Three-dimensional space field prediction method and device based on large model and data compression, equipment and medium

Abstract

The application discloses a three-dimensional space field prediction method and device based on a large model and data compression, equipment and a medium, and relates to the field of space field prediction. The method comprises the following steps: simulating to generate a dynamic three-dimensional space field data set of each irregular grid under different boundary conditions, mapping the irregular grid to a regular grid, and compressing the dynamic three-dimensional space field data by using a down-sampling method; constructing a training set according to the three-dimensional space field compressed data and the boundary conditions and the three-dimensional space field compressed data at a next historical moment, thereby training a large model pre-training network, the network combines a 3D convolutional neural network and an auto-encoder structure, and a three-dimensional space field prediction large model is obtained; and the model is applied to predict the three-dimensional space field distribution at a future moment. The application can accurately and efficiently predict a three-dimensional space field.

Description

Technical Field

[0001] This application relates to the field of space field prediction, and in particular to a method, apparatus, device and medium for three-dimensional space field prediction based on large models and data compression. Background Technology

[0002] Heat transfer problems are widely used in energy, aviation, aerospace, and machinery industries. However, due to limitations in sensor placement, it is often difficult to obtain comprehensive spatial information. With the rapid development of computer technology and fluid mechanics theory, numerical simulation has been widely applied to solving multiphase flow problems, especially in engineering practice, where computational fluid dynamics (CFD) has become a common method for spatial reconstruction of flow fields. However, CFD models are complex and computationally time-consuming, making it difficult to meet the requirements of real-time optimization control. Furthermore, the strong coupling, high nonlinearity, and time-delay characteristics of flow field information make it difficult for existing solution methods to balance accuracy and real-time performance. Current numerical simulation methods are mostly used for high-precision calculations of the static state at specific operating points, and their application in dynamic regulation and real-time operation has certain limitations.

[0003] Currently, most studies on spatial flow fields rely on traditional steady-state prediction methods. For example, in thermal power, these methods are mainly used to analyze steady-state data of NOx distribution, temperature field, and velocity field, with a focus on the spatial distribution characteristics of various physical quantities under steady-state conditions. However, combustion is inherently a dynamic flow process, and in practical applications, it often exhibits complex time-varying characteristics. Therefore, in-depth analysis and prediction of dynamic heat transfer processes have significant theoretical and practical implications. Traditional methods often rely on simulation software such as Ansys to construct flow field information. This involves first simulating the object's operating state by dividing the grid, setting boundary conditions, and configuring simulation parameters. However, this method is often computationally slow. For instance, simulating the entire process from boiler ignition to steady-state can take up to 80,000 seconds of computation time, making traditional simulation methods inefficient for dynamic prediction. Therefore, there is an urgent need to explore a highly efficient and accurate dynamic modeling and prediction method to effectively guide the dynamic control of heat transfer processes. Unlike steady-state prediction, dynamic prediction needs to consider spatiotemporal variations, involving higher-dimensional data and complex time-series analysis. However, due to the spatiotemporal superposition effect, the amount of data involved in dynamic prediction usually increases significantly, which poses a huge challenge to model training. Common neural network methods often require a long training time when dealing with large-scale spatiotemporal data, and may fail to converge or learn sufficiently due to excessive data volume or improper network design. Summary of the Invention

[0004] The purpose of this application is to provide a method, apparatus, device, and medium for predicting three-dimensional spatial fields based on large models and data compression, which can accurately and efficiently predict three-dimensional spatial fields.

[0005] To achieve the above objectives, this application provides the following solution:

[0006] Firstly, this application provides a method for predicting three-dimensional spatial fields based on large models and data compression, including:

[0007] Simulations are performed using an object model with an irregular mesh to generate a dynamic three-dimensional spatial field dataset for each irregular mesh under different boundary conditions; the dynamic three-dimensional spatial field dataset includes dynamic three-dimensional spatial field data at multiple historical moments; the object model is a structural model that generates the three-dimensional spatial field.

[0008] Map the dynamic 3D spatial field data in each irregular grid dynamic 3D spatial field dataset to dynamic 3D spatial field data in a regular grid;

[0009] The dynamic 3D spatial field data of each regular grid is compressed using a downsampling method to obtain the 3D spatial field compressed data of each regular grid at multiple historical moments;

[0010] Define the input and output formats;

[0011] Based on the 3D spatial field compression data and boundary conditions of each regular grid at a preset number of historical moments, input data is obtained according to the input format, and based on the 3D spatial field compression data of each regular grid at the next historical moment, output data is obtained according to the output format. The input data and output data together constitute the training set; the next historical moment is the moment after the last moment in the preset number of historical moments.

[0012] The training set is used to train the large model pre-training network to obtain a large model for predicting three-dimensional spatial fields; the large model pre-training network combines a 3D convolutional neural network and an autoencoder structure.

[0013] Based on the spatial field data of each regular grid at a preset number of historical moments under actual boundary conditions and the actual boundary conditions, the input data is obtained according to the input format and input into the three-dimensional spatial field prediction large model, and the three-dimensional spatial field compressed data of each regular grid at future moments is output.

[0014] After upsampling the compressed 3D spatial field data of each regular grid at a future time, it is mapped to an irregular grid to obtain the 3D spatial field distribution at the future time.

