System for controlling fluid space field in hearth based on large model

Abstract

The invention discloses a system for controlling a fluid space field in a hearth based on a large model, and relates to the field of fluid space field control. The distributed sensor network, the data acquisition device, the distributed storage device and the edge computing device are arranged, the small prediction model and the large correction model are deployed in the edge computing device, the small prediction model is used for prediction, and the large correction model is used for correcting the prediction result, so that the prediction precision and stability are improved, and the prediction efficiency is improved. Based on the fluid space field data measured by the limited sensors in the hearth, prediction of the fluid space field in the hearth is completed, the fluid space field in the hearth is controlled based on the prediction result, and the control precision and stability of the fluid space field in the hearth are guaranteed.

Description

Technical Field

[0001] This application relates to the field of fluid space field control, and particularly to a fluid space field control system in a furnace based on a large model. Background Technique

[0002] The rapid and accurate prediction of the fluid space field is crucial for understanding the fluid change trend and optimizing the production process. However, due to cost and operation complexity, the full deployment of physical sensors in most fluid spaces is limited. Therefore, researchers usually rely on numerical calculation methods to simulate the flow characteristics of the fluid space field. For example, the Navier-Stokes (NS) equation, as a partial differential equation describing fluid dynamics characteristics, although applicable in some simple flow scenarios, there is a contradiction between efficiency and accuracy when solving most complex fluid flow problems. Traditional turbulence modeling methods face many challenges, such as limited accuracy, long calculation time, and huge consumption of computing resources, which limit their effectiveness and universality in practical applications.

[0003] In existing research methods, performing flow field calculation and simulating the spatial physical field based on Ansys Fluent is a common solution. Ansys Fluent is a widely used Computational Fluid Dynamics (CFD) software package with advanced numerical calculation methods and powerful pre- and post-processor functions, capable of performing professional analyses of fluid dynamics, heat transfer processes, combustion processes, and multiphase flow processes. However, since this method relies on dynamically solving spatial field information and uses an iterative calculation form, it results in relatively high consumption of computing resources and faces the limitation of being difficult to generalize to a wider range of application scenarios. Summary of the Invention

[0004] The purpose of this application is to provide a fluid space field control system in a furnace based on a large model to improve the prediction accuracy of the fluid space field in the furnace and reduce the calculation time and resource consumption.

[0005] To achieve the above purpose, this application provides the following solutions.

[0006] This application provides a fluid space field control system in a furnace based on a large model, including: a distributed sensor network, a data acquisition device, a distributed storage device, and an edge computing device;

[0007] The distributed sensor network is deployed inside the furnace;

[0008] The data acquisition device is connected to the distributed sensor network, the distributed storage device is connected to the data acquisition device, and the distributed storage device is used to store the fluid space field data measured by the distributed sensor network;

[0009] The edge computing device is used to predict the fluid space field in the furnace based on the measured fluid space field data.

[0010] A prediction small model and a correction large model are deployed on the edge computing device. The prediction small model is used to preliminarily predict the fluid space field in the furnace based on the measured fluid space field data to obtain preliminary predicted fluid space field data. The correction large model is used to correct the preliminary predicted fluid space field data to obtain predicted fluid space field data, and control the fluid space field in the furnace according to the predicted fluid space field data.

[0011] According to the specific embodiments provided by the present application, the present application has the following technical effects.

[0012] The present application provides a control system for the fluid space field in the furnace based on a large model. The present application sets up a distributed sensor network, a data acquisition device, a distributed storage device, and an edge computing device. A prediction small model and a correction large model are deployed on the edge computing device. The prediction small model is used for prediction, and the correction large model is used to correct the prediction result, thereby improving the accuracy and stability of the prediction. Based on the fluid space field data measured by the limited sensors in the furnace, the prediction of the fluid space field in the furnace is completed, and the control of the fluid space field in the furnace is carried out based on the prediction result, ensuring the accuracy and stability of the control of the fluid space field in the furnace. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for use in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0014] Figure 1 It is a schematic structural diagram of a control system for the fluid space field in the furnace based on a large model provided by an embodiment of the present application.

[0015] Figure 2 It is a schematic structural diagram of an encoded neural network model provided by an embodiment of the present application.

[0016] Figure 3 It is a schematic flow diagram of a prediction method for a fluid space field provided by an embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0017] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present application without creative efforts shall fall within the protection scope of the present application.

[0018] To make the above objects, features, and advantages of the present application more obvious and understandable, the present application will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0019] With the continuous development of artificial intelligence technology, neural networks have been widely applied to spatial field prediction tasks due to their powerful feature extraction and pattern recognition capabilities. In many studies, effective prediction models based on data characteristics have been constructed and achieved good application effects in multiple fields. However, most of the existing neural network models are not fully applicable to the time series prediction tasks of spatial fields. The time series data of spatial fields usually have the characteristics of high dimensionality and large data volume, which makes it easy for conventional neural network models to face difficulties in feature capture and non-convergence during processing. Since the time series data of spatial fields not only contain complex spatio-temporal dependence relationships but may also be affected by factors such as non-linearity and noise, the expression ability and generalization ability of conventional models are limited when dealing with these complex features. Therefore, how to effectively design a model suitable for high-dimensional and time series characteristics has become a major challenge in current research.

[0020] Combining data dimensionality reduction and compression techniques with neural network methods can effectively solve the problems of high dimensionality and large data volume faced in spatio-temporal prediction of spatial fields. Traditional Proper Orthogonal Decomposition (POD) and Reduced-order Models (ROMs) represent high-dimensional and large amounts of data through low-dimensional models and have been widely applied in fields such as fluid mechanics. However, the temporal data of fluid spatial fields usually exhibit complex non-linear characteristics and cannot be effectively reduced in dimension through linear mapping. In addition, the training of ROMs usually requires a large amount of sample data, and the spatio-temporal prediction of fluid spatial fields involves data at two scales of space and time, which makes the training process of traditional ROMs quite difficult. In contrast, as an unsupervised learning method, the autoencoder can effectively compress the input data in dimension by having the number of neurons in the hidden layer less than that in the input layer, thus achieving feature expression and dimensionality reduction. At the same time, the 3D Convolutional Neural Network (3DCNN) performs excellently in spatio-temporal feature extraction and can effectively identify the dependencies in space and time. Combining the advantages of the autoencoder and 3DCNN, this method can solve the prediction problem of high-dimensional and large-data-volume temporal data. Through network hierarchy design, comprehensive learning of multi-modal input data can also be achieved, thereby improving the prediction accuracy and generalization ability of the model.

