Converter valve temperature field reconstruction method and device based on reduced-order model

Abstract

According to the converter valve temperature field reconstruction method and device based on the reduced-order model provided by the invention, the model precision and the response speed are improved by constructing a reduced-order modeling architecture in which the primary function dynamic optimization network and the double-layer coefficient prediction network cooperate. Specifically, working condition parameters are input into a pre-trained primary function dynamic optimization network to obtain a primary function correction, and an initial primary function is corrected according to the primary function correction, so that the primary function can adapt to different working condition changes, and the problem of prediction precision attenuation caused by primary function mismatch is effectively solved. Furthermore, a target modal coefficient is determined through a double-layer coefficient prediction network combining a global layer and a local layer, and local nonlinear features are finely compensated while the macroscopic trend is captured. And finally, performing linear superposition on the target basis function and the target modal coefficient, and reconstructing to obtain the temperature field of the converter valve. Therefore, on the premise that details of the high-precision three-dimensional field are reserved, the reconstruction speed of the temperature field is increased, and the reconstruction precision is considered at the same time.

Description

Technical Field

[0001] This application relates to the field of power system technology, and in particular to a method and apparatus for reconstructing the temperature field of a converter valve based on a reduced-order model. Background Technology

[0002] As a core hub of the power system, converter stations operate their critical equipment under severe conditions of high load and high stress for extended periods, placing extremely high demands on the reliability and timeliness of operation and maintenance support. Promoting the construction of intelligent operation and maintenance systems with reduced or even unmanned operation has become an urgent goal for the industry. However, existing intelligent operation and maintenance technologies still have significant bottlenecks in data acquisition and status awareness, making it difficult to fully support the realization of this goal. Against this backdrop, digital twin technology has emerged. By integrating advanced communication technologies and high-fidelity modeling methods, it constructs a virtual model highly synchronized with the physical entity, providing a new technical path for deep perception of equipment status and intelligent decision-making.

[0003] Currently, digital twin modeling mainly relies on three types of models: mathematical models based on theoretical mechanisms, data-driven models based on measured data, and simulation models based on multiphysics simulation. However, in practical applications, the degree of integration between these heterogeneous models is still relatively limited, failing to fully realize their synergistic value. Especially in scenarios with strong multiphysics coupling, such as valve towers, traditional one-dimensional simplified models are unable to accurately describe their multi-field interaction mechanisms, including thermal, electromagnetic, and structural aspects. While high-precision simulations can provide detailed characterization, they consume significant computational resources and have slow solution speeds. In summary, existing modeling methods struggle to balance model accuracy and response efficiency. Summary of the Invention

[0004] The purpose of this application is to address at least one of the aforementioned technical deficiencies, particularly the technical deficiency that the degree of integration between heterogeneous models in the prior art is still relatively limited, failing to fully realize their synergistic value and making it difficult to balance model accuracy and response efficiency.

[0005] In a first aspect, this application provides a method for reconstructing the temperature field of a converter valve based on a reduced-order model, the method comprising:

[0006] The operating parameters of the converter valve are obtained and input into a pre-trained basis function dynamic optimization network to obtain the basis function correction amount;

[0007] Determine the initial basis functions, and modify the initial basis functions according to the basis function modification amount to obtain the target basis functions;

[0008] A pre-trained coefficient prediction network is obtained, and the operating condition parameters are respectively input into the global layer and local layer of the coefficient prediction network to obtain the first mode coefficient output by the global layer and the second mode coefficient output by the local layer;

[0009] The target modal coefficients are determined based on the first modal coefficients and the second modal coefficients, and the target basis function is linearly superimposed with the target modal coefficients to obtain the temperature field of the converter valve.

[0010] In one embodiment, the basis function dynamic optimization network and the coefficient prediction network are trained collaboratively. The training process of the basis function dynamic optimization network and the coefficient prediction network includes:

[0011] Acquire operating condition parameter samples and their corresponding temperature field data;

[0012] Based on the working condition parameter samples, the preset first network and the preset second network are iteratively trained. During the iterative training process, the reconstructed temperature field is determined based on the output results of the first network and the second network. The difference between the reconstructed temperature field and the temperature field data is used as the training target to optimize the first network and the second network.

[0013] When the preset training conditions are met, the final first network is determined to be the basis function dynamic optimization network, and the final second network is determined to be the coefficient prediction network.

[0014] In one embodiment, determining the initial basis function and modifying the initial basis function according to the basis function modification amount to obtain the target basis function includes:

[0015] Based on the heat dissipation mechanism of the valve tower of the converter valve, several key physical quantities are determined, and the sampling space is determined according to each physical quantity.

[0016] Based on the objectives of maximizing information entropy and minimizing parameter correlation constraints, the information distribution of the sampling space is optimized to obtain the target sampling space;

[0017] Samples are taken from the target sampling space to determine the sample matrix, and the sample matrix is ​​subjected to eigenorthogonal decomposition to obtain the initial basis functions.

[0018] In one embodiment, the step of inputting the operating condition parameters into the global layer and local layer of the coefficient prediction network respectively to obtain the first modal coefficient output by the global layer and the second modal coefficient output by the local layer includes:

[0019] The operating parameters are respectively input into the global layer and local layer of the coefficient prediction network;

[0020] In the global layer, the operating parameters are globally regressed using a wide-area radial basis function to capture the overall trend of the changes in the modal coefficients, so as to preliminarily predict the first modal coefficients.

[0021] In the local layer, a regression structure of a local adaptive radial basis function is constructed based on the residual part of the output of the global layer to enhance the ability to capture nonlinear disturbances and local complex changes in the operating parameters, and the second mode coefficient is output.

