A large model-based furnace internal fluid space field control system

Abstract

The application discloses a large model-based furnace inner fluid space field control system and relates to the field of fluid space field control. The application is provided with 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 in the edge computing device, prediction is performed by using the prediction small model, and the prediction result is corrected by using the correction large model, so that the accuracy and stability of prediction are improved, the prediction of the furnace inner fluid space field is completed based on the fluid space field data measured by the limited sensors in the furnace, and the control of the furnace inner fluid space field is performed based on the prediction result, so that the accuracy and stability of the furnace inner fluid space field control are ensured.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of fluid space field control, in particular to a furnace internal fluid space field control system based on a large model. BACKGROUND

[0002] Fast and accurate prediction of fluid space field is crucial for understanding fluid trend and optimizing production process. However, due to the cost and operational complexity, physical sensors are not fully deployed in most fluid spaces, so researchers usually rely on numerical calculation methods to simulate the flow characteristics of fluid space field. For example, Navier-Stokes (NS) equation as a partial differential equation describing the hydrodynamic characteristics, although applicable in some simple flow scenarios, but in solving most complex fluid flow problems, there is a contradiction between efficiency and accuracy. 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, Ansys Fluent is a common scheme for flow field calculation and simulation of space physical field. Ansys Fluent is a widely used Computational Fluid Dynamics (CFD) software package with advanced numerical calculation methods and powerful pre- and post-processor functions, which can perform professional analysis of fluid dynamics, heat transfer process, combustion process and multiphase flow process. However, since this method relies on dynamic solution of space field information and uses iterative calculation form, it leads to high consumption of computing resources, and faces the limitation of being difficult to generalize to more extensive application scenarios. SUMMARY

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

[0005] To achieve the above-mentioned purpose, the present application provides the following solutions.

[0006] The present application provides a furnace internal fluid space field control system based on a large model, comprising: 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 with the distributed sensor network, and the distributed storage device is connected with the data acquisition device, and the distributed storage device is used for storing the fluid space field data measured by the distributed sensor network.

[0009] The edge computing device is configured to predict the fluid space field in the hearth based on the measured fluid space field data;

[0010] The edge computing device is deployed with a prediction small model and a correction large model, the prediction small model is configured to preliminarily predict the fluid space field in the hearth based on the measured fluid space field data, and obtain preliminary prediction fluid space field data; the correction large model is configured to correct the preliminary prediction fluid space field data, and obtain prediction fluid space field data, and control the fluid space field in the hearth according to the prediction fluid space field data.

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

[0012] The application provides a hearth fluid space field control system based on a large model, the application is provided with 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 for correcting the prediction result, so as to improve the accuracy and stability of the prediction, based on the fluid space field data measured by the limited sensors in the hearth, the prediction of the fluid space field in the hearth is completed, and the control of the fluid space field in the hearth is based on the prediction result, so as to ensure the accuracy and stability of the control of the fluid space field in the hearth. BRIEF DESCRIPTION OF DRAWINGS

[0013] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the following will briefly introduce the drawings needed in the embodiments. Obviously, the drawings in the following description are only some embodiments of the application, and for those skilled in the art, other drawings can also be obtained without creative labor based on these drawings.

[0014] Figure 1 A structural schematic diagram of a hearth fluid space field control system based on a large model provided by an embodiment of the application.

[0015] Figure 2 A structural schematic diagram of an encoding neural network model provided by an embodiment of the application.

[0016] Figure 3 A flowchart of a fluid space field prediction method provided by an embodiment of the application. DETAILED DESCRIPTION

[0017] With reference to the drawings and embodiments described below, the above and other objects, features and advantages of the present application will become more apparent.