[0021] To further improve the accuracy of the spatial field prediction model, by analyzing empirical data, regions where data feature learning is insufficient can be identified, and then targeted measures can be taken. Based on the unsupervised learning ability, tokenization ability, emergence ability, and powerful generalization ability of large language models, they can effectively handle the complex multi-dimensional relationships between data and have the advantage of integrating multi-modal information. Through instruction fine-tuning with language descriptions, the learning direction of the model can be flexibly adjusted, thereby improving the generalization ability and accuracy of the prediction. Especially when targeting regions with large differences, large language models can serve as an auxiliary mechanism to effectively supplement the prediction deficiencies of small models in these regions. Large language models can not only enhance the learning ability of the model by analyzing the mutual relationships between regions, but also optimize the data processing methods of the model in different regions through an adaptive adjustment mechanism, thus achieving more accurate temporal prediction. In this process, the emergence characteristics of the language model enable it to automatically generate highly adaptable inference paths when facing complex and dynamically changing fluid data, providing a more accurate and reliable solution for spatial field prediction.

[0022] In an exemplary embodiment, such as Figure 1As shown, a fluid space field control system in a furnace based on a large model is provided, including: a distributed sensor network, a data acquisition device, a distributed storage device, and an edge computing device; the distributed sensor network is deployed inside the furnace; the data acquisition device is connected to the distributed sensor network, the distributed storage device is connected to the data acquisition device, and the distributed storage device is used to store the fluid space field data measured by the distributed sensor network; the edge computing device is used to predict the fluid space field in the furnace based on the measured fluid space field data; a prediction small model and a correction large model are deployed on the edge computing device, the prediction small model is used to make a preliminary prediction of the fluid space field in the furnace based on the measured fluid space field data to obtain preliminary predicted fluid space field data; the correction large model is used to correct the preliminary predicted fluid space field data to obtain predicted fluid space field data, and control the fluid space field in the furnace according to the predicted fluid space field data.

[0023] In another exemplary embodiment, the edge computing device includes a CPU unit and two GPU units; the CPU unit is connected to the distributed storage device, and the CPU unit is also connected to the two GPU units; the CPU unit is used to process the measured fluid space field data, the prediction small model is deployed on one of the GPU units, and the correction large model is deployed on the other GPU unit; the distributed storage device includes a relational database and a non-relational database, the relational database is used to store structured numerical data, and the non-relational database is used to store unstructured or semi-structured graphical data and / or log data.

[0024] In another exemplary embodiment, the fluid space field control system in a furnace based on a large model further includes a power module and a heat dissipation system; the power module is respectively connected to the data acquisition device, the distributed storage device, and the edge computing device, and the power module is used to supply power to the data acquisition device, the distributed storage device, and the edge computing device; the heat dissipation system is used to dissipate heat from the power module and the edge computing device.

[0025] In the above device embodiments, to improve the real-time prediction ability and system stability of the correction large model, a comprehensive hardware architecture is proposed, which involves the collaborative work of sensor and data acquisition hardware, storage system, adaptive computing unit, and edge computing device. First, the system adopts a distributed sensor network to collect key physical quantity data input inside and outside the furnace in real time. The sensors are arranged in various areas of the furnace to ensure comprehensive and efficient collection of key data. The data is transmitted to the central system through a highly reliable communication protocol to ensure the stability of the data transmission process. To process and store large-scale historical data and real-time collected data, the embodiments of this application design an efficient distributed storage device and combine relational and non-relational database systems for data management and optimized access. The relational database (RDBMS) is based on the traditional relational model, and the data is organized in the form of rows and columns, and the tables are associated through foreign keys, which is suitable for storing structured numerical data; while the non-relational database (NoSQL) supports flexible data models and does not require a strict table structure, and is suitable for storing unstructured or semi-structured data such as graphic data and log data.

[0026] In terms of computing and processing, the embodiments of this application are equipped with an adaptive computing unit in the edge computing device, which can dynamically adjust computing resources according to different computing requirements to support real-time inference and prediction tasks of large-scale combustion models.

[0027] For example, a small prediction model can be deployed on the CPU, while a complex correction large model can be accelerated using the GPU. In the embodiments of this application, the data collection of the prediction model comes from sensors. The implementation of the prediction requires the memory to store data. Then the hardware box needs to include a CPU and a GPU. The small model is on the CPU, and the large model is on the GPU. The whole device is called an edge computing device, which can be deployed on-site to predict according to the real-time incoming data.

[0028] For another example, the embodiments of this application can also set up a structure with one CPU unit and two GPU units. One is deployed with a small model for preliminary prediction of the spatial field, and the other is deployed with a correction large model for key areas (the core of the furnace, the area of coal injection and air injection). The edge computing device plays a crucial role in the system. By being deployed on-site, it reduces data transmission latency, realizes real-time data processing and rapid prediction, and optimizes the operation control of the furnace.

[0029] In another exemplary embodiment, the fluid space field control system in the furnace based on the large model further includes a central computing unit; the central computing unit is used to train an encoded neural network model to obtain a prediction small model, and construct a correction large model based on the Qwen large model;

[0030] Among them, in terms of training the encoded neural network model to obtain a prediction sub-model, the central computing unit is specifically used to execute the following steps 101 to 105.

[0031] Step 101: Use Ansys software to simulate the fluid space field of the furnace to obtain the fluid space field data at each time step from the initial simulation moment to the fluid stable state.

