A converter valve hall temperature field simulation method and device, a terminal and a medium

Abstract

The application discloses a converter valve hall temperature field simulation method and device, a terminal and a medium, and relates to the technical field of converter valve simulation. The scheme provided by the application fuses convolution operation and a multi-head attention mechanism, extracts local thermal features such as a high-temperature area of a valve tower and a tuyere diffusion path in original temperature field sample data by using a convolution layer, and then models a global temperature field distribution formed by cold and hot airflow interaction through a self-attention layer, fuses local temperature gradient details and global heat transfer rules, forms complementary feature expression, and thus improves simulation accuracy under complex working conditions while ensuring calculation efficiency, and provides reliable technical support for real-time monitoring of a converter valve hall temperature field.

Description

Technical Field

[0001] This application relates to the field of converter valve simulation technology, and in particular to a method, device, terminal and medium for simulating the temperature field of a converter valve chamber. Background Technology

[0002] The converter valve hall is a building in a converter station used to house equipment related to converter valves. It is the core of the converter station's architecture. The temperature field simulation method for the converter valve hall is mainly used to simulate the temperature distribution and dynamic changes within the hall. This method achieves quantitative analysis and trend prediction of the temperature field by constructing a corresponding model and combining it with measured data.

[0003] In principle, simulation methods can be broadly categorized into two types: physical model-based methods and data-driven methods. Physical model-based methods focus on the fundamental laws of heat transfer, establishing coupled mathematical models between different operating conditions and using numerical calculations to solve the governing equations to obtain the temperature field distribution. While these methods can accurately reproduce the physical processes of temperature field formation, they involve large computational loads, complex model building, and low computational efficiency. Data-driven methods, on the other hand, are based on monitoring data and use machine learning or deep learning algorithms to construct mapping relationships between the temperature field and multi-dimensional features, enabling rapid temperature field prediction. Their core advantage lies in the fact that they do not require analyzing complex physical mechanisms; results can be output simply through data association modeling, resulting in high computational efficiency. However, their simulation accuracy is poor when facing complex engineering scenarios. Therefore, to meet the needs of real-time monitoring and rapid decision-making in engineering, providing a temperature field simulation scheme for converter valve halls that balances computational efficiency and accuracy has become a pressing technical problem for those skilled in the art. Summary of the Invention

[0004] This application provides a method, device, terminal, and medium for simulating the temperature field of a converter valve hall, which aims to achieve both efficiency and accuracy in the simulation calculation of the temperature field of a converter valve hall.

[0005] To address the aforementioned technical problems, the first aspect of this application provides a method for simulating the temperature field of a converter valve chamber, comprising: Obtain raw temperature field sample data of the converter valve chamber; The original temperature field sample data is input into an initial convolutional attention fusion model to train the model using the original temperature field sample data and a preset adaptive learning optimization logic. The convolutional attention fusion model includes a convolutional operation module and an attention operation module. The convolutional operation module extracts features from the input temperature field data through a multi-layer convolutional structure, concatenates the extracted features along the channel dimension into a spatiotemporal feature sequence representing a composite feature, and transmits it to the attention operation module. The attention operation module, based on the spatiotemporal feature sequence, combines a multi-head attention mechanism and feature association enhancement processing to capture the long-distance spatial dependency features of the spatiotemporal feature sequence. When the convergence degree or number of iterations of the convolutional attention fusion model meets the preset training termination condition, the convolutional attention fusion model is obtained. The real-time temperature data of the converter valve hall is input into the convolutional attention fusion model, so that the temperature field simulation results of the converter valve hall can be obtained through the calculation of the convolutional attention fusion model.

[0006] Preferably, the attention operation module includes: a position encoding layer, an encoder layer, and a decoder layer, wherein the encoder layer includes: a multi-head self-attention submodule and a feedforward fully connected network submodule; The position encoding layer is used to map the spatiotemporal feature sequence to the embedded feature dimension through a fully connected layer, and then convert the spatiotemporal feature sequence into the corresponding position vector through a preset sine-cosine position encoding mechanism, and then embed the position vector into the input feature matrix through element-wise addition; The multi-head self-attention submodule is used to generate a QKV matrix through linear transformation based on the input feature matrix, then divide the QKV matrix into several self-attention heads according to the embedded feature dimension, then map the high-dimensional vector formed by concatenating the self-attention heads to the embedded feature dimension, and then combine the attention weight calculation method and weighted summation method of the multi-head attention mechanism to obtain the associated feature vector. The feedforward fully connected network submodule is used to perform nonlinear enhancement processing on the associated feature vector output by the multi-head self-attention submodule and then transmit it to the decoder layer.

[0007] Preferably, the position encoding formula of the position encoding layer is as follows: ; ; In the formula, This represents the location index of the sample data, where i represents the location. The i-th dimension at position, For position The position vector located in the 2i-th dimension. denoted as the embedding feature dimension.

[0008] Preferably, the function expression of the feedforward fully connected network submodule is: ; In the formula, Let x be the associated feature vector after processing by the feedforward fully connected network submodule, and let x be the associated feature vector before processing by the feedforward fully connected network submodule. For Swish activation function, and These are the weight coefficients of the two linear layers in the feedforward fully connected network submodule. and These are the bias term coefficients of the two linear layers in the feedforward fully connected network submodule, respectively.

[0009] Preferably, the multi-layer convolutional structure is a two-layer convolutional structure.

[0010] Preferably, after obtaining the original temperature field sample data of the converter valve chamber, the method further includes: The original temperature field sample data is preprocessed, wherein the preprocessing includes data cleaning and normalization.