[0015] Optionally, a downsampling method is used to compress the dynamic three-dimensional spatial field data of each regular grid to obtain the three-dimensional spatial field compressed data of each regular grid at multiple historical moments. Specifically, this includes: using a downsampling method to reduce the dimension of the dynamic three-dimensional spatial field data of each regular grid, and using a cubic interpolation method to generate the three-dimensional spatial field compressed data of each regular grid at multiple historical moments.

[0016] Optionally, the input format of the three-dimensional spatial field compressed data in the input data is x×y×z×N×N t The input format for the boundary conditions in the input data is (N c ,1); where x, y, and z represent the dimensions of the three-dimensional spatial field compression data in the X, Y, and Z directions, respectively, and N represents the number of three-dimensional spatial fields to be predicted. t N represents the preset number of historical moments; c The number of boundary conditions is indicated; the output format is x×y×z×N×1.

[0017] Optionally, the large model pre-trained network includes an encoder, a Transformer module, and a decoder connected in sequence; the encoder is used to receive 3D spatial field compressed data in the input data and extract the spatial features of the 3D spatial field compressed data; the Transformer module adopts a multi-head self-attention mechanism to model the interrelationships between the spatial features and capture long-distance dependencies to obtain a feature map; the decoder is used to receive the feature map and boundary conditions in the input data, concatenate the feature map with the boundary conditions, and obtain output data according to the output format based on the concatenated feature map and boundary conditions.

[0018] Optionally, the encoder includes a stacked 3D convolutional layer, a first residual layer, a downsampling layer, and a first LeakyReLU activation function connected in sequence; the decoder includes a fully connected layer, a deconvolutional layer, a second residual layer, an upsampling layer, a second LeakyReLU activation function, and an output layer connected in sequence; the Transformer module is connected between the first LeakyReLU activation function and the deconvolutional layer.

[0019] Optionally, when the object model is a boiler furnace model, the boundary conditions include primary air velocity, secondary air velocity, burnout air velocity, coal feed rate, and coal feed rate.

[0020] Optionally, the three-dimensional spatial field is one or more of a temperature field, a velocity field, and a concentration field.

[0021] Secondly, this application provides a three-dimensional spatial field prediction device based on a large model and data compression, comprising:

[0022] The simulation module is used to perform simulations using an object model divided into irregular meshes, generating a dynamic three-dimensional spatial field dataset for each irregular mesh under different boundary conditions; the dynamic three-dimensional spatial field dataset includes dynamic three-dimensional spatial field data at multiple historical moments; the object model is a structural model that generates the three-dimensional spatial field.

[0023] The mapping module is used to map the dynamic 3D spatial field data in each irregular grid dynamic 3D spatial field dataset to dynamic 3D spatial field data of regular grids;

[0024] The compression module is used to compress the dynamic three-dimensional spatial field data of each regular grid using a downsampling method, so as to obtain the three-dimensional spatial field compressed data of each regular grid at multiple historical moments;

[0025] Define the module, which is used to define the input and output formats;

[0026] The module is used to obtain input data according to the input format based on the three-dimensional spatial field compression data and boundary conditions of each regular grid at a preset number of historical moments, and to obtain output data according to the output format based on the three-dimensional spatial field compression data of each regular grid at the next historical moment. The input data and output data together constitute the training set; the next historical moment is the moment after the last moment in the preset number of historical moments.

[0027] The training module is used to train the large model pre-trained network using the training set to obtain a large model for predicting three-dimensional spatial fields; the large model pre-trained network combines a 3D convolutional neural network with an autoencoder structure.

[0028] The prediction module is used to obtain input data according to the input format based on the spatial field data of each regular grid at a preset number of historical times under actual boundary conditions and the actual boundary conditions, and input the data into the three-dimensional spatial field prediction model, and output the three-dimensional spatial field compressed data of each regular grid at future times.

[0029] The recovery module is used to upsample the compressed 3D spatial field data of each regular grid at a future time and then map it to an irregular grid to obtain the 3D spatial field distribution at a future time.

[0030] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the three-dimensional spatial field prediction method based on large model and data compression as described above.

[0031] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the three-dimensional spatial field prediction method based on large model and data compression as described above.

[0032] According to the specific embodiments provided in this application, this application has the following technical effects:

[0033] This application provides a method, apparatus, device, and medium for predicting three-dimensional spatial fields based on large models and data compression. The large model pre-trained network combines a 3D convolutional neural network with an autoencoder structure, enabling it to learn information about the three-dimensional spatial field at multiple time points simultaneously and accurately predict the three-dimensional spatial field at future time points. By mapping irregular grid dynamic three-dimensional spatial field data to regular grids and downsampling, the spatial resolution of the data can be significantly reduced, thereby reducing the computational burden on the large model pre-trained network during training and prediction, and significantly improving computational efficiency. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the embodiments 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.

[0035] Figure 1 A flowchart illustrating a three-dimensional spatial field prediction method based on a large model and data compression, provided as an embodiment of this application;

[0036] Figure 2 This is a schematic diagram of the network structure of a large-scale pre-trained model provided in an embodiment of this application;

[0037] Figure 3 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0038] 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.