[0022] In one embodiment, determining the target mode coefficient based on the first mode coefficient and the second mode coefficient includes:

[0023] Obtain the fusion weights;

[0024] Based on the fusion weights, the first modal coefficients and the second modal coefficients are weighted and summed to obtain the target modal coefficients.

[0025] In one embodiment, determining the target mode coefficient based on the first mode coefficient and the second mode coefficient includes:

[0026] Obtain the fusion weights, and perform a weighted summation of the first modal coefficients and the second modal coefficients based on the fusion weights to obtain the weighted modal coefficients;

[0027] The weighted modal coefficients are input into a preset adaptive denoising model to obtain adaptively denoised weighted modal coefficients, and these weighted modal coefficients are determined as target modal coefficients.

[0028] The adaptive denoising model constructs a total loss function using reconstruction consistency loss, spectral smoothing loss, and physical constraint loss, and is trained based on the total loss function.

[0029] In one embodiment, after obtaining the temperature field of the converter valve, the method further includes:

[0030] Based on the temperature field of the converter valve, an operation and maintenance early warning plan is generated.

[0031] Secondly, this application provides a converter valve temperature field reconstruction device based on a reduced-order model, the device comprising:

[0032] The parameter acquisition module is used to acquire the operating parameters of the converter valve and input the operating parameters into a pre-trained basis function dynamic optimization network to obtain the basis function correction amount;

[0033] A basis function correction module is used to determine an initial basis function and correct the initial basis function according to the basis function correction amount to obtain a target basis function;

[0034] The coefficient prediction module is used to acquire a pre-trained coefficient prediction network and input the operating condition parameters into the global layer and local layer of the coefficient prediction network respectively to obtain the first mode coefficient output by the global layer and the second mode coefficient output by the local layer.

[0035] The temperature field reconstruction module is used to determine the target mode coefficients based on the first mode coefficients and the second mode coefficients, and to linearly superimpose the target basis function with the target mode coefficients to obtain the temperature field of the converter valve.

[0036] Thirdly, this application provides a storage medium storing computer-readable instructions, which, when executed by one or more processors, cause the one or more processors to perform the steps of the converter valve temperature field reconstruction method based on the reduced-order model as described in any of the above embodiments.

[0037] Fourthly, this application provides a computer device, including: one or more processors, and a memory;

[0038] The memory stores computer-readable instructions, and when the one or more processors execute the computer-readable instructions, they perform the steps of the converter valve temperature field reconstruction method based on the reduced-order model as described in any of the above embodiments.

[0039] As can be seen from the above technical solutions, the embodiments of this application have the following advantages:

[0040] The method and apparatus for reconstructing the temperature field of a converter valve based on a reduced-order model provided in this application improves model accuracy and response speed by constructing a reduced-order modeling architecture that coordinates a dynamic optimization network for basis functions and a two-layer coefficient prediction network. Specifically, the operating parameters of the converter valve are input into a pre-trained dynamic optimization network for basis functions to obtain basis function correction values. The initial basis functions are then corrected based on these correction values, enabling the basis functions to adapt to different operating conditions and effectively solving the problem of prediction accuracy decay caused by basis function mismatch. Furthermore, a two-layer coefficient prediction network combining global and local layers is used to determine the target modal coefficients. This captures macroscopic trends while finely compensating for local nonlinear characteristics, effectively enhancing the modal coefficients' ability to represent complex coupled fields. Finally, the target basis functions and target modal coefficients are linearly superimposed to reconstruct the temperature field of the converter valve. This method utilizes the above-mentioned reduced-order modeling architecture to reconstruct the temperature field while preserving high-precision three-dimensional field details, thereby improving the reconstruction speed while maintaining reconstruction accuracy. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 A flowchart illustrating a method for reconstructing the temperature field of a converter valve based on a reduced-order model, provided in an embodiment of this application;

[0043] Figure 2 Example flowcharts for valve module-level and valve tower system-level degradation provided in the embodiments of this application;

[0044] Figure 3 A schematic diagram of a converter valve temperature field reconstruction device based on a reduced-order model provided in this application embodiment;

[0045] Figure 4 This is an internal structural diagram of a computer device provided in an embodiment of this application. Detailed Implementation

[0046] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0047] In one embodiment, this application provides a method for reconstructing the temperature field of a converter valve based on a reduced-order model. The following embodiments illustrate the application of this method to a server. It is understood that the method for reconstructing the temperature field of a converter valve based on a reduced-order model can be performed on a single server or on a server cluster consisting of multiple servers; this application does not impose any specific limitations on this.

[0048] like Figure 1 As shown, this application provides a method for reconstructing the temperature field of a converter valve based on a reduced-order model, the method comprising:

[0049] S101: Obtain the operating parameters of the converter valve and input the operating parameters into the pre-trained basis function dynamic optimization network to obtain the basis function correction amount.

[0050] Operating parameters refer to a set of key external conditions and electrical control variables that describe the operating state of the converter valve, such as cooling medium flow rate and cooling medium inlet temperature. The basis function dynamic optimization network is a neural network model with operating condition awareness capabilities. It is used to output a correction vector for the shape parameters of the initial basis function based on the operating parameters, i.e., the basis function correction amount.

[0051] In this step, when the temperature field of the converter valve needs to be rapidly reconstructed, the server can read the measured values ​​of operating parameters such as cooling water flow rate, inlet temperature, and bridge arm current in real time or near real time through sensors deployed in the converter valve's cooling system and electrical circuit. These data are then collected and standardized to form a complete operating parameter vector. This operating parameter vector, representing the current operating state of the converter valve, is then input into a pre-trained basis function dynamic optimization network to obtain the basis function correction.