[0032] Step 102: Preprocess the fluid space field data at each time step from the initial simulation moment to the fluid stable state to obtain the preprocessed fluid space field data at each time step from the initial simulation moment to the fluid stable state, and construct a fluid space field data sequence.

[0033] Step 103: Divide the fluid space field data sequence by using a sliding window method to construct a training set.

[0034] Step 104: Construct a loss function.

[0035] Step 105: Use the training set and the loss function to train the encoded neural network model to obtain a trained encoded neural network model as a prediction sub-model.

[0036] Through the close cooperation between the edge computing device and the central computing unit in this application, the above device can efficiently complete model inference and parameter adjustment, ensuring the real-time acquisition and processing of internal data and working condition parameters. This hardware architecture design not only effectively supports the efficient operation of the large space prediction model, but also ensures low latency and high precision in data processing, greatly improving the real-time prediction ability and system stability of the combustion large model.

[0037] Taking the configuration of two GPU units as an example, the working process of the furnace internal fluid space field control system based on the large model of the application will be described below.

[0038] First, through the data acquisition device, collect the physical quantity data with coordinate information collected from the simulation software or sensors, and store it in the data storage module. In the embodiment of the present application, an efficient distributed storage device is designed, and a relational and non-relational database system is combined for data management and optimized access. The relational database (RDBMS) is based on the traditional relational model, and the data is organized in the form of rows and columns. Tables are associated through foreign keys and are suitable for storing structured numerical data; while the non-relational database (NoSQL) supports flexible data models and does not require a strict table structure, making it suitable for storing unstructured or semi-structured data such as graphic data and log data. Then, the calculation unit module reads the memory data and processes the transformation from the irregular grid to the regular grid on the CPU unit. The prediction model can be accelerated using the GPU unit. In this embodiment, a structure with two GPU units is used. One is deployed with a small model for the preliminary prediction of the spatial field, and the other is deployed with a modified large model for the key areas (furnace core, air injection and coal injection areas). The adaptive computing unit equipped in the embodiment of the present application can dynamically adjust the computing resources according to different computing requirements to support the real-time inference and prediction tasks of large-scale combustion models. The prediction results obtained from the calculation unit are then sent into the memory for subsequent multi-step prediction or analysis. In addition, through the configuration of the battery module and the power amplifier module in the power supply system, it is responsible for the power supply management of the device, ensuring the stable operation of GPU and CPU calculations and preventing calculation interruptions. The device is configured with a cooling system to relieve the large amount of heat generated by the GPU, CPU, and power supply module, maintaining the stable operation of the hardware and avoiding loss of the device's lifespan.

[0039] The entire large model-based fluid space field control system in the furnace can be deployed on-site and predict according to the real-time incoming data. The main modules are required as shown in the following table:

[0040] Table 1 System parameter requirements

[0041]

[0042]

[0043] In another exemplary embodiment of the present application, a three-dimensional space model of the furnace is constructed based on Ansys software and an irregular grid is generated. Different boundary conditions are set according to different working conditions for dynamic simulation. During multiple time steps from the initial simulation moment to the fluid stable state, fluid parameter data is obtained. These data are recorded by the method of exporting layer by layer along the Z-axis, and the grid information and fluid parameter information are stored together in the CSV file format for convenient subsequent data processing and analysis.

[0044] By calculating the time series mutual information To quantify the temporal dependency between fluid time series data, and Represent the fluid spatial field data at the t0th time step and the t0-k0th time step respectively. This process can reveal the mutual relationship between fluids at different time steps and provide a basis for time series data modeling. Time series mutual information can effectively capture the implicit information in the dynamic changes of fluids, especially when the fluid behavior has nonlinear and complex time-varying characteristics. Time series mutual information As shown below:

[0045]

[0046] in, is the fluid space field data of the t0-k0 time step The entropy of is the fluid space field data at the t0th time step and the fluid space field data at the t0-k0 time step The cross entropy between Represents the fluid space field data at the t0th time step The entropy of , which is the data distribution measure of the fluid at the t0th time step, is as follows:

[0047]

[0048] in, is the fluid space field data at the t0th time step middle The probability distribution of occurrence.

[0049] Rearrange the data according to the calculated temporal dependency. Assuming the temporal dependency is k, the data needs to be arranged in the following format: the input data X of the training sample in the training set i =(D i ,D i+1 ,...,D i+k-1 ), label Y i =(D i+k ), where i is (1, 2, ..., I), and the relationship I + k is the total step length. This data organization method can provide rich and time-dependent training samples for subsequent deep learning models.

[0050] In another exemplary embodiment, since spatial fluid data usually has high dimensionality, large scale, and complexity, in order to improve the training efficiency and accuracy of subsequent prediction models, the data must be effectively compressed. Considering the irregularity of the grids generated by Ansys software, therefore, a point mapping method is adopted to map its irregular grid data and fluid parameter data one by one into regular cubic grids to ensure data consistency and adaptability. Assume the shape of this cubic grid is (Ax, Ay, Az), that is, the size of the cubic grid in three-dimensional space. Through this mapping method, high-dimensional and complex fluid data can be transformed into a unified format suitable for input into deep learning models.

[0051] In the grids generated by Ansys software, the number of grids in the boundary region is usually large. Although this improves the prediction accuracy in the boundary region, it also results in a relatively low grid density in the central region, which may in turn affect the prediction performance of the model in the central region. Therefore, a strategy must be introduced to balance the attention on the boundary and central regions. Specifically, the Euclidean distance from the central region can be defined, the distance of each grid point from the central region can be calculated, and different weights can be assigned to each region during the training process based on this distance. By adding a weight value linearly related to the distance to the loss function of the prediction model, the attention of the model to each region can be effectively adjusted, especially the attention to the central region, thereby optimizing the overall prediction effect. These distance information will also be mapped into the regular cubic grid to ensure data consistency. The Euclidean distance calculation formula is as follows:

[0052]

[0053] where d m is the Euclidean distance between the m-th grid point of the fluid space field and the center point of the fluid space field, α and β are the coefficients of the linear relationship, x m , y m and z m are the x-axis, y-axis, and z-axis coordinates of the m-th grid point of the fluid space field respectively, and x c , y c and z c are the x-axis, y-axis, and z-axis coordinates of the center point of the fluid space field respectively.