[0039] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0040] In one exemplary embodiment, such as Figure 1 As shown, a method for predicting three-dimensional spatial fields based on large models and data compression is provided, including the following steps 101 to 108. Wherein:

[0041] Step 101: Simulate using an object model with an irregular mesh to generate a dynamic three-dimensional spatial field dataset for each irregular mesh under different boundary conditions; the dynamic three-dimensional spatial field dataset includes dynamic three-dimensional spatial field data at multiple historical moments; the object model is a structural model that generates the three-dimensional spatial field.

[0042] Step 102: Map the dynamic 3D spatial field data in the dynamic 3D spatial field dataset of each irregular grid to dynamic 3D spatial field data of a regular grid.

[0043] Step 103: Compress the dynamic 3D spatial field data of each regular grid using a downsampling method to obtain the 3D spatial field compressed data of each regular grid at multiple historical moments.

[0044] Step 104: Define the input and output formats.

[0045] Step 105: Based on the 3D spatial field compression data and boundary conditions of each regular grid at a preset number of historical moments, obtain input data according to the input format, and based on the 3D spatial field compression data of each regular grid at the next historical moment, obtain output data according to the output format. The input data and output data together constitute the training set; the next historical moment is the moment after the last moment in the preset number of historical moments.

[0046] Step 106: Train the large model pre-training network using the training set to obtain a large model for predicting three-dimensional spatial fields; the large model pre-training network combines a 3D convolutional neural network with an autoencoder structure.

[0047] Step 107: Based on the spatial field data of each regular grid at a preset number of historical moments under the actual boundary conditions and the actual boundary conditions, obtain the input data according to the input format, and input it into the three-dimensional spatial field prediction large model, and output the three-dimensional spatial field compressed data of each regular grid at future moments.

[0048] Step 108: After upsampling the compressed 3D spatial field data of each regular grid at a future time, map it to the irregular grid to obtain the 3D spatial field distribution at the future time.

[0049] In another exemplary embodiment of this application, the simulation in step 101 is specifically implemented as follows: a model of the research object is constructed based on Ansys software, and a spatial mesh is generated through mesh generation. Simulations are performed using the Fluent module for different boundary conditions to obtain dynamic three-dimensional spatial field data. Subsequently, the spatial field fluid information and mesh coordinates are exported using the Post module. The spatial field fluid information includes physical quantities related to fluid flow and heat transfer, such as velocity field, temperature field, pressure field, turbulence intensity, fluid density, and energy transfer. This information reflects the dynamic state of the fluid throughout the entire research area. Dynamic three-dimensional spatial field data refers to the fluid state at various moments in the three-dimensional space that change over time, typically including the changing trends of the aforementioned fluid information over time. It is a time-series dataset of spatial field fluid information. The mesh coordinates are used to locate and map the simulation calculation results to specific spatial locations. In subsequent spatial fluid information field prediction, it is necessary to correlate the simulation results (such as temperature, velocity, pressure, etc.) with the mesh coordinates to ensure that the results correctly reflect the spatial location. In addition, the mesh coordinates help determine the specific location of each simulation node and visualize the fluid data. By combining grid coordinates with fluid information, the distribution of flow field, temperature field, pressure field, etc., can be presented in three-dimensional space.

[0050] In another exemplary embodiment of this application, the specific implementation of step 102 mapping is as follows: In order to input the spatial field fluid information into the model, grid data compression is required in this step. Since the spatial coordinates of each point in the original temporal spatial field data constitute an irregular grid, and the shape of the research object is mostly irregular, it needs to be converted into a regular spatial three-dimensional grid to facilitate network input processing. When processing the scattered coordinate data of the research object, all points are first sorted in ascending order according to the X, Y, and Z dimensions to generate a table containing coordinates and feature values. Then, these sorted data are mapped onto the regular grid. The data is first arranged step by step in the XY plane, and after filling one layer, it is then superimposed layer by layer in the Z dimension to finally form a cubic grid. The number of points in the spatial three-dimensional grid that is greater than that in the irregular grid is filled with the minimum value of the spatial field information.

[0051] In another exemplary embodiment of this application, to reduce computational costs and optimize training efficiency, a downsampling method is used to reduce the spatial field data at each time step to 36×36×36. Downsampling reduces the amount of data by lowering the resolution or size of the data, thereby reducing the computational burden while maintaining prediction accuracy.

[0052] Downsampling employs an interpolation method, specifically cubic interpolation within the `scipy.ndimage.zoom` function. Cubic interpolation, based on four adjacent data points, smooths the transitions between data points by fitting a cubic polynomial, thus generating a smooth curve. This method is suitable for processing continuous data, such as images.

[0053] Specifically, the cubic interpolation formula is as follows:

[0054] f(x) = a3x 3 +a2x 2 +a1x+a0;

[0055] Where x is the target position (spatial coordinates after downsampling), and a3, a2, a1 and a0 are coefficients calculated from known points.

[0056] In another exemplary embodiment of this application, in step 104 above, the input format of the three-dimensional spatial field compression data in the input data is x×y×z×N×N t The input format for boundary conditions in the input data is (N c ,1); The output format is x×y×z×N×1. Where x, y, and z represent the dimensions of the 3D spatial field compressed data in the X, Y, and Z directions, respectively, and N represents the number of 3D spatial fields to be predicted. t N represents the preset number of historical moments; c Indicates the number of boundary conditions.