[0011] Preferably, the adaptive learning optimization logic is specifically: Adamax adaptive learning optimization logic.

[0012] The second aspect of this application provides a device for simulating the temperature field of a converter valve chamber, comprising: The sample data acquisition unit is used to acquire the original temperature field sample data of the converter valve chamber; The fusion model training unit is used to input the original temperature field sample data into an initial convolutional attention fusion model, so as to train the convolutional attention fusion model through the original temperature field sample data and a preset adaptive learning optimization logic. The convolutional attention fusion model includes a convolution operation module and an attention operation module. The convolution operation module is used to extract features from the input temperature field data step by step through a multi-layer convolutional structure, and to concatenate the extracted features along the channel dimension into a spatiotemporal feature sequence with composite feature representation and transmit it to the attention operation module. The attention operation module is used to capture the long-distance spatial dependency features of the spatiotemporal feature sequence based on the spatiotemporal feature sequence, combined with a multi-head attention mechanism and feature association enhancement processing. The model convergence determination unit is used to obtain the convolutional attention fusion model when the convergence degree or number of iterations of the convolutional attention fusion model meets the preset training termination condition. The temperature field simulation control unit is used to input the real-time temperature operating data of the converter valve hall into the convolutional attention fusion model, so as to obtain the temperature field simulation results of the converter valve hall through the calculation of the convolutional attention fusion model.

[0013] A third aspect of this application provides a simulation terminal for the temperature field of a converter valve hall, comprising: a memory and a processor; The memory is used to store program code, which corresponds to the valve hall temperature field simulation method of the converter valve as provided in the first aspect of this application. The processor is used to read and execute the program code to implement the method for simulating the temperature field of the converter valve hall.

[0014] The fourth aspect of this application provides a computer-readable storage medium storing program code, which is read and executed by a processor to implement the converter valve hall temperature field simulation method provided in the first aspect of this application.

[0015] As can be seen from the above technical solutions, this application has the following advantages: The solution provided in this application integrates convolution operations with a multi-head attention mechanism. It utilizes convolutional layers to extract local thermal features such as the high-temperature zone of the valve tower and the diffusion path of the air outlet from the original temperature field sample data. Then, it uses self-attention layers to model the global temperature field distribution formed by the interaction of hot and cold airflows. By integrating local temperature gradient details with global heat transfer laws, complementary feature representations are formed. This improves the simulation accuracy under complex working conditions while ensuring computational efficiency, providing reliable technical support for real-time monitoring of the temperature field of the converter valve hall. Attached Figure Description

[0016] 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.

[0017] Figure 1 This is a flowchart illustrating an embodiment of a method for simulating the temperature field of a converter valve chamber provided in this application.

[0018] Figure 2 A schematic diagram of the model architecture of the convolutional attention fusion model provided in this application.

[0019] Figure 3 The input-output relationship curve of the Swish activation function.

[0020] Figure 4The simulation cloud diagram of the converter valve chamber module is obtained based on the method provided in this application.

[0021] Figure 5 A comparison of loss data between the convolutional attention fusion model provided in this application and other existing model algorithms.

[0022] Figure 6 This is a schematic diagram of the architecture of an embodiment of a converter valve hall temperature field simulation device provided in this application.

[0023] Figure 7 This is a schematic diagram of the architecture of a converter valve hall temperature field simulation terminal embodiment provided in this application. Detailed Implementation

[0024] In existing technologies, temperature field simulation of converter valve halls mainly relies on physical model-based or data-driven methods. Physical model-based methods require establishing complex heat transfer equations and performing numerical solutions, resulting in low computational efficiency. While data-driven methods can provide rapid predictions, their accuracy is insufficient when facing complex operating conditions. In actual converter station operation, the temperature gradient formed by the high-temperature region of the valve tower and the air conditioning supply path exhibits significant spatial differences. Traditional methods struggle to simultaneously consider both local thermal characteristics and global temperature distribution patterns, leading to simulation results that fail to meet the requirements for real-time monitoring and rapid decision-making.

[0025] To address the aforementioned issues, a model capable of extracting both local thermal features and capturing global temperature dependence needs to be constructed. Therefore, this application provides a method, apparatus, terminal, and medium for simulating the temperature field of a converter valve hall, aiming to achieve both efficiency and accuracy in the simulation calculation of the temperature field of the converter valve hall.

[0026] To make the inventive objectives, features, and advantages of this application more apparent and understandable, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0027] First, a detailed description of an embodiment of the temperature field simulation method for a converter valve chamber provided in this application is as follows: Please see Figure 1 The method for simulating the temperature field of the converter valve chamber provided in this embodiment includes: Step 101: Obtain the original temperature field sample data of the converter valve chamber.

[0028] Step 102: Input the original temperature field sample data into the initial convolutional attention fusion model to train the convolutional attention fusion model using the original temperature field sample data and the preset adaptive learning optimization logic. The convolutional attention fusion model includes a convolutional operation module and an attention operation module. The convolutional operation module is used to extract features from the input temperature field data step by step through a multi-layer convolutional structure, and to concatenate the extracted features along the channel dimension into a spatiotemporal feature sequence with composite feature representation and transmit it to the attention operation module. The attention operation module is used to capture the long-distance spatial dependency features of the spatiotemporal feature sequence based on the spatiotemporal feature sequence, combined with a multi-head attention mechanism and feature association enhancement processing.

[0029] Step 103: When the convergence or number of iterations of the convolutional attention fusion model meets the preset training termination condition, the convolutional attention fusion model is obtained.