[0052] Specifically, when the converter valve operates under severe conditions of high current and low flow, the basis function dynamic optimization network senses this state, which is prone to local overheating, and outputs a correction value accordingly. This correction value makes the basis function shape describing the "high temperature hotspot" more prominent or sensitive, while potentially suppressing those basis functions characterizing "uniform low temperature," thus providing a more targeted and dynamically optimized basis for the next step of high-precision temperature field reconstruction.

[0053] S102: Determine the initial basis functions and correct them according to the basis function correction amount to obtain the target basis functions.

[0054] The initial basis function refers to a fixed and general set of spatial patterns extracted from a set of high-fidelity simulation samples covering typical working conditions using the intrinsic orthogonal decomposition method. It includes multiple basis functions.

[0055] In this step, the initial basis functions are first determined based on the object whose temperature field reconstruction is to be performed, i.e., the initial basis functions corresponding to the converter valve are determined. Then, the matrix corresponding to the basis function correction is superimposed element-wise with the matrix corresponding to the initial basis functions to obtain the target basis functions.

[0056] Specifically, when determining the initial basis functions, a set of sample data covering typical operating conditions of the converter valve can be obtained first using an appropriate sampling method, and the corresponding sample matrix can be determined based on this sample data. Then, the sample matrix is ​​subjected to eigenorthogonal decomposition to obtain the initial basis functions.

[0057] Understandably, traditional static basis functions are average values ​​obtained through training under a wide range of operating conditions, making it difficult to accurately adapt to all operating states, especially when operating conditions deviate from the design point, where significant accuracy degradation occurs. This method, however, introduces basis function correction parameters, enabling the target basis function to possess real-time adaptive capabilities, flexibly adjusting its spatial shape to better match the physical field characteristics determined by the current operating parameters.

[0058] In one example, the process of modifying the initial basis function according to the basis function modification to obtain the target basis function can be represented as follows:

[0059]

[0060] In the formula, Describe the objective basis function. Denotes the initial basis functions. This indicates that the basis function dynamically optimizes the network. Indicates operating parameters, This represents the parameters of the network dynamically optimized by the basis functions.

[0061] S103: Obtain the pre-trained coefficient prediction network, and input the operating condition parameters into the global layer and local layer of the coefficient prediction network respectively to obtain the first mode coefficient output by the global layer and the second mode coefficient output by the local layer.

[0062] The coefficient prediction network employs a two-layer structure consisting of a global layer and local layers to establish the complex nonlinear relationship between the operating condition space and the modal coefficient space. The global layer captures the macroscopic, overall trends in the relationship between operating condition parameters and modal coefficients. The local layers capture the subtle, local nonlinear characteristics and disturbances between the operating condition parameters and modal coefficients.

[0063] In this step, the trained coefficient prediction network is obtained, and its global layer is used to make preliminary predictions of the operating parameters, yielding the first modal coefficients. Simultaneously, the local layers of this network are used to obtain local correction values, i.e., the second modal coefficients, based on the operating parameters. Specifically, the first modal coefficients refer to the predicted modal coefficient values ​​output by the global layer, reflecting the fundamental weights of the basis functions of the temperature field under macroscopic physical laws. The second modal coefficients refer to the predicted modal coefficient values ​​output by the local layers, representing the required local correction and fine compensation for the first modal coefficients under the current specific operating conditions.

[0064] For example, when the valve body operates under high current conditions, the global layer can accurately predict that the overall temperature level will generally increase (this is reflected in the change of the first modal coefficient). However, some tiny hot spots caused by specific flow channel structures or local eddies, which cannot be fully explained by macroscopic laws, need to be detected by the local layer. The local layer outputs the second modal coefficient for these specific combinations of operating parameters, fine-tuning the global prediction results to ensure that the final reconstructed temperature field can both reflect the overall trend and accurately locate these key local features related to safety.

[0065] S104: Determine the target mode coefficient based on the first and second mode coefficients, and linearly superimpose the target basis function with the target mode coefficient to obtain the temperature field of the converter valve.

[0066] In this step, the robust trend of the global layer output (first mode coefficient) is combined with the fine-tuning captured by the local layer (second mode coefficient) using a preset or adaptively learned weight to obtain the target mode coefficient. Then, the target mode coefficient is linearly superimposed with the target basis function. This process uses an efficient matrix multiplication operation to combine the optimized target basis function, representing the spatial morphology, with the target mode coefficient, representing the weight allocation, to reconstruct the temperature field of the converter valve.

[0067] For example, assuming the converter valve is currently operating under high load and slightly insufficient cooling, a reduced-order model composed of a basis function dynamic optimization network and a coefficient prediction network can be used to obtain the target basis function optimized for this operating condition (which better describes potential hotspots) and the target modal coefficients that integrate macroscopic trends and local compensations. By performing linear superposition, a complete, high-resolution three-dimensional temperature cloud map of the valve body can be reconstructed instantly. Maintenance personnel can clearly see the exact location and value of the highest temperature point, as well as the overall temperature gradient distribution. This result highly matches the results of detailed CFD simulations that take several hours, but is hundreds or thousands of times faster, providing a direct basis for real-time condition assessment and early warning decisions.

[0068] In the above embodiments, a reduced-order modeling architecture combining a dynamic basis function optimization network and a two-layer coefficient prediction network is constructed to improve model accuracy and response speed. Specifically, the operating parameters of the converter valve are input into the pre-trained dynamic basis function optimization network to obtain basis function correction values. The initial basis functions are then corrected based on these correction values, enabling the basis functions to adapt to different operating conditions and effectively solving the problem of prediction accuracy decay caused by basis function mismatch. Furthermore, a two-layer coefficient prediction network combining global and local layers is used to determine the target modal coefficients. This captures macroscopic trends while finely compensating for local nonlinear characteristics, effectively enhancing the modal coefficients' ability to represent complex coupled fields. Finally, the target basis functions and target modal coefficients are linearly superimposed to reconstruct the temperature field of the converter valve. This reduced-order modeling architecture, while preserving high-precision three-dimensional field details, improves the temperature field reconstruction speed while maintaining reconstruction accuracy.