[0018] The above and other objects, features and advantages of the present application will become more apparent from the following detailed description when taken in conjunction with the accompanying drawings in which:

[0019] With the continuous development of artificial intelligence technology, neural networks have been widely used in 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 have achieved good application results in multiple fields. However, most existing neural network models are not completely suitable for spatial field time series prediction tasks. Spatial field time series data usually has the characteristics of high dimension and large data volume, which makes it difficult for conventional neural network models to capture features and train to converge. Since the spatial field time series data not only contains complex spatio-temporal dependence relationships, but also may be affected by non-linear and noise factors, the expression and generalization capabilities of conventional models are limited when dealing with these complex features. Therefore, how to effectively design a model that adapts to high-dimensional and time-series characteristics has become a major challenge in current research.

[0020] Combining data dimensionality reduction compression techniques with neural network methods can effectively solve the high dimensionality and large data volume problems faced in spatial field time series prediction. Traditional Proper Orthogonal Decomposition (POD) and Reduced-order Models (ROMs) represent high-dimensional and large data through low-dimensional models and have been widely used in fluid mechanics and other fields. However, the time series data of fluid spatial fields often exhibit complex nonlinear characteristics, which cannot be effectively reduced by linear mapping. In addition, the training of ROMs usually requires a large amount of sample data, and the time series prediction of fluid spatial fields involves data in both spatial and temporal scales, making the traditional ROMs training process difficult. In contrast, as an unsupervised learning method, the Autoencoder can effectively compress the dimensionality of the input data by having fewer hidden layer neurons than input layer neurons, thereby achieving feature expression and dimensionality reduction. At the same time, the 3D Convolutional Neural Network (3DCNN) performs well in spatio-temporal feature extraction, effectively identifying spatial and temporal dependencies. By combining the advantages of Autoencoder and 3DCNN, this method can solve the problem of high-dimensional and large-data time series data prediction. Through network level design, it can also achieve comprehensive learning of multi-modal input data, thereby improving the prediction accuracy and generalization ability of the model.

[0021] To further improve the accuracy of spatial field prediction models, we can analyze the empirical data to identify areas where data feature learning is insufficient, and then take targeted measures. Based on the unsupervised learning ability, Tokenization ability, emergent ability, and strong generalization ability of large language models, they can effectively handle complex multi-dimensional relationships between data and have the advantage of integrating multi-modal information. Through language description for instruction fine-tuning, the learning direction of the model can be flexibly adjusted, thereby improving the generalization ability and accuracy of prediction. Especially in areas with large differences, large language models can serve as an auxiliary mechanism to effectively supplement the prediction deficiencies of small models in these areas. Large language models not only enhance the learning ability of the model by analyzing the relationships between regions, but also optimize the data processing method of the model in different regions through an adaptive adjustment mechanism, thereby achieving more accurate time series prediction. In this process, the emergent properties of language models enable them to automatically generate adaptive reasoning paths when faced with complex and dynamically changing fluid data, providing more accurate and reliable solutions for spatial field prediction.

[0022] In one exemplary embodiment, as Figure 1As shown, a large model-based furnace internal fluid space field control system is provided, comprising: 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 with the distributed sensor network, the distributed storage device is connected with the data acquisition device, and the distributed storage device is used for storing the fluid space field data measured by the distributed sensor network; the edge computing device is used for predicting 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 for preliminary prediction of the fluid space field inside the furnace based on the measured fluid space field data, to obtain preliminary prediction fluid space field data; the correction large model is used for correcting the preliminary prediction fluid space field data to obtain prediction fluid space field data, and controlling the fluid space field inside the furnace according to the prediction 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 with the distributed storage device, and the CPU unit is also connected with the two GPU units; the CPU unit is used for processing 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 for storing structured numerical data, and the non-relational database is used for storing unstructured or semi-structured graph data and / or log data.

[0024] In another exemplary embodiment, the large model-based furnace internal fluid space field control system further comprises a power module and a heat dissipation system; the power module is connected with the data acquisition device, the distributed storage device and the edge computing device respectively, and the power module is used for supplying power to the data acquisition device, the distributed storage device and the edge computing device; the heat dissipation system is used for dissipating heat of the power module and the edge computing device.