[0054] Set the weight value of the linear relationship as w m , and the weight value linearly related to the distance is calculated as follows:

[0055] w m = α·d m + β.

[0056] Among them, α and β are the coefficients of the linear relationship, controlling the degree of change of the control weight with distance. α adjusts the sensitivity of the weight to distance, and β is the offset used to adjust the reference level of the weight.

[0057] In the loss function of the model, by introducing these weights, the importance of different regions during the training process can be adjusted. For example, assume the loss function is L m , then the loss function L with weights w is as follows:

[0058]

[0059] where L m is the loss at the m-th grid point of the fluid spatial field, and w m is the linear weight at the m-th grid point of the fluid spatial field.

[0060] To further improve the computational efficiency and reduce the computational burden of the model, downsampling techniques can be adopted to map high-dimensional data to a lower-dimensional spatial field (B x , B y , B z ). The calculation formula for the sampling interval R is as follows:

[0061] R = A / B.

[0062] where A and B represent the original data and the resolution of the downsampled data respectively.

[0063] Downsampling can not only reduce the consumption of computing resources, but also effectively remove noise, highlight the main features of the data, and improve the model's understanding ability of the fluid spatial field. This process can ensure that the model can still retain key fluid dynamic information in the lower-dimensional spatial field, providing sufficient data support for subsequent time series prediction tasks.

[0064] In another exemplary embodiment, the above prediction sub-model is trained based on an encoded neural network model. This encoded neural network model uses a high-dimensional cube grid as input and performs dynamic prediction of fluid parameters by combining a three-dimensional convolutional neural network (3D Convolutional Neural Network, 3DCNN) and an autoencoder. Specifically, through the combination of multiple 3D convolutional layers, residual connections, and skip connections, the spatio-temporal features of the input spatial data are extracted. The convolutional layers are used to capture local features, and the residual connections and skip connections help to avoid the problem of gradient disappearance in deep networks, thereby improving the stability and performance of the model. This encoded neural network model includes an encoder, a Transformer layer, and a decoder. The encoder maps the input data to a low-dimensional space, the Transformer layer serves as a feature compression layer, and the decoder restores the compressed information.

[0065] To enable the model to handle the prediction of dynamic fluid parameters under different working conditions, it is necessary to consider the influence of working condition parameters on fluid dynamics. Therefore, working condition parameters are introduced into the Transformer layer of the encoding neural network model, and the working condition parameters are fused with the feature vectors extracted by the encoder through a concatenation operation. This operation can enhance the adaptability of the model to different working conditions, enabling the network to learn the variation laws of fluid parameters under each working condition. The Transformer layer combined with working condition parameters not only plays the role of feature compression, but also can effectively incorporate external working condition information into the time series prediction of the spatial field, improving the generalization ability of the model in practical applications.

[0066] As Figure 2 shown, the encoding neural network model of the embodiment of the present application includes: a data input layer, a working condition parameter input layer, an encoder, a Transformer layer, a decoder, and an output layer. Among them, the data input layer, the encoder, the Transformer layer, the decoder, and the output layer are connected in sequence; the encoder and the decoder are also connected in a skip connection manner; the working condition parameter input layer is connected to the decoder.

[0067] The encoder includes a first 3D convolutional layer, a first residual block, a downsampling layer, and a second 3D convolutional layer connected in sequence; the decoder includes a transmission layer, a third 3D convolutional layer, a second residual block, and a deconvolution layer connected in sequence.

[0068] The prediction small model adopted in the present application combines a three-dimensional convolutional neural network (3DCNN) and an autoencoder (AutoEncoder) model with a Transformer structure, that is, an encoding neural network model. The input is spatial field data with a shape of (D1, D1, D1, N) at multiple moments after regular grid processing and working condition parameters with a shape of (N c , 1). The model extracts features through the encoder, changes features through the Transformer layer, and finally restores the prediction result through the decoder.

[0069] Among them, the data input layer inputs spatial field data with a shape of (D1, D1, D1, N), and the working condition parameter input layer inputs data with a shape of (N c , 1). Among them, D1 is the dimension information of the regular grid, and N represents data at the previous N moments.

[0070] For the encoder part, two 3D convolutional layers are used for preliminary feature learning and extraction. The residual block consists of 12 residual units, and each residual unit contains two layers of 3D convolution. The features are added through skip connections to keep the input and output dimensions consistent.

[0071] Downsampling layer: The data dimension is gradually reduced through 3 sampling blocks. Each downsampling block reduces the spatial size through convolution operations, and finally outputs a low-dimensional shape of (D2, D2, D2, N M ).

[0072] The operating condition parameter input layer expands the operating condition parameters with a shape of (N c , 1) to (D2, D2, D2, N M ) through a fully connected layer. It is concatenated with the feature map output by the encoder to affect the output of the decoder.

[0073] The transmission layer transforms the data after concatenating the output of the encoder and the operating condition parameters through the multi-head attention mechanism (Multi-HeadAttention).

[0074] Decoder part: The output of the encoder is upsampled through a transposed convolution layer to gradually restore the spatial resolution of the data. The decoder is restored to the final size of (D1, D1, D1, 1) through the transposed convolution layer, that is, to predict the spatial field data of the next step.

[0075] As Figure 2 shown, Figure 2 in it, the residual connection is represented by the + sign, indicating the addition of the input and the output after passing through the layer operation. The data flow direction is marked with an arrow, and a part of the output of the encoder is connected to the decoder through a skip connection.

[0076] The calculation formulas of the convolution and pooling layers in the 3D convolutional layer in the prediction small model are as follows:

[0077] Y l = Conv3D(X l-1 , W l , b l ).

[0078] Y l pool = MaxPool3D(Y l ).