[0057] For example, when the dimension of the three-dimensional spatial field compression data is 36×36×36, the input format of the three-dimensional spatial field compression data is 36×36×36×N×N. t The output format is 36×36×36×N×1. Assuming N is 1, meaning only one spatial field needs to be predicted, the training set format is 36×36×36×N. t The test set format is 36×36×36×1.

[0058] In another exemplary embodiment of this application, a neural network is a complex system composed of a large number of widely connected neurons. It acquires knowledge by learning from known samples and stores this knowledge in a distributed manner in the network weights. With most key parameters of flow field information now available in real time, neural networks have gradually become a representative modeling tool in flow field spatial modeling. This method departs from the mechanistic analysis of traditional heat transfer processes, instead training the neural network with actual collected data samples, enabling it to learn and reveal the potential inherent laws between input and output data, thereby reflecting the characteristics of the heat transfer process. This data-driven modeling method not only overcomes the complexity and limitations of traditional mechanistic model construction but also demonstrates strong adaptability and accuracy in practical applications. However, with the continuous deepening of theoretical research and the expansion of application fields, neural networks have gradually revealed some limitations: such as dependence on large samples, overfitting, nondeterminism of network structure, slow learning speed, and training failures due to getting trapped in local maxima.

[0059] Further improvements to the neural network model structure include employing more efficient training strategies, such as using spatiotemporal convolutional networks for feature extraction and incorporating adaptive mechanisms into the network structure to enhance training efficiency and prediction accuracy. Furthermore, dynamic prediction requires the model to adapt to changes in input data in real time, ensuring effective handling of complex and nonlinear heat transfer processes.

[0060] The application of large-scale models in 3D modeling tasks is still in the exploratory stage, but they exhibit significant advantages compared to traditional small-scale models. Through a large number of parameters, tokenization, and adaptive feature extraction and fine-tuning capabilities based on attention mechanisms, large-scale models can more accurately capture spatial and temporal features when handling complex tasks, providing higher accuracy and generalization ability. Tokenization enables them to effectively handle high-dimensional data, while the attention mechanism enhances the ability to focus on key features, especially excelling in spatiotemporal data modeling. Although training resources are more demanding, the flexibility, accuracy, and adaptability of large-scale models make them promising candidates for handling large-scale complex tasks. With the continuous advancement of computing resources, large-scale models will demonstrate even greater potential in 3D dynamic modeling.

[0061] Furthermore, autoencoders, as an efficient data compression structure, effectively reduce the dimensionality of input data through encoding and decoding processes, thereby reducing the computational burden during model training and prediction. The encoder part of an autoencoder maps high-dimensional data to a low-dimensional latent space through a series of nonlinear transformations, extracting key features from the latent space. This compression not only reduces data storage and processing requirements but also eliminates redundant information, ensuring data quality while improving the model's prediction efficiency. By simplifying data dimensionality, autoencoders reduce the complexity of input data, making subsequent decoding and prediction processes more efficient. Therefore, using autoencoders for data compression can not only optimize the utilization of computational resources but also accelerate the prediction process, especially when dealing with large-scale high-dimensional data, significantly improving the overall system performance and real-time performance.

[0062] To enable the network to learn the characteristics of spatial field fluid information, multi-dimensional feature learning is required through network layer design. This application designs a multi-layered, large-scale pre-trained model architecture that combines a 3D Convolutional Neural Network (3DCNN) with an autoencoder structure to process network data with spatiotemporal characteristics. The model includes an encoder, a decoder, and an intermediate Transformer module, aiming to perform time-series predictions using historical spatial field data and control parameters. This architecture aims to capture the complexity of spatiotemporal data, thereby improving the predictive ability for dynamic system behavior. The network structure of the large-scale pre-trained model is as follows: Figure 2 As shown, the encoder receives compressed 3D spatial field data from the input data and extracts its spatial features. The Transformer module employs a multi-head self-attention mechanism to model the relationships between the spatial features and capture long-distance dependencies to obtain a feature map. The decoder receives the feature map and boundary conditions from the input data, concatenates the feature map with the boundary conditions, and obtains output data according to the output format based on the concatenated feature map and boundary conditions.

[0063] Encoder design: The encoder consists of a stacked 3D convolutional layer, a first residual layer, a downsampling layer, and a first LeakyReLU activation function connected in sequence.

[0064] The encoder extracts spatial features from the input grid using 3D convolutional layers and residual connections. By stacking convolutional layers, the encoder progressively extracts high-level features and compresses the data dimensionality. Specifically, convolutional layers extract local spatial features through convolution operations, with an input size of 36×36×36, and the size of the output feature map of each layer is progressively compressed. Residual blocks effectively avoid the vanishing gradient problem during deep network training through skip connections.