[0025] In the above device embodiment, in order to improve the real-time prediction ability and system stability of the modified large model, a comprehensive hardware architecture is proposed, which involves the cooperative work of sensors and data acquisition hardware, storage systems, adaptive computing units and edge computing devices. First, the system uses a distributed sensor network to collect key physical quantity data in real time inside and outside the furnace. 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 high-reliability communication protocol to ensure the stability of the data transmission process. In order to process and store large amounts of historical data and real-time collected data, the application embodiment designs an efficient distributed storage device, and combines a relational and non-relational database system 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 strict table structure, which is suitable for storing unstructured or semi-structured data such as graphic data and log data.

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

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

[0028] For another example, the application embodiment can also be configured with a structure of one CPU unit and two GPU units, one block deploying a small model for preliminary prediction of the spatial field, and the other block deploying a modified large model for the key area (furnace core, coal injection area). The edge computing device plays a crucial role in the system, reduces data transmission delay, realizes real-time data processing and fast 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 comprises a central computing unit; the central computing unit is used for training the coded neural network model to obtain a prediction small model, and constructing a modified large model based on the Qwen large model;

[0030] In the aspect of training the encoding neural network model to obtain the prediction small model, the central computing unit is specifically configured to perform the following steps 101-105.

[0031] In step 101, the fluid space field of the furnace is simulated by using the Ansys software to obtain the fluid space field data of each time step from the initial simulation time to the fluid stable state.

[0032] In step 102, the fluid space field data of each time step from the initial simulation time to the fluid stable state is preprocessed to obtain the preprocessed fluid space field data of each time step from the initial simulation time to the fluid stable state, and a fluid space field data sequence is constructed.

[0033] In step 103, the fluid space field data sequence is divided in a sliding window manner to construct a training set.

[0034] In step 104, a loss function is constructed.

[0035] In step 105, the training set and the loss function are used to train the encoding neural network model to obtain a trained encoding neural network model as a prediction small model.

[0036] The above device can efficiently complete model reasoning and parameter adjustment through the close cooperation of the edge computing device and the central computing unit, and ensure the real-time collection and processing of internal data and working condition parameters. This hardware architecture design not only effectively supports the efficient operation of the space prediction large model, but also guarantees low delay and high precision of data processing, greatly improving the real-time prediction ability and system stability of the combustion large model.

[0037] The working process of the furnace fluid space field control system based on the large model of the application will be described below taking two GPU unit configurations as an example.

[0038] Firstly, the data acquisition device collects physical quantity data with coordinate information collected from simulation software or sensors and stores it in the data storage module. The embodiment of the application designs an efficient distributed storage device, and combines a relational and non-relational database system 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. The non-relational database (NoSQL) supports flexible data models and does not require strict table structure, which is suitable for storing unstructured or semi-structured data such as graphic data and log data. Then the memory data is read by the calculation unit module, and the irregular grid to regular grid change processing is performed on the CPU unit, and the prediction model can be accelerated by the GPU unit. The embodiment uses the structure of two GPU units, one of which deploys a small model to preliminarily predict the spatial field, and the other deploys a large correction model for the key area (furnace core, coal injection area). The adaptive calculation unit provided in the embodiment can dynamically adjust the calculation resources according to different calculation requirements to support real-time reasoning and prediction tasks of large-scale combustion models. The prediction results obtained from the calculation unit are sent to 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, the device can be responsible for power management to ensure the stable operation of the GPU and CPU calculation and prevent calculation interruption. The device is equipped with a cooling system to relieve the large amount of heat generated by the GPU, CPU and power module, maintain stable hardware operation, and avoid equipment wear and tear.