[0069] In one embodiment, the basis function dynamic optimization network and the coefficient prediction network are trained collaboratively. The training process of the basis function dynamic optimization network and the coefficient prediction network includes:

[0070] S1: Obtain operating condition parameter samples and their corresponding temperature field data.

[0071] S2: Iteratively train the preset first network and the preset second network based on the working condition parameter samples. During the iterative training process, determine the reconstructed temperature field based on the output results of the first network and the second network, and optimize the first network and the second network by taking the difference between the reconstructed temperature field and the temperature field data as the training target.

[0072] S3: When the preset training conditions are met, the final first network is determined as the basis function dynamic optimization network, and the final second network is determined as the coefficient prediction network.

[0073] In this embodiment, operating condition parameter samples and their corresponding temperature field data can be obtained from historical data. Then, the operating condition parameter samples are used as training data, and the temperature field data is used as label data to collaboratively train the first and second networks. Specifically, in one iteration, the operating condition parameter samples are input into the first network to obtain the basis function correction, which is used to determine the target basis function. Simultaneously, the operating condition parameter samples are input into the second network to obtain the target modal coefficients. The target basis function and target modal coefficients are further combined to obtain the reconstructed temperature field. A loss function is constructed based on the difference between the reconstructed temperature field and the actual temperature field data, and the first and second networks are collaboratively optimized based on this loss function. This process continues until preset training conditions are met. At this point, the first network is determined as the basis function dynamic optimization network, and the second network is determined as the coefficient prediction network.

[0074] Specifically, the training condition refers to the stopping condition for iterative training, which can be set to reach a preset number of iterations or to the absolute value of the difference between the losses of two consecutive iterations being less than a preset value. This application does not impose specific restrictions on this. Furthermore, the basis function dynamic optimization network can adopt a KAN (Kolmogorov-Arnold Network) structure, and the coefficient prediction network can adopt a two-layer ARBF (Adaptive Radial Basis Function) network.

[0075] It is understandable that by adopting a collaborative training approach, the outputs of two networks are used together to reconstruct the temperature field, and the difference between the reconstructed result and the real data is used as a unified training objective for joint optimization. This forces the two networks to learn to cooperate and divide tasks during training, thereby avoiding the error accumulation and functional disconnect that may occur with step-by-step independent training. This enables the trained ensemble model (i.e., the reduced-order model) to achieve higher accuracy in temperature field reconstruction.

[0076] In one embodiment, an initial basis function is determined, and the initial basis function is modified according to the basis function modification amount to obtain the target basis function, including:

[0077] S1: Based on the heat dissipation mechanism of the valve tower of the converter valve, several key physical quantities are determined, and the sampling space is determined according to each physical quantity.

[0078] S2: Based on the objectives of maximizing information entropy and minimizing parameter correlation constraints, optimize the information distribution of the sampling space to obtain the target sampling space.

[0079] S3: Sample from the target sampling space to determine the sample matrix, and perform eigenorthogonal decomposition on the sample matrix to obtain the initial basis functions.

[0080] In this embodiment, based on the heat dissipation mechanism and engineering design requirements of the flexible DC converter valve tower, key physical quantities are determined. These key physical quantities include, but are not limited to, temperature field distribution, heat flux density, cooling medium velocity, and turbulence intensity. These key physical quantities collectively characterize the fluid-thermal-solid multi-field coupling behavior during the heat dissipation process of the valve tower and form the basis for subsequent modeling and prediction. Subsequently, an improved Latin hypercube sampling method is used to sample within the sampling space constituted by these key physical quantities. Specifically, before sampling, the spatial uniformity and information content of the sample distribution in the sampling space are optimized based on the objectives of maximizing information entropy and minimizing parameter correlation constraints. Finally, sampling is performed in the optimized sampling space to obtain a sample matrix, and then the sample matrix is ​​subjected to eigenorthogonal decomposition to obtain the initial basis functions.

[0081] In one example, the objective of maximizing information entropy and minimizing parameter-related constraints can be represented by the following expression:

[0082]

[0083] In the formula, Information entropy represents the set of samples and is used to measure the richness of the sample distribution. Indicates the correlation between sample parameters. represents the weighting coefficients. The optimized target sample space has a uniform distribution of samples in the multidimensional space, providing sufficient coverage and improving the global representativeness and generalization performance of the reduced-order model.

[0084] Subsequently, for each set of sample parameters (including cooling medium flow rate, inlet temperature, heating power, ambient temperature, material thermal conductivity, etc.), experimentally validated high-fidelity CFD / conjugate heat transfer simulations were performed to ensure that the error of key results was less than 5%. Finally, a high-quality sample matrix was constructed. ,in, The number of discrete physical field points. This represents the number of sampling conditions.

[0085] Furthermore, regarding the sample matrix The process of performing eigenorthogonal decomposition can be represented as:

[0086]

[0087] in, The spatial mode matrix, i.e., the basis functions, This is the modal coefficient matrix, i.e., the modal coefficients. Subsequently, a modal truncation criterion based on energy contribution rate is adopted, and the optimal number of modes K is determined through this criterion, achieving data dimensionality reduction while ensuring reconstruction accuracy.

[0088] In one embodiment, operating parameters are input into the global layer and local layer of the coefficient prediction network, respectively, to obtain the first mode coefficients output by the global layer and the second mode coefficients output by the local layer, including:

[0089] S1: Input the operating parameters into the global and local layers of the coefficient prediction network respectively.