[0030] Step 104: Input the real-time temperature condition data of the converter valve hall into the convolutional attention fusion model, so as to obtain the temperature field simulation results of the converter valve hall through the calculation of the convolutional attention fusion model.

[0031] It should be noted that the Convolutional Neural Network (CNN) module is a component that extracts spatial features of the temperature field step by step through a multi-layer convolutional structure. Specifically, it can be implemented by stacking convolutional layers with different kernel sizes, used to capture local temperature change patterns caused by heat dissipation from the valve tower surface and air conditioning airflow. The Transformer module is a component that processes spatiotemporal feature sequences based on a multi-head attention mechanism. Specifically, it can be implemented by generating a query matrix Q, a key matrix K, and a value matrix V through linear transformation and calculating association weights, used to model the temperature influence of the air conditioning airflow path on distant areas. The adaptive learning optimization logic refers to an algorithm that dynamically adjusts the parameter update rate during model training. Specifically, it can be implemented using an optimizer based on gradient moment estimation, used to balance the impact of different feature dimensions on the model convergence speed.

[0032] The model architecture of the convolutional attention fusion model provided in this application is as follows: Figure 2 As shown, this model is a hybrid neural network model combining CNN and Transformer, comprising an input layer, convolutional layer, encoding layer, Transformer layer, fully connected layer, and output layer. The convolutional layer extracts key local features of the temperature field; the encoding layer encodes and transforms these features to adapt to subsequent modeling requirements; the Transformer layer performs global feature correlation modeling; and the fully connected layer integrates the features to support the predicted output.

[0033] In a specific example of model training, the following hyperparameter configuration is used: the optimizer is Adama, the initial learning rate is set to 0.001, and a cosine annealing decay strategy is employed; the loss function is mean squared error (MSE); the batch size is set to 64; the maximum number of training epochs is 500, and an early stopping mechanism is implemented, terminating training when the validation set loss does not decrease for 30 consecutive epochs. During training, each iteration includes the following steps: First, a batch of samples is extracted from the training set and input into the convolutional attention fusion model for forward propagation to obtain the predicted temperature field; second, the MSE loss between the predicted temperature field and the finite element simulation labels is calculated; then, the Adama optimizer is used to update all trainable parameters of the CNN module and the Transformer module according to the gradient of the loss function. To prevent overfitting, a Dropout layer with a ratio of 0.3 is introduced before the fully connected layer.

[0034] In practical applications, the internal computation process of the model includes: after the preprocessed operating data enters the neural network through the input layer, the convolutional layer first performs preliminary feature extraction: it uses the local connectivity and weight sharing characteristics of the convolutional kernel to capture the local features of the input data, and uses the Swish activation function to enhance the nonlinear expressive power; then, the convolutional features are mapped to a fixed-dimensional vector space through the embedding layer, and positional encoding is superimposed to supplement the positional information of the features; subsequently, multiple stacked Transformer layers use the self-attention mechanism to capture long-distance dependencies in the data, completing the global association and deep modeling of features; after global feature modeling, the average pooling layer performs dimensionality compression and feature aggregation to reduce computational complexity; then, the pooled features are linearly transformed through the fully connected layer to further mine the potential information in the features; finally, the features are mapped to the corresponding output space to output the simulation results of the temperature field of the converter valve hall.

[0035] Specifically, the process begins by acquiring raw temperature field sample data, primarily including load conditions such as valve tower heat loss in the valve hall, air supply velocity of the air conditioning system in the valve hall, and air supply temperature of the air conditioning system in the valve hall. This acquired sample data is then input into a convolutional attention fusion model. The model's convolutional operation module scans the input data through a multi-layer convolutional structure, extracting temperature gradient features at different locations and concatenating them along the channel dimension to form a spatiotemporal feature sequence. The attention operation module receives this sequence and calculates the correlation weights between feature vectors using a multi-head attention mechanism, capturing the impact of the air conditioning system's diffusion path within the valve hall space on the overall temperature distribution. During training, an adaptive optimization algorithm dynamically adjusts the learning rate based on the loss function gradient, accelerating model convergence. The trained model then performs forward computation on real-time operating data, outputting simulation results that include details of the temperature field distribution and global trends.

[0036] This solution is based on data-driven simulation principles. It automatically extracts local thermal features through convolutional layers, avoiding manual feature engineering. An attention mechanism models long-distance temperature dependence, overcoming the limitation of the receptive field in traditional convolutional networks. This enables end-to-end learning of spatial correlation features of the temperature field. Through these technical solutions, it can effectively integrate local temperature gradient details with global heat transfer patterns. This avoids the drawbacks of conventional data-driven methods that only fit the input-output relationship through fully connected networks and ignore the spatial correlation of the temperature field. While ensuring computational efficiency, it improves the simulation accuracy under complex operating conditions, providing reliable technical support for real-time monitoring of the temperature field in the converter valve hall.

[0037] Based on the above basic embodiments, this application further provides a more detailed description of the attention operation module. This embodiment proposes that the attention operation module may specifically include: a position encoding layer, a multi-head self-attention submodule, and a feedforward fully connected network submodule; the position encoding layer is used to map the spatiotemporal feature sequence to the embedded feature dimension through the fully connected layer, and then convert the spatiotemporal feature sequence into the corresponding position vector through a preset sine-cosine position encoding mechanism, and then embed the position vector into the input feature matrix through element-wise addition; the multi-head self-attention submodule is used to generate a QKV matrix through linear transformation based on the input feature matrix, and then divide the QKV matrix into several self-attention heads according to the embedded feature dimension, and then map the high-dimensional vector formed by concatenating the various self-attention heads to the embedded feature dimension, and then combine the attention weight calculation method and weighted summation method of the multi-head attention mechanism to obtain the associated feature vector; the feedforward fully connected network submodule is used to perform nonlinear enhancement processing on the associated feature vector output by the multi-head self-attention submodule.