[0079] Among them, X l-1 is the output of the (l - 1)th layer, W l is the convolution kernel of the lth layer, and b lis the bias term of the l-th layer, Conv3D is a three-dimensional convolution operation, and MaxPool3D is a pooling operation. Through multiple convolutions and poolings stacked, a low-rank representation ZCNN with spatio-temporal features is obtained, which contains the deep features of the spatial fluid field.

[0080] The structures of the encoder, Transformer layer, and decoder are represented as follows:

[0081] Z encoder = f encoder (Z CNN ).

[0082] Z bottleneck = f bottleneck (Z encode ).

[0083]

[0084] Among them, Z encoder is the low-dimensional space mapped by Z CNN , f encoder () is the encoding function of the encoder, Z CNN is the output of the data input layer, Z bottleneck is the compressed representation of the Transformer layer, f bottleneck () is the compression function of the Transformer layer, is the prediction result, and f decoder () is the decoding function of the decoder.

[0085] To further improve the adaptability of the model under different working conditions, the working condition parameter P can be incorporated into the representation of the Transformer layer. Through the concatenation operation, the working condition parameter is fused with the Transformer layer features, and the formula is as follows:

[0086] Z bottleneck,combined = Concat(Z bottleneck , P).

[0087] Among them, Concat is the concatenation function, and Z bottleneck,combined is the fused feature of the working condition parameter and the compressed representation of the Transformer layer.

[0088] The final model output Y is:

[0089] Y = f final (Z bottleneck,combined ).

[0090] Y is the prediction result, and f final () is the decoding and prediction function.

[0091] In the embodiments of the present application, in order to further improve the prediction accuracy of the model, skip connections and residual blocks can also be added during network design. These structures help to better transmit feature information, slow down the attenuation of information in the deep network, and enable the model to better capture important features in long-term time series data. During the model training process, the network will perform backpropagation based on historical data under different working conditions, thereby learning the influence mechanism of different working conditions on the fluid field, and then predicting the fluid state at future moments.

[0092] In an exemplary embodiment, as Figure 3 shown, in terms of predicting the fluid space field in the furnace based on measured fluid space field data, the edge computing device is specifically configured to perform the following steps 201 to 206.

[0093] Step 201, initialize the value of t to 0.

[0094] Step 202, obtain the fluid space field data of the furnace at the current time step and the previous k - 1 time steps before the current time step, and form the input data for the t-th time step; k represents the time series dependence.

[0095] Step 203, input the input data of the t-th time step into the prediction sub-model to obtain the preliminary predicted fluid space field data for the (t + 1)-th time step.

[0096] Step 204, use the correction large model to correct the preliminary predicted fluid space field data for the (t + 1)-th time step to obtain the fluid space field data for the (t + 1)-th time step.

[0097] Step 205, use the fluid space field data for the (t + 1)-th time step to update the input data for the t-th time step to obtain the input data for the (t + 1)-th time step.

[0098] Step 206, increase the value of t by 1, and return to the step of "inputting the input data of the t-th time step into the prediction sub-model to obtain the preliminary predicted fluid space field data for the (t + 1)-th time step" until the fluid in the furnace reaches a stable state.

[0099] Implementing the above steps 201 to 206 can improve the prediction accuracy of the fluid space field and reduce the calculation time and resource consumption.

[0100] In another exemplary embodiment, the method for obtaining the fluid space field data of the furnace at the current time step and the previous k - 1 time steps before the current time step in step 202 is measurement and / or simulation.

[0101] In another exemplary embodiment, after obtaining the spatial field data predicted by the prediction small model, this data will be input into the large language model based on Qwen for further correction. Since the spatial field data is usually numerical data, while the input of the Qwen model is character data, when these numerical values are input into Qwen, the model will break down each numerical value into individual characters for processing. This approach may lead to insufficient learning of the input data by the model, thereby causing instability or divergence in training. Therefore, it is necessary to appropriately convert the input numerical values so that they can be compatible with the input format of the Qwen model.

[0102] To solve this problem, a method of mapping numerical values to special Tokens is adopted. The Qwen model has expandable Token positions, ranging from <|extra_0|> to <|extra_204|>, which can be used to represent different numerical values or numerical ranges. Since the number of Tokens in Qwen is limited, directly mapping all numerical values may result in too many Tokens, increasing the complexity of training. Therefore, it is necessary to divide the numerical values into multiple intervals and map each interval to the corresponding Token. Specifically, by designing a suitable interval division method, the continuous numerical space can be mapped into these special Tokens, thereby reducing the number of Tokens fed into the large model while retaining the effective information of the numerical values. Let the input spatial field data be S = {s1, s2, …, s J}, where represents the j-th data point (such as temperature, pressure, etc.). Assume that the numerical range [s min , s max of the fluid spatial field data is divided into N intervals, and each interval corresponds to a Token T n , and the mapping function MAP() is:

[0103]

[0104] where, (s max - s min ) / N is the size of each interval, represents the rounding operation.

[0105] This mapping method not only solves the problem of format mismatch between numerical input and Qwen character input, but also effectively reduces the number of Tokens and improves the training efficiency of the large model. By optimizing the mapping strategy and the number of Tokens, and further fine-tuning and correcting the network in combination with the training set data, the accuracy of the prediction results has been significantly improved. In this process, the Qwen model can fully learn the complex relationship between numerical values and the spatial field, improving its generalization ability and accuracy for spatial field prediction tasks.

[0106] In the above step 206, after the spatial field data obtains the preliminary prediction result through the prediction small model, these data will be input into the correction large model for further correction to generate the spatial field prediction of the next moment. Subsequently, the obtained new spatial field data will form new input together with the previous historical data. Through the sliding window method, the overall window slides backward and serves as the new input to continue the prediction of the next moment, and is corrected again through the correction large model. This iterative prediction mechanism enables the model to continuously perform efficient prediction and correction in the continuously updated time series, thereby continuously refining the spatial field data of future moments.

[0107] Suppose the initial spatial field prediction data Y is obtained from the small model t , and it is sent into the correction large model to obtain the corrected result The calculation formula is as follows:

[0108]

[0109] Among them, is the large model mapping function, and θ is the parameter of the large model.