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

[0040] Table 1 System parameter requirements

[0041]

[0042]

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

[0044] Through calculation of time series mutual information To quantify the temporal dependencies between fluid time series data, where and These represent the fluid spatial field data at time step t0 and time step t0-k0, respectively. This process can reveal the interrelationships of the fluid at different time steps, providing a basis for time series data modeling. Time series mutual information can effectively capture the implicit information in fluid dynamic changes, especially when fluid behavior exhibits nonlinear and complex time-varying characteristics. Time series mutual information. As shown below:

[0045]

[0046] in, Fluid spatial field data for time steps t0-k0 entropy, Fluid spatial field data at time step t0 Fluid spatial field data at time steps t0-k0 Cross-entropy between This represents the fluid spatial field data at time step t0. The entropy, which is a measure of the data distribution of the fluid at time step t0, is expressed by the following formula:

[0047]

[0048] in, This is the fluid spatial field data at time step t0. middle The probability distribution of occurrence.

[0049] Based on the calculated temporal dependency, the data is reorganized. Assuming the temporal dependency is k, the data needs to be organized in the following format: Input data X of the training samples in the training set. i =(D i D i+1 ,...,D i+k-1 ), label Y i =(D i+k ), where i takes the value (1, 2, ..., I), and the relation I+k is the total step size. This data processing method can provide rich and time-dependent training samples for subsequent deep learning models.

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

[0051] In the grid generated by the Ansys software, the number of grids in the boundary region is usually large, which although improves the prediction accuracy of the boundary region, but also leads to a low grid density in the center region, which may affect the prediction performance of the model in the center region. Therefore, a strategy must be introduced to balance the attention of the boundary and center regions. Specifically, the Euclidean distance from the center region can be defined, the distance of each grid point from the center region is calculated, and different weights are given to each region in the training process according to this distance. By adding a weight value with a linear relationship with the distance in the loss function of the prediction model, the attention of the model to each region, especially the center region, can be effectively adjusted, thereby optimizing the overall prediction effect. These distance information will also be mapped to the regular cubic grid to ensure the consistency of the data. The Euclidean distance calculation formula is as follows:

[0052]

[0053] where d m is the Euclidean distance between the mth grid point of the fluid space field and the center point of the fluid space field, and a and b 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 mth 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] The weight value of the linear relationship is set as w m , and the weight value with a linear relationship with the distance is calculated as follows:

[0055] w m = a·d m + b.

[0056] where a and b are coefficients of linear relationship, controlling the degree of weight value change with distance. a adjusts the sensitivity of weight value to distance, and b is the offset for adjusting the baseline level of weight value.

[0057] By introducing these weights, the importance of different regions in the training process can be adjusted in the loss function of the model. For example, assuming the loss function is L m , then the weighted loss function L w is as follows:

[0058]

[0059] where L m is the loss of the mth grid point of the fluid space field, and w m is the linear weight of the mth grid point of the fluid space field.

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

[0061] R = A / B.

[0062] where A and B represent the original data resolution and the resolution after downsampling, respectively.

[0063] Downsampling not only reduces the consumption of computing resources, but also effectively removes noise, highlights the main features of the data, and improves the model's understanding of the fluid space field. This process ensures that the model can still retain key fluid dynamic information in a lower-dimensional space field, providing sufficient data support for subsequent time series prediction tasks.

[0064] In another exemplary embodiment, the above-mentioned small prediction model is trained based on an encoding neural network model that uses a high-dimensional cubic 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 encoder (Autoencoder). Specifically, by combining multiple 3D convolutional layers, residual connections, and skip connections, the spatio-temporal features of the input spatial data are extracted. Convolutional layers are used to capture local features, while residual connections and skip connections help to avoid the gradient vanishing problem in deep networks, thereby improving the stability and performance of the model. The encoding 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] In order to enable the model to process dynamic fluid parameter prediction under different working conditions, the influence of working condition parameters on fluid dynamics needs to be considered. Therefore, the 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 splicing operation. This operation can enhance the adaptability of the model to different working conditions, so that the network can learn the variation law of fluid parameters under different working conditions. The Transformer layer combined with the working condition parameters not only plays a role in feature compression, but also effectively integrates external working condition information into the time series prediction of the spatial field, improving the generalization ability of the model in actual application.