[0038] It should be noted that the attention module Transformer in this embodiment can be understood as consisting of an encoder and a decoder, both of which implement sequence-to-sequence mapping based on a multi-layered stacked sub-layer architecture. The encoder is responsible for converting the input sequence into a context-aware feature representation, while the decoder generates the target sequence based on the encoder's output. The encoder structure achieves multi-level feature abstraction through four stacked layers, each containing two core sub-layers: a multi-head self-attention layer and a feedforward fully connected layer. Each sub-layer is followed by residual connections and layer normalization to ensure stable gradient propagation and accelerate convergence. The decoder structure is similar to the encoder, consisting of four decoding layers. Each decoding layer contains a masked multi-head self-attention layer, an encoder-decoder attention layer, and a feedforward neural network layer. Among them, the Encoder-Decoder attention layer is a key sub-layer unique to the decoder, used to dynamically focus on relevant regions of the encoder output during decoding. The output of each sub-layer is processed through residual connections and layer normalization, and finally mapped to the target sequence dimension through a fully connected layer.

[0039] The position encoding layer refers to the module that generates position vectors through mathematical functions and embeds them into the feature matrix. Specifically, it can be implemented using a sine-cosine position encoding mechanism, which generates absolute position identifiers through sine and cosine functions of different frequencies to solve the problem of missing position information in the spatial dimension of temperature field data.

[0040] The multi-head self-attention submodule refers to a module that divides the input features into multiple subspaces for parallel attention computation. Specifically, it can be achieved by generating a QKV matrix through linear transformation and dividing it into multiple self-attention heads, thereby enhancing the model's ability to capture the correlation between different regions of the temperature field.

[0041] The feedforward fully connected network submodule refers to the module that performs nonlinear transformation on the attention output. Specifically, it can be implemented using a two-layer fully connected structure containing the Swish activation function, which enhances the feature representation of the core heat dissipation area while preserving the weak temperature gradient features.

[0042] Specifically, the spatiotemporal feature sequence is first mapped to a preset embedding feature dimension, and then a feature matrix with location identifiers is generated through a sine-cosine function. This matrix undergoes a linear transformation to generate a QKV matrix, which is then divided into multiple self-attention heads for parallel computation. These heads are then concatenated and projected to form associated feature vectors. A feedforward network performs two linear transformations and Swish activation on the associated features, ultimately outputting an enhanced temperature field feature representation.

[0043] More specifically, during the operation of the converter valve hall, the flow path of cold air from the vent to the valve tower causes the overall temperature gradient to shift. The temperature at distant measuring points is indirectly affected by upstream heat loss and airflow. The attention computing module can capture this long-distance spatial dependence through the attention mechanism, achieving accurate modeling of the global temperature field. The specific process can be seen in the following example: The spatiotemporal feature sequences extracted by the CNN module are mapped to the embedded feature dimension through a fully connected layer. Then, a sine-cosine position encoding mechanism is used to generate absolute position identifiers through sine-cosine functions of different frequencies, the mathematical definition of which is as follows: ; ; in, This represents the position index of the sample, where i represents the position. The i-th dimension at position i, specifically the index value of the embedded feature dimension, is used to distinguish the position vectors of different dimensions, generating a position vector PE. Next, the position vector containing order information is embedded into the input latent features through element-wise addition to form the input feature matrix, and finally the output is fed to the encoder layer.

[0044] The position encoding formula, by alternating the use of sine and cosine functions of different frequencies, can generate feature vectors with multi-scale position information in a high-dimensional embedding space. This allows for a more accurate modeling of the correlation between local hotspots and global heat dissipation paths in the temperature field, and avoids the problem of decreased accuracy in temperature field simulation due to insufficient position information encoding.

[0045] The query is generated by linear transformation of the input feature matrix. ,key ,value Three matrices. The multi-head attention mechanism integrates Q, K, and V along the feature dimension. The upper part is divided into h heads, each head has a dimension of The calculation method for each attention head is as follows: ; These are learnable weight parameters. The output of all heads... When concatenated, they are projected through a learnable output matrix. The concatenated high-dimensional vector is mapped back to the d_model dimension.

[0046] ; Attention score is calculated by the dot product of query and key: ; Where, dk The scaling factor is used to mitigate the vanishing gradient problem caused by excessively large dot products. The softmax function normalizes the scores into a probability distribution, representing the attention weight of each position to other positions, and finally obtains the context-aware output vector through weighted summation.

[0047] The feedforward fully connected network consists of two linear layers and a Swish activation function, used to non-linearly enhance the associated features of the attention output, as follows: ; After the feedforward fully connected network completes the nonlinear enhancement processing of the associated features, the processed associated features are input into the decoder layer, and finally the output is mapped to the target sequence dimension to obtain the target feature sequence.

[0048] It needs further explanation that the reason for using Swish as the activation function in the feedforward fully connected network submodule is as follows: Compared to the ReLU activation function commonly used in traditional methods for processing feedforward network features, the ReLU activation function tends to directly truncate the weak temperature gradient signal in the negative range, failing to effectively reflect the impact of local heat dissipation anomalies on the global temperature field. However, in the temperature field simulation scenario of the converter valve hall, the data features exhibit significant differentiation in strength, showing both significant temperature changes around core equipment such as the valve tower, and weak temperature differences in local airflow disturbances and equipment edge areas. The Swish activation function possesses better gradient smoothness, such as... Figure 3 As shown, Swish retains finely tuned feature responses in the negative range, avoiding the direct discarding of weak temperature difference signals and ensuring that the model does not miss potentially influential details. Meanwhile, its feature propagation enhancement capability in the non-negative range highlights the temperature change patterns in the core heat dissipation area, allowing the model to focus on dominant heat transfer features. This design not only significantly reduces the probability of neuron inactivation during deep network training but also adapts to complex data distributions. Therefore, Swish's characteristics are more suitable for the temperature field environment of the converter valve chamber.