[0090] S2: In the global layer, the operating parameters are globally regressed using the wide-area radial basis function to capture the overall trend of the changes in the modal coefficients, so as to preliminarily predict the first modal coefficients.

[0091] S3: In the local layer, based on the residual part of the output of the global layer, a regression structure of the local adaptive radial basis function is constructed to enhance the ability to capture nonlinear disturbances and local complex changes in the operating parameters, and the second mode coefficient is output.

[0092] Among them, a wide-area radial basis function refers to a radial basis function with a large fixed bandwidth. A local adaptive radial basis function is a radial basis function with adaptive or smaller bandwidth, whose bandwidth can be dynamically adjusted according to the location or characteristics of the input condition, making its response range more concentrated and flexible.

[0093] In this embodiment, the operating condition parameter vector representing the current operating state is simultaneously fed into the global and local layers of the coefficient prediction network. In the global layer, wide-area radial basis functions are activated, each covering a broad region of the input space. Through their smooth response characteristics, a global regression is performed on the operating condition parameters to obtain the first mode coefficients. Based on this, a regression structure of local adaptive radial basis functions is constructed using the residual portion of the global layer's output. Then, based on these local adaptive radial basis functions, adaptive regression is performed on the currently input operating condition parameters to obtain the second mode coefficients. This compensates for and corrects nonlinear disturbances and complex changes that the global layer might ignore or smooth out in this local region. This division of labor—with the global layer performing wide-area regression and the local layer performing adaptive compensation—allows the global layer to specifically capture the dominant overall trend, while the local layer, based on its prior knowledge of the residual distribution, dynamically focuses on and fits local nonlinear features that the global model struggles to describe. This improves the complete and accurate representation of the global temperature field characteristics of the converter valve under complex operating conditions.

[0094] In one example, the process of performing global regression on the operating parameters using wide-area radial basis functions in the global layer can be represented as:

[0095]

[0096] In the formula, Represents the first modal coefficient. Let p be a Gaussian or polynomial radial kernel function, where p represents the operating parameters and M represents the number of radial basis functions in the global layer. Let be the learnable weights of the j-th radial basis function in the global layer. This represents the center of the j-th radial basis function. This refers to the global layer bandwidth parameter.

[0097] In the local layer, the process of constructing a regression structure of a local adaptive radial basis function based on the residual part of the global layer's output to enhance the ability to capture nonlinear disturbances and complex local changes in the operating parameters can be represented as follows:

[0098]

[0099] In the formula, Represents the second modal coefficient. Let p be a Gaussian or polynomial radial kernel function, where p represents the operating parameters and M represents the number of radial basis functions in the local layer. Let be the learnable weights of the j-th radial basis function in the local layer. This represents the center of the j-th radial basis function. This refers to the local layer bandwidth parameter. It can be dynamically adjusted according to the input, achieving adaptive weighting for high gradient regions.

[0100] In one embodiment, determining the target modal coefficients based on the first modal coefficients and the second modal coefficients includes:

[0101] S1: Obtain the fusion weights.

[0102] S2: Based on the fusion weights, the first mode coefficient and the second mode coefficient are weighted and summed to obtain the target mode coefficient.

[0103] Here, the fusion weight refers to the weight of the first mode coefficient or the second mode coefficient.

[0104] In this embodiment, the first and second modal coefficients are weighted and summed by fusing weights, which enhances the adaptive capability and overall prediction accuracy of the final target modal coefficients. This ensures that the reduced-order model can output the optimal result that does not distort macroscopic physical laws and accurately reflects local details under various complex operating conditions.

[0105] In one example, the process of obtaining the target mode coefficient by weighted summation of the first and second mode coefficients according to the fusion weights can be represented as:

[0106]

[0107] in, This represents the target modal coefficients, and k represents the index of the modal coefficients. Represents the first modal coefficient. Represents the second modal coefficient. The fusion weights can be adaptively determined through cross-validation or by minimizing the reconstruction error. It is important to note that the first mode coefficients, second mode coefficients, and target mode coefficients are all vector matrices containing multiple mode coefficients, corresponding to multiple basis functions in the target basis functions.

[0108] In one embodiment, determining the target modal coefficients based on the first modal coefficients and the second modal coefficients includes:

[0109] S1: Obtain the fusion weights, and perform a weighted summation of the first and second modal coefficients based on the fusion weights to obtain the weighted modal coefficients.

[0110] S2: Input the weighted modal coefficients into the preset adaptive denoising model to obtain the weighted modal coefficients after adaptive denoising, and determine the weighted modal coefficients as the target modal coefficients.

[0111] The adaptive denoising model constructs a total loss function using reconstruction consistency loss, spectral smoothing loss, and physical constraint loss, and is trained based on the total loss function.

[0112] In this embodiment, after obtaining the weighted modal coefficients, the matrix may contain high-frequency noise components due to discretization errors and turbulence disturbances. Directly using it for modeling may lead to unstable predictions. Therefore, an adaptive denoising model is used to reduce the noise of the weighted modal coefficients, thereby obtaining more accurate target modal coefficients and improving the accuracy of temperature field reconstruction.

[0113] Specifically, the loss of consistency during reconstruction To ensure that the modal coefficients after denoising can accurately reconstruct the original physical field, it can be expressed as:

[0114]

[0115] In the formula, Represents the modal coefficients after denoising. Represents a physical field. This represents the objective basis function.

[0116] Spectral smoothing loss Used to control high-frequency noise components and enhance the spatiotemporal continuity of modes, it can be expressed as:

[0117]

[0118] Based on the principles of energy conservation and boundary continuity, physical constraints are introduced, resulting in physical constraint losses. It can be represented as:

[0119]

[0120] in, Thermal conductivity, This is for fever.