[0066] As shown in Figure 2 , the encoding neural network model of the embodiment of the present application comprises 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 in a skip connection manner. The working condition parameter input layer is connected with the decoder.

[0067] The encoder comprises a first 3D convolutional layer, a first residual block, a down-sampling layer and a second 3D convolutional layer connected in sequence. The decoder comprises a transmission layer, a third 3D convolutional layer, a second residual block and a deconvolutional layer connected in sequence.

[0068] The prediction small model adopted in the present application combines a three-dimensional neural convolutional network (3DCNN) and a Transformer structure auto-encoder (AutoEncoder) model, i.e. an encoding neural network model. The input is spatial field data shaped as (D1, D1, D1, N) at multiple times after regular grid processing and working condition parameters shaped as (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] The data input layer inputs data shaped as (D1, D1, D1, N) spatial field data, and the working condition parameter input layer inputs data shaped as (N c , 1). Wherein, D1 is the dimension information of the regular grid, and N represents the data of the last N times.

[0070] For the encoder part, two 3D convolutional layers are used for initial feature learning and extraction. The residual block includes 12 residual units, each containing two 3D convolutional layers. Features are summed through skip connections to keep the input and output dimensions consistent.

[0071] Downsampling layer: The data dimension is gradually reduced by three sampling blocks. Each downsampling block reduces the spatial size through a convolution operation, and the final output has a low-dimensional shape of (D2,D2,D2,N). M ).

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

[0073] The transport layer transforms the data after splicing the encoder output and operating parameters through a multi-head attention mechanism.

[0074] Decoder section: The encoder output is upsampled through deconvolution layers to gradually restore the spatial resolution of the data. The decoder restores the data to the final size (D1,D1,D1,1) through deconvolution layers, which is used to predict the next spatial field data.

[0075] like Figure 2 As shown, Figure 2 The plus sign is used to indicate a residual connection, which means that the input and the output after the layer operation are added together. Arrows indicate the direction of data flow. A portion of the encoder output is connected to the decoder through skip connections.

[0076] The formulas for calculating convolution and pooling layers in the 3D convolutional layers of a small prediction 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 It is the output of layer (l-1), W l It is the convolution kernel of the l-th layer, b lis the bias term of the first layer, Conv3D is the three-dimensional convolution operation, MaxPool3D is the pooling operation, through multiple convolution and pooling stacking, a set of low-rank representation ZCNN with spatio-temporal features is obtained, which contains the deep features of the spatial fluid field.

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

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

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

[0083]

[0084] wherein Z encoder is the low-dimensional space of the Z CNN mapping, 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, f decoder () is the decoding function of the decoder.

[0085] In order to further improve the adaptability of the model under different working conditions, the working condition parameters P can be combined into the representation of the Transformer layer. Through the splicing operation, the working condition parameters are fused with the Transformer layer features, and the formula is as follows:

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

[0087] wherein Concat is the splicing function, and Z bottleneck,combined is the fusion feature of the working condition parameters 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 connection and residual block can also be added during network design. These structures help better transmit feature information, slow down the decay of information in deep network, and enable the model to better capture important features in long-term data. During the model training process, the network will perform back propagation 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 the future time.

[0092] In one exemplary embodiment, as shown in Figure 3 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 configured to perform steps 201-206.

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

[0094] Step 202, obtain the fluid space field data of the furnace at the current time step and the k-1 time steps before the current time step to form the input data of the tth time step; k represents the time sequence dependence degree.

[0095] Step 203, input the input data of the tth time step into the prediction small model to obtain the preliminary prediction fluid space field data of the t+1th time step.

[0096] Step 204, correct the preliminary prediction fluid space field data of the t+1th time step using the correction large model to obtain the fluid space field data of the t+1th time step.

[0097] Step 205, update the input data of the tth time step using the fluid space field data of the t+1th time step to obtain the input data of the t+1th time step.

