The method comprises the following steps: acquiring operating state data and corresponding structural size data of the target converter transformer; inputting the operating state data and the structural size data into the second converter transformer modal prediction model; obtaining inherent frequency, modal damping ratio and modal shape distribution data of the target converter transformer through forward calculation of the second converter transformer modal prediction model.

6. A converter transformer modal prediction system, characterized by, The method comprises the following steps: a response relationship construction module is configured to construct a modal superposition vibration analytical model through Laplace transform based on a vibration differential equation of a target converter transformer, so as to establish a theoretical frequency response relationship between an excitation vector and a response vector of the target converter transformer; a model construction module is configured to take the theoretical frequency response relationship as a learning target, train a neural network model through excitation data and response data of the target converter transformer under a plurality of operating conditions, and construct a first converter transformer modal prediction model; a model optimization module is configured to perform joint optimization of the first converter transformer modal prediction model based on geometric and structural mechanics similarity constraints, and obtain a second converter transformer modal prediction model; a modal prediction module is configured to collect real-time operating parameters of the target converter transformer and input the real-time operating parameters into the second converter transformer modal prediction model, and obtain corresponding modal prediction data.

7. The converter transformer mode prediction system of claim 6, wherein, The response relationship construction module is specifically configured to: establish a multi-degree-of-freedom vibration differential equation of the target converter transformer, wherein the vibration differential equation comprises a mass matrix, a stiffness matrix and a damping matrix; perform Laplace transform on the vibration differential equation to obtain a generalized impedance matrix of a power system in which the target converter transformer is located in a frequency domain, and obtain a generalized admittance matrix by inverting the generalized impedance matrix; construct a frequency response function matrix representing a quantitative mapping relationship between the excitation vector and the response vector of the target converter transformer based on the generalized admittance matrix, as a frequency response relationship.

8. The converter transformer mode prediction system of claim 6, wherein, The model construction module is specifically configured to: obtain excitation data and response data of the target converter transformer under a plurality of operating conditions; perform preprocessing on the excitation data and the response data, and construct an excitation matrix and a response matrix based on the preprocessed data; The excitation matrix is taken as an input quantity, and the response matrix is taken as an output quantity, and a preset deep neural network is supervised trained, the supervised training process is designed as follows: a forward propagation calculation network output, an error gradient between a predicted output and an actual response matrix is calculated through a back propagation algorithm, and a gradient descent method is used to iteratively update network weight parameters until a model prediction error converges to a preset threshold, so that the first converter transformer modal prediction model is obtained.

9. A computer device, comprising: A processor, a memory, and a computer program stored in the memory and configured to be executed by the processor are included, and the processor implements the converter transformer modal prediction method according to any one of claims 1 to 5 when the computer program is executed.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and when a device where the computer readable storage medium is located executes the computer program, the converter transformer modal prediction method according to any one of claims 1 to 5 is implemented.

Images