[0065] In residual learning, the network learns not the complete output y, but the difference between the input x and an offset (i.e., the residual). This difference is modeled through convolutional layers in the network. Each residual block contains two convolutional layers, where the output of the second convolutional layer is added to the input of the first convolutional layer, forming a residual connection. The specific implementation steps are as follows: First, the input x passes through the first convolutional layer (e.g., Conv3D) to obtain the feature map conv1(x); then, the second convolutional layer applies a second convolutional operation (e.g., Conv3D) to further process the output conv1(x) of the first convolutional layer, obtaining the feature map conv2(x). Finally, a skip connection is used to add the input x to the output of the second convolutional layer as the final output y, calculated as follows:

[0066] y = conv2(conv1(x)) + x;

[0067] In this way, residual blocks can effectively alleviate the gradient vanishing problem in deep networks and improve the training effect of the model.

[0068] The output after the residual connection passes through a downsampling layer. This downsampling layer progressively reduces the spatial resolution of the input through each downsampling block (using convolution operations and convolutions with a stride of 2), while simultaneously introducing non-linear features through the LeakyReLU activation function, thereby enhancing the model's expressive power. In the encoder section, a Transformer layer is incorporated as an improvement, further enhancing the feature representation capability, especially when handling complex temporal data and control parameters. This Transformer module employs a multi-head attention mechanism, which effectively models the relationships between input features and captures long-distance dependencies, thus enhancing the model's performance in complex tasks.

[0069] The Transformer, as the foundational framework for large model architectures, leverages its efficient parallel computing and modeling capabilities to effectively handle long-range dependencies. This Transformer module employs a multi-head attention mechanism, which effectively models the relationships between input features and captures long-distance dependencies, thereby enhancing the model's performance in complex tasks. The Transformer mechanism is as follows:

[0070]

[0071] Here, Q is the query matrix, representing the information that needs to be retrieved from the current input vector. The dimension of Q is typically (n, d). k ), where n is the length of the input sequence, d k This refers to the query dimension. K is a key matrix, representing the "key" of each input element, used to match it against the query matrix and calculate relevance. The dimension of K is also (n, d). k V is a value matrix, representing the actual information corresponding to each key. The dimensions of V are typically (n, d). v ), where d v It is the dimension of the value. QK T This represents the dot product of the query matrix and the key matrix. This dot product calculates the similarity between each pair of queries and keys, resulting in an (n,n) matrix that represents the relationship between each position in the input sequence. Indicates prevention of QK T When the inner product is too large, it is usually used for scaling, d k The key dimension helps avoid the vanishing gradient problem during softmax computation. The softmax operation normalizes the scaled dot product, calculates the importance (or weight) of each position, and ensures that the weights sum to 1. The softmax function is applied to the result of each row, producing an (n,n) matrix representing the degree of match between the query and the key. The final attention output is obtained by multiplying the softmax-normalized weights by the value matrix V. This operation aggregates information, resulting in a weighted sum representation. Through a multi-head mechanism, multiple attention outputs can be computed in parallel and concatenated to further improve the model's learning and expressive capabilities.

[0072] Decoder Design: The decoder consists of a fully connected layer, a deconvolutional layer, a second residual layer, an upsampling layer, a second LeakyReLU activation function, and an output layer, connected in sequence. The Transformer module is connected between the first LeakyReLU activation function and the deconvolutional layer.

[0073] The decoder receives the encoder's output and control parameters as input. The control parameters (e.g., airflow, coal consumption) are first expanded to the same spatial dimension as the encoder output through a fully connected (Dense) layer, and then concatenated spatially with the encoder's output feature map. Next, the decoder gradually restores the spatial resolution through a series of convolutional operations and transposed convolutions, ultimately generating the reconstructed grid data. This includes deconvolutional layers, residual blocks, upsampling layers, the LeakyReLU activation function, and an output layer. The deconvolutional layers upsample the encoder's output features, restoring the low-dimensional feature map to its original spatial resolution. Residual blocks help prevent information loss and make the output smoother and more accurate. The upsampling layers progressively enlarge the feature map, gradually restoring the low-dimensional feature map to the original data size. The LeakyReLU activation function helps the network maintain the diversity and fluidity of information during upsampling, avoiding information loss in certain regions due to zero activation. The output layer can be configured according to task requirements.

[0074] In the processing of control parameters, the control parameters are first expanded through a fully connected layer, and then transformed into a shape suitable for concatenation with the encoder output through a reshape layer. This ensures that the control parameters can be effectively fused with the feature map of the encoder output, improving the decoder's ability to reconstruct the input data. The reshape layer is a commonly used operation in neural networks, its function being to change the shape of data. It is often used to transform the data dimension between different network layers so that the network can handle different data representations.

[0075] Then, through transpose convolution (also known as deconvolution), the spatial dimensions of the input data are gradually restored. The formula is shown below:

[0076] Conv3DTranspose(x)=upsampling(x)×W;

[0077] Here, upsampling(x) is the upsampling operation, and W is the convolution kernel.

[0078] The entire model employs an autoencoder architecture, aiming to compress the input grid data using an encoder and then reconstruct it into an output similar to the target grid using a decoder. The autoencoder takes historical temperature field data containing control parameters as input and outputs a prediction of the future temperature field. Training the autoencoder is typically optimized by minimizing the reconstruction loss, with the mean squared error (MSE) being a commonly used loss function. Its formula is as follows:

[0079]

[0080] Among them is The decoder's predicted output, y i N represents the true value, and N is the number of data samples.