[0121] Therefore, the total loss function of KAN is:

[0122]

[0123] in, and These are the hyperparameters used to balance the loss.

[0124] Furthermore, the adaptive denoising model can employ a KAN network, and its training process can include: using modal coefficients as input data, iteratively training the pre-trained model, and constructing a total loss function based on reconstruction consistency loss, spectral smoothing loss, and physical constraint loss during the iterative training process, updating the pre-trained model according to the total loss function, until the training conditions are met, and then determining the final pre-trained model as the adaptive denoising model.

[0125] In one embodiment, after obtaining the temperature field of the converter valve, the converter valve temperature field reconstruction method based on the reduced-order model further includes:

[0126] Based on the temperature field of the converter valve, an operation and maintenance early warning plan is generated.

[0127] In this embodiment, the temperature field of the converter valve obtained from the reconstruction can be used to analyze whether the junction temperature of the IGBT is within a reasonable range under the current operating conditions and the trend of temperature changes. Then, an operation and maintenance early warning plan can be generated based on this analysis. Furthermore, monitoring and early warning can be performed based on the temperature field obtained from the real-time reconstruction. For example, when the temperature is not within its corresponding reasonable range, an early warning signal can be automatically issued so that operation and maintenance personnel can intervene in a timely manner.

[0128] In one embodiment, such as Figure 2 As shown, Figure 2 This is a flowchart illustrating the valve module-level and valve tower system-level cooling process provided in this application. Specifically, the valve tower cooling system generally includes two parts: liquid cooling and air cooling, used to dissipate heat from the IGBT modules and other auxiliary heat-generating components (such as capacitors, voltage equalizing resistors, control boards, busbars, etc.). Liquid cooling devices typically use water as the cooling medium, connected to each IGBT module via flow channels to achieve direct cooling; air cooling relies on air convection for heat exchange.

[0129] Because the valve tower and IGBT cooling system involve strong fluid-structure interaction and heat transfer processes, the main steps include: clarifying the physical relationships between the components in the liquid cooling and air cooling systems; considering the impact of water temperature changes on heat dissipation performance in the liquid cooling system; and considering airflow organization and convective heat transfer effects in the air cooling system. The two subsystems, liquid cooling and air cooling, are then coupled to integrate a complete cooling system model. It is understood that the reconstruction method provided in this application can also be applied to objects of different scales, such as valve modules and valve tower systems. The application process for valve module-level and valve tower system-level order reduction is as follows:

[0130] (1) Valve module level downgrade

[0131] The valve module is the core heat-generating unit of the valve tower, and its heat dissipation level directly affects the heat load of the entire valve tower. It is also the area with the highest temperature. IGBTs, as the main heat source, primarily dissipate heat through water cooling, with the vast majority of heat being discharged through water-cooled plates and cooling pipes; air cooling is negligible. Therefore, voltage / current, cooling water temperature and flow rate, and ambient temperature can be selected as input features, with the overall temperature distribution of the valve module as the output target. Batch simulations can be conducted across the entire operating range to generate a training dataset for reduced-order modeling. Based on this data, a reduced-order model of the valve module's thermal characteristics (i.e., a combination of a basis function dynamic optimization network and a coefficient prediction network) can be constructed using a reduction-order method, ultimately forming a mapping relationship from electrical parameters and cooling conditions to the temperature field.

[0132] (2) Valve tower stage downgrading

[0133] Based on the reduced-order model of the valve module, and combined with the equivalent models of other heating and non-heating components, a system-level thermo-fluid macroscopic reduced-order model of the entire valve tower can be further constructed. Using the bridge arm current / valve module equivalent current, total cooling flow rate / inlet water temperature, and ambient temperature as inputs, and the overall three-dimensional temperature field of the valve tower as output, full-condition batch simulations are performed to generate training data. The flow distribution data of each module can be obtained through coupling with the established pipeline reduced-order model, thereby reducing the overall parameter scale through indirect fitting. Based on the batch simulation data, a valve tower-level thermal reduced-order model (i.e., a combination of a basis function dynamic optimization network and a coefficient prediction network) can be constructed to achieve rapid mapping from system inputs to the overall temperature field.

[0134] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0135] The following describes the converter valve temperature field reconstruction device based on the reduced-order model provided in the embodiments of this application. The converter valve temperature field reconstruction device based on the reduced-order model described below can be referred to in correspondence with the converter valve temperature field reconstruction method based on the reduced-order model described above.

[0136] like Figure 3As shown, this application provides a converter valve temperature field reconstruction device 200 based on a reduced-order model, the device comprising:

[0137] The parameter acquisition module 201 is used to acquire the operating parameters of the converter valve and input the operating parameters into the pre-trained basis function dynamic optimization network to obtain the basis function correction amount;

[0138] The basis function correction module 202 is used to determine the initial basis function and correct the initial basis function according to the basis function correction amount to obtain the target basis function;

[0139] The coefficient prediction module 203 is used to acquire a pre-trained coefficient prediction network and input the working condition parameters into the global layer and local layer of the coefficient prediction network respectively to obtain the first mode coefficient output by the global layer and the second mode coefficient output by the local layer.

[0140] The temperature field reconstruction module 204 is used to determine the target mode coefficients based on the first mode coefficients and the second mode coefficients, and to linearly superimpose the target basis function with the target mode coefficients to obtain the temperature field of the converter valve.