[0049] This scheme introduces Swish as an activation function, which retains finely tuned feature responses in the negative range. This allows the model to capture low-intensity temperature fluctuations at the end of the air supply path or the edge of the valve tower, alleviating the gradient vanishing problem and balancing the relationship between preserving weak temperature signals and enhancing core thermal features. This avoids the loss of key detail features. Furthermore, it can enhance the high-temperature diffusion characteristics of the core heat dissipation area in the non-negative range through its approximately linear properties, thereby improving the model's accuracy in modeling the dominant heat transfer laws.

[0050] Furthermore, based on the above-described basic embodiment, this embodiment further proposes a multi-layer convolutional structure, specifically a two-layer convolutional structure.

[0051] It should be noted that the high-temperature region formed by the valve tower's heat dissipation, the low-temperature airflow diffusion from the cooling vents, and the temperature gradient between adjacent measuring points all exhibit local spatial continuity. The CNN module achieves efficient feature extraction in the temperature field simulation of the converter valve hall through a local receptive field and weight sharing mechanism. Its hierarchical abstraction capability can also automatically filter redundant information. It uses double-layer convolution to extract features from the input feature map step by step, concatenating them along the channel dimension to form a composite feature representation, which is then fed into the subsequent Transformer model. This provides the Transformer module with a more discriminative and compact feature representation, thereby enhancing global modeling efficiency.

[0052] Specifically, in the simulation scenario of the converter valve hall temperature field, the dual-layer convolutional structure captures local thermal features such as the temperature gradient on the valve tower surface and the cold air diffusion from the air conditioning vents through the first layer of convolution. The second layer of convolution then establishes the heat conduction correlation pattern between adjacent regions based on the aforementioned features. The feature maps output by the two convolutional layers are concatenated along the channel dimension to form a spatiotemporal feature sequence with composite feature representation. This sequence contains both low-level local details and mid-level regional correlation information, providing effective input for the subsequent attention operation module to capture long-distance spatial dependencies.

[0053] For example, suppose the input sequence contains m samples, each sample has n features (such as core parameters like valve tower heat loss, cooling system air velocity, and air supply temperature), denoted as Each layer has k convolutional kernels, with kernel size p and stride q. The convolution process of the h-th convolutional kernel (h=1,2,…,k) on the input sequence is as follows: First, on the sample... Perform convolution to output features Then, the convolution kernel is slid by a stride of q, and the convolution operation is repeated until all samples are covered. For the s-th sample, a single output feature... The calculation formula can be expressed as: ; in, Let be the weight parameters of the h-th convolutional kernel. The output sequence length r is determined by the number of input samples m, the kernel size p, and the stride q, as shown in the following expression: ; Compared to existing technologies, traditional temperature field simulation models often employ single-layer convolutional structures or deep convolutional structures with three or more layers. Single-layer convolutions, due to their limited receptive field, struggle to capture regional correlation features, while deep convolutional structures, although capable of extracting complex features, significantly increase computational complexity. Dual-layer convolutional structures, while ensuring accuracy in local feature extraction, balance computational efficiency and feature representation capabilities through reasonable control of network depth, making them more suitable for the response speed requirements of real-time temperature field simulations in engineering scenarios. This technical solution effectively addresses the problems of high computational complexity and slow inference speed in temperature field simulations using traditional multi-layer convolutional structures. Dual-layer convolutional structures can extract local thermal features from both the high-temperature and low-temperature diffusion regions of the valve hall, and avoid the gradient vanishing risk associated with excessively deep networks through hierarchical feature fusion, enabling efficient modeling of the spatiotemporal evolution of the temperature field with limited computational resources.

[0054] Furthermore, this embodiment also proposes that after performing step 101, the original temperature field sample data be preprocessed, including data cleaning and normalization.

[0055] Data cleaning refers to identifying and correcting unreasonable values ​​in the original data through preset physical constraints. Specifically, this can be achieved by setting a temperature threshold range. When the temperature value exceeds this range, it is considered abnormal data. A substitute value is generated between adjacent normal data points using linear interpolation, thereby eliminating outlier interference caused by sensor noise or transmission errors. For example, the upper limit of valve tower surface temperature is 85 degrees Celsius, and the lower limit of air conditioning supply temperature is 18 degrees Celsius. When there are outliers in the original data that exceed the boundaries, normal data from adjacent time steps or adjacent monitoring points are used for linear interpolation to fill them in.

[0056] Normalization refers to transforming input features with different dimensions to a unified numerical range. Specifically, it can be achieved using the Min-Max linear transformation method. By scaling based on the maximum and minimum values ​​of each feature dimension, the negative impact of dimensional differences on model parameter updates is eliminated. This ensures that the sequential feature data of the input hybrid model have consistent dimensions, avoids deviations in model weight calculation caused by excessive differences in feature values, and provides a stable and efficient input basis for the rapid simulation of the temperature field of the converter valve hall.