[0110] After each prediction, the newly predicted result is added to the history to form new input data. The updated input data X t+1 contains the latest prediction data and the previous historical data:

[0111]

[0112] The iterative process can continue until the set required step size is reached.

[0113] Compared with the traditional numerical simulation method, the iterative prediction mechanism has significant advantages in terms of computational efficiency and resource consumption. The traditional numerical simulation method usually needs to gradually solve the entire fluid spatial field, involving complex equations and a large amount of computational work. Especially when facing high-dimensional spatio-temporal data, the demand for computational time and resources is extremely large. By using the iterative prediction mechanism, through gradually updating the historical data and prediction results, it can effectively reduce the dependence on global calculations and greatly save computational resources and time overhead.

[0114] In addition, the iterative prediction mechanism can correct and optimize the prediction result at each time step, thereby improving the accuracy and stability of the prediction. Each iteration not only uses the previous historical data but also takes the prediction result of the previous moment as the input, forming a closed-loop update process. This method enables the prediction model to flexibly adapt to dynamic working conditions, gradually approach the true fluid dynamic characteristics, and accumulate experience and adjust strategies in each iteration, improving the accuracy and reliability of the prediction result.

[0115] In another exemplary embodiment, a specific application is provided to illustrate the technical effects of the above control system as follows.

[0116] Taking the prediction of the spatial temperature field of the dynamic combustion in the furnace as an implementation case, the effectiveness of the proposed large model for three-dimensional spatial field prediction is tested. In the embodiment of the present application, in the application of the furnace, the fluid spatial field can be a spatial temperature field or an air velocity field. In the embodiment of the present application, the spatial temperature field is taken as an example for illustration. The control system of the present application can be applied not only to the control of the fluid spatial field in the furnace, but also to the control of the fluid spatial field in other spaces, which will not be elaborated here.

[0117] Based on the boiler model constructed by Ansys software, its length, width, and height are 24 meters, 55.8 meters, and 16 meters respectively, and the number of grids reaches 1,885,691. The boundary conditions of the model cover the primary air velocity, secondary air velocity, burnout air velocity, coal feeding amount and its rate. By changing different boundary conditions for simulation, from the start of powder injection to the stable process, the boiler temperature is adjusted to 1000K to ignite the coal material, and the temperature field data at each step from ignition to the stable combustion state are exported, and finally a data set of 70 working conditions is generated. In order to further analyze the correlation of the temperature field time series data, the time series mutual information of these data is calculated. Through this process, the influence law of the temperature field data at the previous 5 moments on the data at the next moment can be identified, and these data are sorted into a training set, which contains the time series field data at 5 moments as input and the temperature field data at the next moment as labels, providing a basis for the subsequent training of the prediction model.

[0118] Due to the overall structural characteristics of the grid data, its fluid spatial field is mapped into a cube data of 48×48×48 one by one. To reduce the data dimension and improve the calculation efficiency, a downsampling method is used to further map it into a cube data of 36×36×36. According to the specific size of the boiler, the position of the center line is first determined. Then, the Euclidean distance from all grid points to the center line is calculated, and the weight is calculated based on this distance and mapped into the cube data of 48×48×48. Finally, the cube data of 48×48×48 with weight information is further compressed into cube data of 36×36×36 by the downsampling method for subsequent model input and training.

[0119] 80% of all the simulation data sets are used as the training set, 14% as the test set, and 6% as the validation set. The large model for spatial field prediction is trained and predicted through 80% of the training set. The whole process is divided into the following steps:

[0120] Prediction small model training and prediction: The input data is the regular cube data of the historical 5 moments, with a shape of 36×36×36×5. At the same time, the operating condition parameter data with a shape of 5×1 needs to be input in the Transformer layer. By concatenating these data in the Transformer layer, the data is finally passed into a 3D convolutional neural network (3DCNN) and an autoencoder for processing, and the spatial temperature field data of the future moment with a shape of 36×36×36×1 is predicted and output. During the training process, the above-mentioned saved Euclidean distance weight function is used to allocate spatial points, which is used as a weighted term of the loss function. This weight function can effectively balance the training focus of the boundary region and the central region and avoid overfitting in the boundary region.

[0121] Since the special Token expansion position of the Qwen large model is limited to 205, that is, from <|extra_0|> to <|extra_204|>, the temperature field range from 0 to 2000K is allocated a special Token position every 10K, and the temperature data is mapped to the corresponding Token representation. These Tokens will be used as identifiers of numerical data. By mapping the output data of each prediction small model to the corresponding real temperature field data, the internal network of the large model is fine-tuned and corrected, so that it can gradually correct the prediction results of the large model and make them closer to the real data. After obtaining the fine-tuned and corrected large model, the temperature field data generated by each prediction small model can be further corrected. By using the large model as an adjustment network, the output results can be manually adjusted in real time using prompt engineering, and the adjusted results are saved to the historical data, so as to provide more accurate and demand-compliant prediction results for users.

[0122] By applying the iterative prediction mechanism, the next-state spatial temperature field data generated by the prediction small model and the corrected large model is adjusted as the training set. As the sliding window advances, continue to train and predict the next state. During the training process, the layer-by-layer mapping and upsampling technology are adopted to restore the prediction result from the 36×36×36 cube data to the 48×48×48 grid data to conform to the grid structure and provide a higher spatial resolution.

[0123] In this experiment, the dynamic temperature field changes for 5 steps were continuously predicted for a certain operating condition. Through the results of the multi-step predicted temperature field, the relative error of the overall spatial field temperature was obtained. The relative error of the highest value was about 4%, the relative error of the lowest value was about 17.83%, and the relative error of the average temperature of all grid points was about 2%. Among them, the errors of 78% of the prediction points fell within the error band of 0-25%, which could ensure a high prediction accuracy while effectively deducing the combustion changes. The errors were mainly concentrated in the regions with large temperature gradient changes (such as the nozzle region), and the errors in other regions were controlled within 25%.