[0081] The spatial field at the next moment is predicted using the large-scale network model described above, with a size of (36,36,36).

[0082] In another exemplary embodiment of this application, after obtaining the three-dimensional spatial field compressed data of each regular grid at a future time in step 107, it is restored to the original three-dimensional grid size through upsampling. Upsampling refers to restoring the size of the original data or generating a finer-grained feature map by increasing the spatial resolution of the input data. Its basic goal is to estimate the new high-resolution data through interpolation methods to minimize the errors generated during the interpolation process. In this study, the cubic interpolation method is used. This method is widely used to expand the spatial dimension of image or grid data, aiming to maintain the smoothness of the data and reduce the errors caused by interpolation. Assuming the original grid is (x,y,z) and the target grid is upsampled_grid(x',y',z'), its interpolation process can be expressed as:

[0083] upsample_grid(x',y',z')=CubicInterpolation(grid,x',y',z').

[0084] By combining the coordinates (X,Y,Z) of the original grid with the prediction results, the spatial grid is mapped, and the predicted spatial temperature field is restored to the irregular grid of the original object, thus completing the prediction task.

[0085] The following section uses the dynamic prediction of the temperature field in the boiler furnace space as an example to provide a more detailed introduction to the method of this application.

[0086] 1. A boiler furnace model was constructed using Ansys software, and meshing was performed, generating a total of 1.88 million irregular spatial grids. Simulations were conducted using the Fluent module under different boundary conditions (including primary air velocity, secondary air velocity, burnout air velocity, coal feed rate, and coal feed quantity) to obtain furnace spatial field data from ignition to steady state. The simulation results were then exported using the Post module. To reduce the computational burden during prediction and avoid prediction bias caused by the number of boundary grids being much larger than the number of grids in the furnace center region, a strategy was adopted to take a set of XZ plane grids at fixed intervals along the Y-axis. Specifically, most of the data within the Y-axis height range of 0 to 38 meters was selected, thus compressing the number of prediction grid points to 110,000, ensuring a more balanced distribution of grids in the boundary and center regions, and avoiding severe uneven distribution problems.

[0087] 2. The original temporal spatial field data contains 110,446 data points, where the X, Y, and Z coordinates form an irregular grid, and the furnace shape is also irregular. To facilitate input processing for the network, it needs to be converted into a regular 48×48×48 grid. To reduce computational costs and optimize training efficiency, a downsampling method was used to reduce the spatial field data at each time step from 48×48×48 to 36×36×36. Downsampling reduces the amount of data by lowering the resolution or size of the data. In this set of experiments, the data resolution was reduced by approximately 25%, thereby reducing the computational burden while maintaining prediction accuracy, providing more data for subsequent training and prediction by the neural network.

[0088] 3. Taking the dynamic prediction of the three-dimensional temperature field of furnace combustion as an example, 80% of the simulation data is used to construct the training set. The data format is 36×36×36×5, where N represents 1 because only one temperature field needs to be predicted, while the historical time is N. t The training set labels consist of spatial field data for the next time step, with a data dimension of 36×36×36×1. Furthermore, the input vector for the control parameters has a dimension of (5,1) and contains five control parameters: primary wind speed, secondary wind speed, burnout wind speed, coal feed rate, and coal feed amount.

[0089] 4. Input the training set data into the designed large-scale pre-trained network to obtain the prediction results, namely the temperature field information distribution at future times. And through upsampling, the 36×36×36 grid is restored to 48×48×48.

[0090] 5. By combining the coordinates of the original furnace grid with the prediction results, the spatial grid mapping is completed. When reading the sorted coordinate data list from the previously stored grid file, this data contains the coordinates (X, Y, Z) of the original furnace grid and their corresponding eigenvalues ​​(prediction results). After mapping this data to an irregular grid, the prediction work is completed. Specifically, according to the previously sorted coordinate list, each prediction result is assigned back to its corresponding coordinate position in the original grid. To ensure the accuracy of the mapping, it is necessary to ensure that the sorted coordinates correspond one-to-one with each point in the original grid. That is, in the previous data processing, the coordinates were mapped to the cubic grid layer by layer according to the XY plane and Z axis. Now, it is necessary to keep these mapping relationships consistent, restore the data to list form in a hierarchical manner, and then match and fill back the coordinate points of the original grid one by one. The predicted spatial temperature field is restored to the irregular grid of the original furnace, thus completing the prediction task.

[0091] The beneficial effects of the method described in this application are as follows:

[0092] 1. This application proposes a multi-dimensional feature learning method that combines 3DCNN with an autoencoder and incorporates a Transformer architecture to effectively extract multi-dimensional features of fluid temperature fields. The 3DCNN captures local spatial features in the spatial data, while the autoencoder performs feature compression and reconstruction. The Transformer architecture further enhances the model's ability to process temporal data, particularly in capturing long-term dependencies and complex temporal variations. Through this architecture, the network can simultaneously learn information about the spatial temperature field at multiple time points, thereby accurately predicting future temporal spatial fields.