[0057] Min-Max normalization maps the original feature values ​​to the [0,1] interval through a linear transformation, and the calculation formula is as follows: ; in, These are the original characteristic values ​​of load conditions such as valve tower heat loss, valve hall air conditioning supply air velocity, and valve hall air conditioning supply air temperature in the valve hall. and These are the minimum and maximum values ​​of the feature in the training data, respectively. These are the normalized eigenvalues.

[0058] This application employs a data cleaning mechanism combining physical constraints and linear interpolation to effectively improve the integrity and reliability of input data. The Min-Max normalization method unifies the numerical range of each feature dimension, resulting in a more stable input distribution across neural network layers, which accelerates model convergence and improves generalization performance. These technical solutions significantly enhance the training efficiency and prediction accuracy of the temperature field simulation model. The data cleaning process combines physical rule constraints with interpolation imputation, preserving real-world operating conditions while eliminating interference from outlier data in model training. Normalization eliminates dimensional differences, enabling the convolutional attention fusion model to learn the influence of different parameters on the temperature field distribution more evenly, thus achieving stable and reliable high-precision temperature field simulation in complex engineering scenarios.

[0059] Furthermore, this embodiment proposes an adaptive learning optimization logic specifically called the Adams adaptive learning optimization logic, with the Adams optimizer selected for model optimization. In the simulation of the temperature field of the converter valve hall, temperature data under some extreme operating conditions may generate large gradients. Adams can effectively avoid the excessive influence of such abnormal gradients on parameter updates, making model training more robust.

[0060] Among them, Adamx adaptive learning optimization logic refers to a variant of Adam optimizer based on the infinite norm. Specifically, it can be implemented by combining gradient pruning and exponential moving average. By constraining the upper limit of the gradient update magnitude, it prevents the gradient explosion problem caused by extreme temperature data. It performs well in gradient stability and sparse feature learning, and is suitable for modeling complex features in the temperature field simulation of the converter valve hall.

[0061] Through the above technical solutions, this application can improve the training stability of the temperature field simulation model of the converter valve hall, avoid gradient anomaly caused by extreme working condition data, and enable the model to maintain the effectiveness and controllability of parameter updates in complex temperature distribution scenarios, thereby ensuring the reliability and engineering applicability of temperature field prediction results.

[0062] The above is a step-by-step explanation of the simulation method for the temperature field of the converter valve hall provided in this application. The following is a detailed description of the simulation verification example based on the above simulation method for the temperature field of the converter valve hall.

[0063] This example involves a valve hall containing six valve towers, each with 168 power modules. AnsysWorkbench was used to geometrically simplify the model, extracting the air domain within the valve hall to construct a conjugate thermal simulation analysis model of the valve hall's fluid flow. During normal operation of the valve towers within the valve hall, the power modules on the towers generate heat losses.

[0064] To address this issue, a random sampling method is first used to divide the training and test sets in an 8:2 ratio to ensure data distribution consistency. Min-Max normalization is then performed to eliminate dimensional differences and accelerate model convergence. Next, the normalized valve hall sample data is input into the convolutional attention fusion model for training. The CNN extracts low-level temporal features through two-layer 1D convolutions, outputting a feature matrix. After dimension mapping and positional encoding embedding, the encoder calculates the association weights of different sample positions using a multi-head self-attention mechanism. The decoder calculates the matching weights between the encoded features and the generated sequence based on Encoder-Decoder attention, and then performs a weighted sum of the global features and the generated sequence features to obtain the prediction result. After model training, the model is rigorously evaluated based on multiple accuracy metrics, and the best proxy model is selected and saved. In the visualization stage, the model can achieve second-level temperature field calculations based on input conditions, accurately simulating the temperature distribution in the converter valve hall. Specific simulation results are shown below. Figure 4 As shown, in the simulation verification of the temperature field in the converter valve hall, the convolutional attention fusion model of this scheme achieved a dual improvement in efficiency and accuracy. Taking a key measuring point in the valve hall as the analysis object, the dynamic changes of the model's predicted value and the actual simulation value are highly consistent. Specifically, under different operating conditions, the mean square error of this measuring point is as low as 0.0028, and the maximum deviation does not exceed 0.73, which fully verifies the model's ability to accurately capture the dynamic fluctuations of heat loss and the temperature response law during the cooling air adjustment process. In terms of efficiency, a comparative analysis with traditional CFD simulation methods is conducted, and the comparison results are summarized in Table 1 below:

[0065] As can be seen from the data summarized in Table 1, under the same working conditions, the traditional method takes 2.5 hours to complete a single calculation, while this solution relies on the efficient collaboration between CNN local feature extraction and Transformer global correlation modeling, and the calculation time is only 0.65 seconds, with an efficiency improvement of more than 95%.

[0066] To further verify the performance superiority of the convolutional attention fusion model described in this application, ANN, RNN, GRU, and LSTM models commonly used in the field of engineering temperature field prediction were selected as benchmark comparison models. Comparative experiments were conducted under the same converter valve hall temperature field prediction task, and the loss function (Loss) change curves during the training process of each model were recorded to intuitively present the convergence effect and stability of different models. The experimental results are as follows: Figure 5 As shown: The training loss curves reveal significant differences in convergence characteristics and final performance among the various comparative models. ANN, as the basic neural network model, exhibits slow convergence during training, with its mean squared error (MSE) decreasing slowly with increasing training epochs, resulting in the highest final loss among all comparative models. This indicates that the static mapping structure of ANN has inherent limitations in capturing the temporal evolution characteristics of the temperature field. In contrast, RNN and its improved models LSTM and GRU show significant improvements in convergence speed; however, RNN is still constrained by the vanishing gradient problem, preventing effective improvement in convergence speed and resulting in a relatively high final loss. LSTM and GRU, by introducing gating mechanisms, effectively alleviate the vanishing gradient problem, thus outperforming standard RNN in both convergence speed and final performance. GRU, while retaining the advantages of gating mechanisms, simplifies the network structure and reduces computational complexity, demonstrating a more stable training process and good prediction performance in this experiment, with a final loss value superior to LSTM.