[0124] Compared with Ansys simulation, the large model based on this spatial field prediction has improved the calculation speed by 98.3%. Specifically, Ansys simulation takes about 1 second for each step of calculation, while using the large model prediction only takes 17 milliseconds, greatly improving the prediction efficiency and real-time performance.

[0125] To accurately monitor the operating state and combustion process of the boiler furnace, the embodiment of this application adopts a distributed sensor network to collect key physical quantity data input from inside and outside the boiler in real time. These sensors cover all areas of the boiler furnace, collecting multi-dimensional data such as furnace temperature, flue gas velocity, pressure, etc., and also including key parameters such as the wind speeds of primary air, secondary air, overfire air, and coal feeding amount. Each sensor module is configured individually according to its physical location and measurement requirements, and transmits the collected data to the central processing unit through a stable communication protocol. For the large-scale data generated during the combustion process of the boiler furnace, this system designs an efficient distributed storage architecture to ensure the efficient storage of historical data and real-time collected data. The system combines a relational database and a non-relational database. The former is used to store structured numerical data (such as temperature, pressure, etc.) generated during the operation of the boiler, and the latter is used to store unstructured data such as sensor data, combustion process logs, and boiler combustion images.

[0126] To support the real-time inference and prediction of large-scale models, this system is equipped with an adaptive computing unit. The adaptive computing unit dynamically adjusts computing resources according to the computational complexity and real-time requirements of tasks, ensuring that the system efficiently processes different types of computing tasks. For example, small prediction models (such as the preliminary prediction of the combustion temperature field) can be deployed to run on low-power CPUs, while the corrected large model involving complex operations (such as correcting grids with large errors) is accelerated by GPUs. In addition, edge computing devices play a crucial role in this system. By deploying edge computing nodes at the boiler site, edge computing devices can reduce data transmission latency, ensuring the immediate processing and real-time feedback of data. Edge computing devices not only perform local data preprocessing, but also can execute tasks such as model inference and adjustment of control parameters, quickly responding to changes during the operation of the boiler. Through the collaborative work of edge computing and the central computing unit, the system achieves low latency and high precision in data collection and processing.

[0127] According to the specific embodiments provided by this application, this application has the following technical effects.

[0128] 1. The prediction small model and the correction large model on which the in-furnace fluid space field control system of this application is based can significantly reduce the calculation time while ensuring high precision. Through the effective compression of grid data, the joint prediction of the prediction small model and the correction large model, and the design of the iterative prediction mechanism, this model comprehensively captures the complex relationship between the operating condition parameters and the historical space field data, thus ensuring the accuracy of the prediction results. In addition, this model can complete the reproduction of space field data once within about 10 milliseconds. Compared with the traditional numerical calculation method, the calculation time is greatly shortened, significantly improving the prediction efficiency and real-time performance.

[0129] 2. The method of introducing a large language model into the prediction model proposed in this application makes full use of the emerging capabilities of the large language model and realizes an efficient training and inference process. This method can timely correct the areas with insufficient accuracy in the previous prediction small model and further optimize the model performance through the guidance of prompt words during the training process. In this way, the model can not only adaptively improve the accuracy but also effectively handle complex multi-modal data inputs, enhancing the robustness and generalization ability of the overall prediction system.

[0130] 3. The prediction small model and the correction large model proposed in this application have high scalability and strong generalization ability. After the encoder part of the autoencoder learns the spatio-temporal features of the historical data of the space field, this model can be applied to other similar models. By adjusting the boundary conditions, the Transformer layer of the prediction small model and the decoder part of the autoencoder can be retrained, and the pre-trained large language model can be fine-tuned without modifying the global parameters. This flexible method enables the model to adapt to different operating conditions and different types of space field learning tasks, thus improving its applicability and accuracy in various application scenarios.

[0131] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, 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, it should be considered as the scope recorded in this specification.

[0132] Specific examples are used in this article to elaborate on the principle and implementation method of this application. The description of the above embodiments is only used to help understand the method and its core idea of this application; at the same time, for those of ordinary skill in the art, based on the idea of this application, there will be changes in the specific implementation method and application scope. To sum up, the content of this specification should not be construed as a limitation to this application.

Claims

1. A fluid space field control system in a furnace based on a large model, characterized in that, Including: A distributed sensor network, a data acquisition device, a distributed storage device, and an edge computing device; The distributed sensor network is deployed inside the furnace; The data acquisition device is connected to the distributed sensor network, the distributed storage device is connected to the data acquisition device, and the distributed storage device is used to store the fluid space field data measured by the distributed sensor network; The edge computing device is used to predict the fluid space field inside the furnace based on the measured fluid space field data; A prediction small model and a correction large model are deployed on the edge computing device. The prediction small model is used to perform a preliminary prediction on the fluid space field inside the furnace based on the measured fluid space field data to obtain preliminary predicted fluid space field data; the correction large model is used to correct the preliminary predicted fluid space field data to obtain predicted fluid space field data, and control the fluid space field inside the furnace according to the predicted fluid space field data.

2. The fluid space field control system in the furnace based on the large model according to claim 1, characterized in that The edge computing device includes a CPU unit and two GPU units; the CPU unit is connected to the distributed storage device, and the CPU unit is also connected to the two GPU units; The CPU unit is used to process the measured fluid space field data. The prediction small model is deployed on one of the GPU units, and the correction large model is deployed on the other GPU unit; The distributed storage device includes a relational database and a non-relational database. The relational database is used to store structured numerical data, and the non-relational database is used to store unstructured or semi-structured graphical data and / or log data.

3. The control system for the fluid space field in the furnace based on the large model according to claim 1, characterized in that, The fluid space field control system inside the furnace based on the large model further includes a power supply module and a heat dissipation system; The power supply module is respectively connected to the data acquisition device, the distributed storage device, and the edge computing device, and the power supply module is used to supply power to the data acquisition device, the distributed storage device, and the edge computing device; The heat dissipation system is used to dissipate heat from the power supply module and the edge computing device.