[0141] In the above embodiments, a reduced-order modeling architecture combining a dynamic basis function optimization network and a two-layer coefficient prediction network is constructed to improve model accuracy and response speed. Specifically, the operating parameters of the converter valve are input into the pre-trained dynamic basis function optimization network to obtain basis function correction values. The initial basis functions are then corrected based on these correction values, enabling the basis functions to adapt to different operating conditions and effectively solving the problem of prediction accuracy decay caused by basis function mismatch. Furthermore, a two-layer coefficient prediction network combining global and local layers is used to determine the target modal coefficients. This captures macroscopic trends while finely compensating for local nonlinear characteristics, effectively enhancing the modal coefficients' ability to represent complex coupled fields. Finally, the target basis functions and target modal coefficients are linearly superimposed to reconstruct the temperature field of the converter valve. This reduced-order modeling architecture, while preserving high-precision three-dimensional field details, improves the temperature field reconstruction speed while maintaining reconstruction accuracy.

[0142] In one embodiment, the basis function dynamic optimization network and the coefficient prediction network are trained collaboratively, and the parameter acquisition module or coefficient prediction module includes:

[0143] The data acquisition submodule is used to acquire operating condition parameter samples and their corresponding temperature field data.

[0144] The collaborative training submodule is used to iteratively train a preset first network and a preset second network based on the working condition parameter samples. During the iterative training process, the reconstructed temperature field is determined based on the output results of the first network and the second network, and the difference between the reconstructed temperature field and the temperature field data is used as the training target to optimize the first network and the second network.

[0145] The network determination submodule is used to determine the final first network as a basis function dynamic optimization network and the final second network as a coefficient prediction network when the preset training conditions are met.

[0146] In one embodiment, the basis function correction module includes:

[0147] The spatial determination submodule is used to determine several key physical quantities based on the heat dissipation mechanism of the valve tower based on the converter valve, and to determine the sampling space based on each physical quantity;

[0148] The spatial optimization submodule is used to optimize the information distribution of the sampling space based on the objectives of maximizing information entropy and minimizing parameter correlation constraints, so as to obtain the target sampling space.

[0149] The matrix factorization submodule is used to sample from the target sampling space to determine the sample matrix, and to perform eigenorthogonal decomposition on the sample matrix to obtain the initial basis functions.

[0150] In one embodiment, the coefficient prediction module includes:

[0151] The parameter input submodule is used to input the operating condition parameters into the global and local layers of the coefficient prediction network, respectively.

[0152] The first output submodule is used to perform global regression on the operating parameters using the wide-area radial basis function in the global layer, capture the overall trend of the changes in the modal coefficients, and make a preliminary prediction of the first modal coefficients.

[0153] The second output submodule is used to construct a regression structure of the local adaptive radial basis function based on the residual part of the output of the global layer in the local layer, so as to enhance the ability to capture nonlinear disturbances and local complex changes in the operating parameters and output the second mode coefficient.

[0154] In one embodiment, the temperature field reconstruction module includes:

[0155] The weight acquisition submodule is used to obtain the fusion weights;

[0156] The coefficient determination submodule is used to perform a weighted summation of the first mode coefficient and the second mode coefficient based on the fusion weight to obtain the target mode coefficient.

[0157] In one embodiment, the temperature field reconstruction module includes:

[0158] The weighted summation submodule is used to obtain the fusion weights and perform a weighted summation of the first and second mode coefficients based on the fusion weights to obtain the weighted mode coefficients.

[0159] The coefficient denoising submodule is used to input the weighted modal coefficients into a preset adaptive denoising model to obtain the adaptively denoised weighted modal coefficients, and to determine the weighted modal coefficients as the target modal coefficients.

[0160] The adaptive denoising model constructs a total loss function using reconstruction consistency loss, spectral smoothing loss, and physical constraint loss, and is trained based on the total loss function.

[0161] In one embodiment, after obtaining the temperature field of the converter valve, the converter valve temperature field reconstruction device based on the reduced-order model further includes:

[0162] The scheme generation module is used to generate operation and maintenance early warning schemes based on the temperature field of the converter valve.

[0163] The division of modules in the converter valve temperature field reconstruction device based on the reduced-order model described above is merely illustrative. In other embodiments, the converter valve temperature field reconstruction device based on the reduced-order model can be divided into different modules as needed to complete all or part of the functions of the device. Each module in the converter valve temperature field reconstruction device based on the reduced-order model can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, 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.

[0164] In one embodiment, this application also provides a storage medium storing computer-readable instructions that, when executed by one or more processors, cause the one or more processors to perform the steps of the converter valve temperature field reconstruction method based on the reduced-order model as described in any of the above embodiments.

[0165] In one embodiment, this application also provides a computer device storing computer-readable instructions that, when executed by one or more processors, cause the one or more processors to perform the steps of the converter valve temperature field reconstruction method based on a reduced-order model as described in any of the above embodiments.

[0166] Indicatively, such as Figure 4 As shown, Figure 4This is a schematic diagram of the internal structure of a computer device 300 provided in an embodiment of this application. The computer device 300 can be provided as a server. (Refer to...) Figure 4 The computer device 300 includes a processing component 302, which further includes one or more processors, and memory resources represented by memory 301 for storing instructions executable by the processing component 302, such as application programs. The application programs stored in memory 301 may include one or more modules, each corresponding to a set of instructions. Furthermore, the processing component 302 is configured to execute instructions to perform the converter valve temperature field reconstruction method based on a reduced-order model of any of the above embodiments.

[0167] The computer device 300 may also include a power supply component 303 configured to perform power management of the computer device 300, a wired or wireless network interface 304 configured to connect the computer device 300 to a network, and an input / output (I / O) interface 305. The computer device 300 may operate on an operating system stored in memory 301, such as Windows Server™, Mac OS X™, Unix™, Linux™, Free BSD™, or similar.

[0168] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0169] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising a…" does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. In this document, the singular forms "a," "an," and "the" may also include the plural forms unless the context clearly indicates otherwise. It should also be understood that the terms “comprising / including” or “having” specify the presence of the stated features, wholes, steps, operations, components, parts or combinations thereof, but do not exclude the possibility of the presence or addition of one or more other features, wholes, steps, operations, components, parts or combinations thereof. Meanwhile, the term “and / or” as used in this specification includes any and all combinations of the associated listed items.