[0067] The above is a detailed description of the simulation verification example of the above-mentioned converter valve hall temperature field simulation method provided in this application. The following is a detailed description of a converter valve hall temperature field simulation device, terminal and storage medium and other related embodiments provided in this application.

[0068] Please see Figure 6 This embodiment provides a device for simulating the temperature field of a converter valve chamber, comprising: The sample data acquisition unit 201 is used to acquire the original temperature field sample data of the converter valve chamber; The fusion model training unit 202 is used to input the original temperature field sample data into the initial convolutional attention fusion model, so as to train the convolutional attention fusion model through the original temperature field sample data and the preset adaptive learning optimization logic. The convolutional attention fusion model includes a convolution operation module and an attention operation module. The convolution operation module is used to extract features from the input temperature field data step by step through a multi-layer convolutional structure, and to concatenate the extracted features along the channel dimension into a spatiotemporal feature sequence with composite feature representation and transmit it to the attention operation module. The attention operation module is used to capture the long-distance spatial dependency features of the spatiotemporal feature sequence based on the spatiotemporal feature sequence, combined with multi-head attention mechanism and feature association enhancement processing. The model convergence determination unit 203 is used to obtain the convolutional attention fusion model when the convergence degree or number of iterations of the convolutional attention fusion model meets the preset training termination condition. The temperature field simulation control unit 204 is used to input the real-time temperature condition data of the converter valve hall into the convolutional attention fusion model, so as to obtain the temperature field simulation results of the converter valve hall through the operation of the convolutional attention fusion model.

[0069] Specifically, after the sample data acquisition unit obtains raw temperature field data from sensors or a database, the fusion model training unit inputs the data into the convolutional attention fusion model for training. The convolutional operation module extracts local thermal features in the temperature field level by level through convolutional layers, such as the heat diffusion pattern in the high-temperature region of the valve tower, and splices the multi-level features along the channel dimension to form a spatiotemporal feature sequence. After receiving this sequence, the attention operation module captures the global temperature dependence of the cold air path through a multi-head attention mechanism, such as the long-range diffusion effect of air conditioning air in the valve hall. The model convergence judgment unit continuously monitors the change of the loss function during the training process and stops training when the preset convergence threshold or the maximum number of iterations is reached. The temperature field simulation control unit inputs the real-time collected operating condition data into the trained model and obtains the temperature field distribution prediction result through forward propagation calculation.

[0070] Furthermore, this application provides a converter valve hall temperature field simulation terminal, such as... Figure 7 As shown, it includes a memory 33 and a processor 31, which can be connected via a communication bus 34; the memory 33 is used to store program code, which corresponds to the simulation method of the temperature field of the converter valve hall; the processor 31 is used to read and execute the program code to implement the simulation method of the temperature field of the converter valve hall.

[0071] In this context, memory 33 refers to a semiconductor device or magnetic medium with data storage capabilities, specifically a solid-state drive or dynamic random access memory, used to persistently store the program instructions and parameter configurations required for training and predicting the temperature field simulation model. Processor 31 refers to an integrated circuit chip that performs arithmetic and logical operations, specifically a multi-core central processing unit or graphics processing unit, used to parse program code and control the flow and computation of data in the convolutional attention fusion model. Program code refers to a computer executable file composed of machine instructions, specifically written and compiled using Python or C++, containing the complete logical implementation of algorithmic steps such as data preprocessing, model training, and result output.

[0072] Specifically, the terminal loads pre-trained convolutional attention fusion model parameters into memory and constructs a computational graph structure containing convolutional and attention operation modules within the processor. When real-time temperature data is input, the processor calls preprocessing logic in the program code to clean and normalize the data. The normalized data is then input into the convolutional operation module for local thermal feature extraction, generating a spatiotemporal feature sequence. This sequence is passed to the attention operation module for global dependency modeling, and finally, the temperature field distribution prediction result is obtained through mapping via a fully connected layer. The entire computation process is executed in parallel within the processor, leveraging hardware acceleration capabilities to achieve low-latency inference.

[0073] In some specific implementations, the memory capacity can be configured to 512GB or more to support the caching requirements of large-scale temperature field sample data, and the processor frequency can be set to 3.5GHz or more to meet the requirements of real-time simulation calculations. The program code can adopt a modular design, separating the data acquisition interface, model inference engine, and result visualization components for easy maintenance and functional expansion in the later stages.

[0074] Furthermore, this application proposes a computer-readable storage medium containing program code for being read and executed by a processor to implement a method for simulating the temperature field of a converter valve chamber.

[0075] Specifically, the program code stored in the computer-readable storage medium comprises four functional modules: data acquisition, model training, convergence determination, and simulation control. When the processor reads the program code, it first executes the data acquisition module to load raw temperature field sample data from the storage medium or external sensors; then, it calls the model training module to input the sample data into the convolutional attention fusion model for iterative training, and adjusts the model parameters through adaptive learning optimization logic; next, it runs the convergence determination module to determine whether training should terminate based on a preset loss function threshold or the maximum number of iterations; finally, it executes the simulation control module to input real-time temperature condition data into the trained model and output the temperature field distribution prediction results. Throughout the process, the storage medium provides the physical basis for the static storage and dynamic loading of the program code, ensuring the repeatable execution of the simulation process.