4. The fluid space field control system in the furnace based on the large model according to claim 1, wherein The fluid space field control system inside the furnace based on the large model further includes a central computing unit; The central computing unit is used to train the encoded neural network model to obtain the prediction small model, and construct the correction large model based on the Qwen large model; Among them, in terms of training the encoded neural network model to obtain the prediction small model, the central computing unit specifically is used for: Using Ansys software to simulate the fluid space field of the furnace to obtain the fluid space field data at each time step from the simulation initial moment to the fluid stable state; Preprocessing the fluid space field data at each time step from the simulation initial moment to the fluid stable state to obtain the preprocessed fluid space field data at each time step from the simulation initial moment to the fluid stable state, and constructing a fluid space field data sequence; Dividing the fluid space field data sequence in a sliding window manner to construct a training set; Constructing a loss function; Train the encoding neural network model using the training set and the loss function to obtain a trained encoding neural network model, which is used as a prediction small model.

5. The furnace internal fluid space field control system based on a large model according to claim 4, characterized in that The training samples in the training set are: X i =(D i , D i+1 ,..., D i+k-1 ), and the label of the training sample is Y i =(D i+k ); X i is the i-th training sample; Y i is the label of the i-th training sample, and D i , D i+1 and D i+k-1 are the preprocessed fluid spatial field data at the i-th, (i + 1)-th, and (i + k - 1)-th time steps respectively, and D i+k is the preprocessed fluid spatial field data at the (i + k)-th time step; The loss function is: w m = α·d m + β; Among them, L w is the loss function, and L m is the loss at the m-th grid point of the fluid space field, and w m is the linear weight at the m-th grid point of the fluid space field; d m is the Euclidean distance between the m-th grid point of the fluid space field and the center point of the fluid space field, α and β are the coefficients of the linear relationship, x m , y m and z m are the x-axis, y-axis and z-axis coordinates of the m-th grid point of the fluid space field respectively, x c , y c and z c are the x-axis, y-axis and z-axis coordinates of the center point of the fluid space field respectively.

6. The fluid space field control system in the furnace based on the large model according to claim 4, characterized in that, Preprocess the fluid space field data for each time step from the simulation initial moment to the fluid stable state to obtain the preprocessed fluid space field data for each time step from the simulation initial moment to the fluid stable state. Specifically, it includes: Use the point mapping method to map the target fluid space field data of the irregular grid into the cube grid to obtain the mapped target fluid space field data; the target fluid space field data is the fluid space field data at any time step from the simulation initial moment to the fluid stable state. Use the downsampling method to downsample the mapped target fluid space field data to obtain the preprocessed target fluid space field data.

7. The fluid space field control system in the furnace based on the large model according to claim 4, wherein The encoding neural network model includes: a data input layer, a working condition parameter input layer, an encoder, a Transformer layer, a decoder, and an output layer; The data input layer, the encoder, the Transformer layer, the decoder, and the output layer are connected in sequence; The encoder and the decoder are also connected by a skip connection method; The working condition parameter input layer is connected to the decoder; The formula of the encoder is: Z encoder = f encoder (Z CNN ); The formula of the Transformer layer is: Z bottleneck = f bottleneck (Z encode ); The formula of the decoder is: Y = f final (Z bottleneck,combined ); Z bottleneck,combined = Concat(Z bottleneck , P); Among them, Z encoder is the low-dimensional space mapped by Z CNN , f encoder () is the encoding function of the encoder, Z CNN is the output of the data input layer, Z bottleneck is the compressed representation of the Transformer layer, f bottleneck () is the compression function of the Transformer layer, Y is the prediction result, f final () is the decoding and prediction function, Z bottleneck,combined is the fusion feature of the working condition parameters and the compressed representation of the Transformer layer, Concat() is the concatenation function, and P is the working condition parameter.

8. The fluid space field control system in the furnace based on the large model according to claim 7, characterized in that, The encoder includes a first 3D convolutional layer, a first residual block, a downsampling layer, and a second 3D convolutional layer connected in sequence; The decoder includes a transmission layer, a third 3D convolutional layer, a second residual block, and a deconvolution layer connected in sequence.

9. The fluid space field control system in the furnace based on the large model according to claim 1, characterized in that, In terms of predicting the fluid space field in the furnace based on the measured fluid space field data, the edge computing device is specifically used for: Initialize the value of t to 0; Obtain the fluid space field data of the furnace at the current time step and the previous k - 1 time steps before the current time step to form the input data for the t-th time step; k represents the time series dependence degree; Input the input data for the t-th time step into the prediction small model to obtain the preliminary predicted fluid space field data for the (t + 1)-th time step; Use the correction large model to correct the preliminary predicted fluid space field data for the (t + 1)-th time step to obtain the fluid space field data for the (t + 1)-th time step; Use the fluid space field data for the (t + 1)-th time step to update the input data for the t-th time step to obtain the input data for the (t + 1)-th time step; Increase the value of t by 1, and return to the step of "input the input data for the t-th time step into the prediction small model to obtain the preliminary predicted fluid space field data for the (t + 1)-th time step" until the fluid in the furnace reaches a stable state.

10. The fluid space field control system in the furnace based on the large model according to claim 9, characterized in that, The time series dependence degree is obtained in the following way: Let the value of k0 be 1; Calculate the time series mutual information amount between the fluid space field data at the t0-th time step and the fluid space field data at the (t0 - k0)-th time step. Judgment formula to determine whether it holds; where I0 is the threshold of the temporal mutual information If it holds, increment the value of k0 by 1 and return to the step of "calculating the temporal mutual information between the fluid spatial field data at the t0-th time step and the fluid spatial field data at the (t0 - k0)-th time step"; If it does not hold, output the value of k0 as the temporal dependence degree; where the calculation formula for the temporal mutual information is: In the formula, is the fluid spatial field data at the t0-th time step and the fluid spatial field data at the (t0 - k0)-th time step The temporal mutual information between them, is the entropy of the fluid spatial field data at the t0-th time step of, is the fluid spatial field data at the (t0 - k0)-th time step of the entropy, is the fluid spatial field data at the t0-th time step and the fluid spatial field data at the (t0 - k0)-th time step The cross entropy between them.

Images