[0093] 2. This application proposes a grid compression operation aimed at effectively reducing computational resource consumption. By performing 3D grid mapping and overall downsampling on the input irregular grid data, the spatial resolution of the data can be significantly reduced, thereby reducing the computational burden on the network during training and prediction. Through this process, the network can complete efficient spatiotemporal feature learning and prediction at a lower computational cost. This method significantly improves computational efficiency without significantly sacrificing prediction accuracy, especially when dealing with large-scale spatiotemporal data, effectively avoiding computational bottlenecks caused by excessive data volume, thus providing a more efficient and feasible solution for practical applications.

[0094] 3. This application proposes a hierarchical architecture for a large-scale network processing model, aiming to improve the accuracy and efficiency of feature capture. By designing a multi-layered, deep network structure, this architecture can process data at different scales and levels, thereby achieving more accurate feature extraction. Through this hierarchical processing and layer-by-layer optimization strategy, this invention not only enhances the model's adaptability to complex data but also improves computational efficiency when processing large-scale datasets, providing strong support for high-precision spatiotemporal prediction.

[0095] Based on the same inventive concept, this application also provides a three-dimensional spatial field prediction device based on large models and data compression for implementing the three-dimensional spatial field prediction method based on large models and data compression described above. The solution provided by this device is similar to the implementation described in the above method. Therefore, the specific limitations of one or more embodiments of the three-dimensional spatial field prediction device based on large models and data compression provided below can be found in the limitations of the three-dimensional spatial field prediction method based on large models and data compression described above, and will not be repeated here.

[0096] In one exemplary embodiment, a three-dimensional spatial field prediction device based on a large model and data compression is provided, comprising: a simulation module, a mapping module, a compression module, a definition module, a composition module, a training module, a prediction module, and a recovery module.

[0097] The simulation module is used to perform simulations using an object model divided into irregular meshes, generating a dynamic three-dimensional spatial field dataset for each irregular mesh under different boundary conditions; the dynamic three-dimensional spatial field dataset includes dynamic three-dimensional spatial field data at multiple historical moments; the object model is a structural model that generates the three-dimensional spatial field.

[0098] The mapping module is used to map the dynamic 3D spatial field data in each irregular grid dynamic 3D spatial field dataset to dynamic 3D spatial field data with regular grids.

[0099] The compression module is used to compress the dynamic 3D spatial field data of each regular grid using a downsampling method, thereby obtaining the 3D spatial field compressed data of each regular grid at multiple historical moments.

[0100] Define the module, which is used to define the input and output formats.

[0101] The module is used to obtain input data according to the input format based on the three-dimensional spatial field compression data and boundary conditions of each regular grid at a preset number of historical moments, and to obtain output data according to the output format based on the three-dimensional spatial field compression data of each regular grid at the next historical moment. The input data and output data together constitute the training set; the next historical moment is the moment after the last moment in the preset number of historical moments.

[0102] The training module is used to train the large model pre-trained network using the training set to obtain a large model for predicting three-dimensional spatial fields; the large model pre-trained network combines a 3D convolutional neural network with an autoencoder structure.

[0103] The prediction module is used to obtain input data according to the input format based on the spatial field data of each regular grid at a preset number of historical times under actual boundary conditions and the actual boundary conditions, and input the data into the three-dimensional spatial field prediction model, and output the three-dimensional spatial field compressed data of each regular grid at future times.

[0104] The recovery module is used to upsample the compressed 3D spatial field data of each regular grid at a future time and then map it to an irregular grid to obtain the 3D spatial field distribution at a future time.

[0105] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 3As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs in the non-volatile storage media to run. The database stores the three-dimensional spatial field distribution at future times. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When executed by the processor, the computer program implements a three-dimensional spatial field prediction method based on a large model and data compression.

[0106] Those skilled in the art will understand that Figure 3 The structures shown are merely block diagrams of some structures related to the present application and do 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 shown in the figures, or combine certain components, or have different component arrangements. In an exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0107] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0108] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0109] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A three-dimensional space field prediction method based on a large model and data compression, characterized by, The three-dimensional space field prediction method is used for dynamic prediction of a boiler furnace space temperature field, and the three-dimensional space field prediction method comprises the following steps: An object model of a divided irregular grid is used for simulation to generate a dynamic three-dimensional space field data set of each irregular grid under different boundary conditions; the dynamic three-dimensional space field data set comprises dynamic three-dimensional space field data of multiple historical time points; the object model is a structure model generating a three-dimensional space field; the object model is a boiler furnace model; different boundary conditions include primary air speed, secondary air speed, overfire air speed, coal supply rate and coal supply amount; the dynamic three-dimensional space field data set comprises furnace space field data from ignition to a stable state; Dynamic three-dimensional space field data in the dynamic three-dimensional space field data set of each irregular grid is mapped into dynamic three-dimensional space field data of a regular grid; A downsampling method is used to compress the dynamic three-dimensional space field data of each regular grid to obtain three-dimensional space field compressed data of each regular grid at multiple historical time points; An input format and an output format are defined; According to the three-dimensional space field compressed data of each regular grid at a preset number of historical time points and the boundary conditions, input data is obtained according to the input format, and output data is obtained according to the three-dimensional space field compressed data of each regular grid at a next historical time point according to the output format, and the input data and the output data together constitute a training set; the next historical time point is a next time point of the last time point in the preset number of historical time points; The training set is used to train a large model pre-training network to obtain a three-dimensional space field prediction large model; the large model pre-training network combines a 3D convolutional neural network and an autoencoder structure; According to the space field data of each regular grid at a preset number of historical time points and actual boundary conditions, input data is obtained according to the input format, and the three-dimensional space field prediction large model is input, and three-dimensional space field compressed data of each regular grid at a future time point is output; After upsampling the three-dimensional space field compressed data of each regular grid at the future time point, the three-dimensional space field compressed data is mapped to an irregular grid to obtain a three-dimensional space field distribution at the future time point.