[0076] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the terminals, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0077] In the several embodiments provided in this application, it should be understood that the disclosed terminals, devices, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection between devices or units through some interfaces, and may be electrical, mechanical, or other forms.

[0078] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the application described herein can be implemented, for example, in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0079] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0080] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0081] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0082] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0083] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for simulating the temperature field of a converter valve chamber, characterized in that, include: Obtain raw temperature field sample data of the converter valve chamber; The original temperature field sample data is input into an initial convolutional attention fusion model to train the model using the original temperature field sample data and a preset adaptive learning optimization logic. The convolutional attention fusion model includes a convolutional operation module and an attention operation module. The convolutional operation module extracts features from the input temperature field data through a multi-layer convolutional structure, concatenates the extracted features along the channel dimension into a spatiotemporal feature sequence representing a composite feature, and transmits it to the attention operation module. The attention operation module, based on the spatiotemporal feature sequence, combines a multi-head attention mechanism and feature association enhancement processing to capture the long-distance spatial dependency features of the spatiotemporal feature sequence. When the convergence degree or number of iterations of the convolutional attention fusion model meets the preset training termination condition, the convolutional attention fusion model is obtained. The real-time temperature data of the converter valve hall is input into the convolutional attention fusion model, so that the temperature field simulation results of the converter valve hall can be obtained through the calculation of the convolutional attention fusion model.

2. The method for simulating the temperature field of a converter valve chamber according to claim 1, characterized in that, The attention operation module includes: a position encoding layer, an encoder layer, and a decoder layer, wherein the encoder layer includes: a multi-head self-attention submodule and a feedforward fully connected network submodule; The position encoding layer is used to map the spatiotemporal feature sequence to the embedded feature dimension through a fully connected layer, and then convert the spatiotemporal feature sequence into the corresponding position vector through a preset sine-cosine position encoding mechanism, and then embed the position vector into the input feature matrix through element-wise addition; The multi-head self-attention submodule is used to generate a QKV matrix through linear transformation based on the input feature matrix, then divide the QKV matrix into several self-attention heads according to the embedded feature dimension, then map the high-dimensional vector formed by concatenating the self-attention heads to the embedded feature dimension, and then combine the attention weight calculation method and weighted summation method of the multi-head attention mechanism to obtain the associated feature vector. The feedforward fully connected network submodule is used to perform nonlinear enhancement processing on the associated feature vector output by the multi-head self-attention submodule and then transmit it to the decoder layer.

3. The method for simulating the temperature field of a converter valve chamber according to claim 2, characterized in that, The specific position encoding formula for the position encoding layer is as follows: ； ； In the formula, This represents the location index of the sample data, where i represents the location. The i-th dimension at position, For position The position vector located in the 2i-th dimension. The embedding feature dimension is denoted as .

4. The method for simulating the temperature field of a converter valve chamber according to claim 2, characterized in that, The function expression for the feedforward fully connected network submodule is: ； In the formula, Let x be the associated feature vector after processing by the feedforward fully connected network submodule, and let x be the associated feature vector before processing by the feedforward fully connected network submodule. For Swish activation function, and These are the weight coefficients of the two linear layers in the feedforward fully connected network submodule. and These are the bias term coefficients of the two linear layers in the feedforward fully connected network submodule, respectively.

5. The method for simulating the temperature field of a converter valve chamber according to claim 1, characterized in that, The multi-layer convolutional structure is specifically a two-layer convolutional structure.

6. The method for simulating the temperature field of a converter valve chamber according to claim 1, characterized in that, After obtaining the original temperature field sample data of the converter valve chamber, the process also includes: The original temperature field sample data is preprocessed, wherein the preprocessing includes data cleaning and normalization.

7. The method for simulating the temperature field of a converter valve chamber according to claim 1, characterized in that, The adaptive learning optimization logic is specifically: Adamax adaptive learning optimization logic.

8. A device for simulating the temperature field of a converter valve chamber, characterized in that, include: The sample data acquisition unit is used to acquire the original temperature field sample data of the converter valve chamber; The fusion model training unit is used to input the original temperature field sample data into an initial convolutional attention fusion model, so as to train the convolutional attention fusion model through the original temperature field sample data and a preset adaptive learning optimization logic. The convolutional attention fusion model includes a convolution operation module and an attention operation module. The convolution operation module is used to extract features from the input temperature field data step by step through a multi-layer convolutional structure, and to concatenate the extracted features along the channel dimension into a spatiotemporal feature sequence with composite feature representation and transmit it to the attention operation module. The attention operation module is used to capture the long-distance spatial dependency features of the spatiotemporal feature sequence based on the spatiotemporal feature sequence, combined with a multi-head attention mechanism and feature association enhancement processing. The model convergence determination unit is used to obtain the convolutional attention fusion model when the convergence degree or number of iterations of the convolutional attention fusion model meets the preset training termination condition. The temperature field simulation control unit is used to input the real-time temperature condition data of the converter valve hall into the convolutional attention fusion model, so as to obtain the temperature field simulation results of the converter valve hall through the calculation of the convolutional attention fusion model.

9. A temperature field simulation terminal for a converter valve chamber, characterized in that, include: Memory and processor; The memory is used to store program code, which corresponds to the converter valve hall temperature field simulation method as described in any one of claims 1 to 7; The processor is used to read and execute the program code to implement the method for simulating the temperature field of the converter valve hall.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium contains program code that is read and executed by a processor to implement the converter valve hall temperature field simulation method as described in any one of claims 1 to 7.

Images