[0170] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can be referred to each other.

[0171] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for reconstructing the temperature field of a converter valve based on a reduced-order model, characterized in that, The method includes: The operating parameters of the converter valve are obtained and input into a pre-trained basis function dynamic optimization network to obtain the basis function correction amount; Determine the initial basis functions, and modify the initial basis functions according to the basis function modification amount to obtain the target basis functions; A pre-trained coefficient prediction network is obtained, and the operating condition parameters are respectively input into the global layer and local layer of the coefficient prediction network to obtain the first mode coefficient output by the global layer and the second mode coefficient output by the local layer; The target modal coefficients are determined based on the first modal coefficients and the second modal coefficients, and the target basis function is linearly superimposed with the target modal coefficients to obtain the temperature field of the converter valve.

2. The method for reconstructing the temperature field of a converter valve based on a reduced-order model according to claim 1, characterized in that, The basis function dynamic optimization network and the coefficient prediction network are trained collaboratively. The training process of the basis function dynamic optimization network and the coefficient prediction network includes: Acquire operating condition parameter samples and their corresponding temperature field data; Based on the working condition parameter samples, the preset first network and the preset second network are iteratively trained. During the iterative training process, the reconstructed temperature field is determined based on the output results of the first network and the second network. The difference between the reconstructed temperature field and the temperature field data is used as the training target to optimize the first network and the second network. When the preset training conditions are met, the final first network is determined to be the basis function dynamic optimization network, and the final second network is determined to be the coefficient prediction network.

3. The method for reconstructing the temperature field of a converter valve based on a reduced-order model according to claim 1, characterized in that, The process of determining the initial basis function and modifying the initial basis function according to the basis function modification amount to obtain the target basis function includes: Based on the heat dissipation mechanism of the valve tower of the converter valve, several key physical quantities are determined, and the sampling space is determined according to each physical quantity. Based on the objectives of maximizing information entropy and minimizing parameter correlation constraints, the information distribution of the sampling space is optimized to obtain the target sampling space; Samples are taken from the target sampling space to determine the sample matrix, and the sample matrix is ​​subjected to eigenorthogonal decomposition to obtain the initial basis functions.

4. The method for reconstructing the temperature field of a converter valve based on a reduced-order model according to claim 1, characterized in that, The step of inputting the operating condition parameters into the global layer and local layer of the coefficient prediction network respectively to obtain the first mode coefficient output by the global layer and the second mode coefficient output by the local layer includes: The operating parameters are respectively input into the global layer and local layer of the coefficient prediction network; In the global layer, the operating parameters are globally regressed using a wide-area radial basis function to capture the overall trend of the changes in the modal coefficients, so as to preliminarily predict the first modal coefficients. In the local layer, a regression structure of a local adaptive radial basis function is constructed based on the residual part of the output of the global layer to enhance the ability to capture nonlinear disturbances and local complex changes in the operating parameters, and the second mode coefficient is output.

5. The method for reconstructing the temperature field of a converter valve based on a reduced-order model according to claim 1, characterized in that, Determining the target modal coefficient based on the first modal coefficient and the second modal coefficient includes: Obtain the fusion weights; Based on the fusion weights, the first modal coefficients and the second modal coefficients are weighted and summed to obtain the target modal coefficients.

6. The method for reconstructing the temperature field of a converter valve based on a reduced-order model according to claim 1, characterized in that, Determining the target modal coefficient based on the first modal coefficient and the second modal coefficient includes: Obtain the fusion weights, and perform a weighted summation of the first modal coefficients and the second modal coefficients based on the fusion weights to obtain the weighted modal coefficients; The weighted modal coefficients are input into a preset adaptive denoising model to obtain adaptively denoised weighted modal coefficients, and these weighted modal coefficients are determined as target modal coefficients. The adaptive denoising model constructs a total loss function using reconstruction consistency loss, spectral smoothing loss, and physical constraint loss, and is trained based on the total loss function.

7. The method for reconstructing the temperature field of a converter valve based on a reduced-order model according to any one of claims 1 to 6, characterized in that, After obtaining the temperature field of the converter valve, the method further includes: Based on the temperature field of the converter valve, an operation and maintenance early warning plan is generated.

8. A converter valve temperature field reconstruction device based on a reduced-order model, characterized in that, The device includes: The parameter acquisition module is used to acquire the operating parameters of the converter valve and input the operating parameters into a pre-trained basis function dynamic optimization network to obtain the basis function correction amount; A basis function correction module is used to determine an initial basis function and correct the initial basis function according to the basis function correction amount to obtain a target basis function; The coefficient prediction module is used to acquire a pre-trained coefficient prediction network and input the operating condition parameters into the global layer and local layer of the coefficient prediction network respectively to obtain the first modal coefficients output by the global layer and the second modal coefficients output by the local layer. The temperature field reconstruction module is used to determine the target mode coefficients based on the first mode coefficients and the second mode coefficients, and to linearly superimpose the target basis function with the target mode coefficients to obtain the temperature field of the converter valve.

9. A storage medium, characterized in that: The storage medium stores computer-readable instructions that, when executed by one or more processors, cause the one or more processors to perform the steps of the converter valve temperature field reconstruction method based on a reduced-order model as described in any one of claims 1 to 7.

10. A computer device, characterized in that, include: One or more processors, and memory; The memory stores computer-readable instructions, which, when executed by the one or more processors, perform the steps of the converter valve temperature field reconstruction method based on a reduced-order model as described in any one of claims 1 to 7.

Images