2. The large model and data compression based three-dimensional space field prediction method according to claim 1, characterized in that, The downsampling method is used to compress the dynamic three-dimensional space field data of each regular grid to obtain three-dimensional space field compressed data of each regular grid at multiple historical time points, specifically including: The downsampling method is used to reduce the dimension of the dynamic three-dimensional space field data of each regular grid, and a cubic interpolation method is used to generate three-dimensional space field compressed data of each regular grid at multiple historical time points. 3.The method of claim 1, wherein, The input format of the three-dimensional space field compression data in the input data is x × y × z × N × N t The input format of the boundary condition in the input data is N c ,1) wherein, x , y , z Xi, Yj, Zk respectively represent the dimension of the three-dimensional space field compression data in the X direction, the Y direction, and the Z direction, N represents the number of three-dimensional space fields that need to be predicted, N t represents a preset number of historical moments; N c represents the number of boundary conditions; The output format is x × y × z × N ×1. 4.The method of claim 1, wherein, The large model pre-training network comprises an encoder, a Transformer module and a decoder connected in sequence; The encoder is used to receive three-dimensional space field compressed data in the input data and extract spatial features of the three-dimensional space field compressed data; The Transformer module adopts a multi-head self-attention mechanism to model the mutual relationship between the spatial features and capture long-distance dependencies to obtain a feature map; The decoder is used to receive the feature map and the boundary conditions in the input data, splice the feature map and the boundary conditions, and obtain output data according to the spliced feature map and the boundary conditions according to the output format.

5. The large model and data compression based three-dimensional space field prediction method according to claim 4, characterized in that, The encoder comprises a stacked 3D convolution layer, a first residual layer, a down-sampling layer and a first LeakyReLU activation function connected in sequence; The decoder comprises a full connection layer, a de-convolution layer, a second residual layer, an up-sampling layer, a second LeakyReLU activation function and an output layer connected in sequence; A Transformer module is connected between the first LeakyReLU activation function and the de-convolution layer.

6. The large model and data compression based three-dimensional space field prediction method according to claim 1, wherein, When the object model is a boiler furnace model, the boundary conditions include primary air speed, secondary air speed, overfire air speed, coal feed rate and coal feed amount.

7. The large model and data compression based three-dimensional space field prediction method according to claim 1, characterized in that, The three-dimensional space field is one or more of a temperature field, a velocity field and a concentration field.

8. A three-dimensional space field prediction device based on a large model and data compression, characterized by, The three-dimensional space field prediction device based on a large model and data compression is used for dynamic prediction of a boiler furnace space temperature field, and comprises: a simulation module configured to simulate an object model divided into irregular meshes to generate a dynamic three-dimensional space field data set of each irregular mesh under different boundary conditions; the dynamic three-dimensional space field data set comprises dynamic three-dimensional space field data at multiple historical time points; the object model is a structure model generating a three-dimensional space field; the object model is a boiler furnace model; the different boundary conditions include primary air speed, secondary air speed, overfire air speed, coal feed rate and coal feed amount; the dynamic three-dimensional space field data set comprises furnace space field data from ignition to a stable state; a mapping module configured to map the dynamic three-dimensional space field data in the dynamic three-dimensional space field data set of each irregular mesh to dynamic three-dimensional space field data of a regular mesh; a compression module configured to compress the dynamic three-dimensional space field data of each regular mesh using a down-sampling method to obtain three-dimensional space field compression data of each regular mesh at multiple historical time points; a definition module configured to define an input format and an output format; a composition module configured to obtain input data according to the three-dimensional space field compression data of each regular mesh at a preset number of historical time points and the boundary conditions in the input format, and obtain output data according to the three-dimensional space field compression data of each regular mesh at a next historical time point in the output format, the input data and the output data together forming a training set; the next historical time point is a next time point of the last time point in the preset number of historical time points; a training module configured to train a large model pre-training network using the training set to obtain a three-dimensional space field prediction large model; the large model pre-training network combines a 3D convolutional neural network and an autoencoder structure; a prediction module configured to obtain input data according to the space field data of each regular mesh at a preset number of historical time points and actual boundary conditions in the input format, and input the three-dimensional space field prediction large model to output three-dimensional space field compression data of each regular mesh at a future time point; a recovery module configured to map the three-dimensional space field compression data of each regular mesh at the future time point to an irregular mesh after up-sampling to obtain a three-dimensional space field distribution at the future time point.

9. A computer device comprising: A memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that the processor executes the computer program to implement the method for predicting a three-dimensional space field based on a large model and data compression according to any one of claims 1-7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the method for predicting a three-dimensional space field based on a large model and data compression according to any one of claims 1-7.

Images