[0098] Step 206, increase the value of t by 1, and return to the step of inputting the input data of the tth time step into the prediction small model to obtain the preliminary prediction fluid space field data of the t+1th time step, until the fluid of the furnace reaches a stable state.

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

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

[0101] In another exemplary embodiment, after obtaining the spatial field data predicted by the small prediction model, this data is fed into a large language model based on Qwen for further refinement. Since spatial field data is typically numerical, while the input to the Qwen model is character-based, the model breaks down each numerical value into individual characters for processing when inputting these values. This approach may lead to insufficient learning of the input data by the model, resulting in training instability or divergence. Therefore, the input numerical values ​​must be appropriately converted to ensure compatibility with the input format of the Qwen model.

[0102] To address this issue, a method of mapping numerical values ​​to special tokens is adopted. The Qwen model has scalable token bits, ranging from <|extra_0|> to <|extra_204|>, which can be used to represent different numerical values ​​or ranges. Since the number of tokens in Qwen is limited, directly mapping all numerical values ​​could lead to an excessive number of tokens, increasing training complexity. Therefore, it is necessary to divide the numerical values ​​into multiple intervals and map each interval to a corresponding token. Specifically, by designing a suitable interval division method, continuous numerical spaces can be mapped to these special tokens, thereby reducing the number of tokens fed into a large model while preserving the effective information of the numerical values. Let the input spatial field data be S={s1,s2,…,s…} J},in This represents the j-th data point (e.g., temperature, pressure, etc.). Assume the numerical range of the fluid spatial field data is [s]. min ,s max The data is divided into N intervals, each interval corresponding to a Token T. n The mapping function MAP() is:

[0103]

[0104] Among them, (s max -s min ) / N is the size of each interval. This indicates the rounding operation.

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

[0106] In step 206, after the spatial field data is obtained by the small prediction model, the data will be input into the correction large model for further correction to generate the next time spatial field prediction. Then, the new spatial field data obtained will be combined with the previous historical data to form new input. Through the sliding window method, the overall window slides backward as new input to continue the prediction of the next time, and is corrected again by the correction large model. This iterative prediction mechanism enables the model to continuously and efficiently predict and correct in the constantly updated time series, thereby continuously refining the spatial field data at future time.

[0107] Assuming that the initial spatial field prediction data Y t is sent to the correction large model to obtain the corrected result The calculation formula is as follows:

[0108]

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

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

[0111]

[0112] The iteration process can continue until the required step length is reached.

[0113] Compared with traditional numerical simulation methods, the iterative prediction mechanism has significant advantages in computational efficiency and resource consumption. Traditional numerical simulation methods usually require step-by-step solution of the entire fluid spatial field, involving complex equations and a large amount of calculation, especially when facing high-dimensional space-time data, the demand for computing time and resources is extremely large. By using the iterative prediction mechanism, the dependence on global calculation is effectively reduced, and the computational resources and time cost are greatly saved.

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

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

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

[0117] A boiler model is constructed based on Ansys software, which has a length, width and height of 24 meters, 55.8 meters and 16 meters respectively, and the number of grids reaches 1,885,691. The boundary conditions of the model include primary air speed, secondary air speed, overfire air speed, coal supply and its rate. By changing different boundary conditions for simulation, the boiler temperature is adjusted to 1000K to ignite the coal from the start of powder injection to the stable process, and the temperature field data of each step from ignition to stable combustion state is exported in turn, 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 of the last 5 time points on the next time data can be identified, and these data are sorted into a training set, which contains 5 time series field data as input and the temperature field data of the next time as label, providing a basis for subsequent prediction model training.

[0118] Due to the overall structural characteristics of the grid data, its fluid space field is mapped to a 48x48x48 cubic data one by one. In order to reduce the data dimension and improve the calculation efficiency, a down-sampling method is used to further map it to a 36x36x36 cubic data. According to the specific size of the boiler, the position of the center line is determined first. Then, the Euclidean distance of all grid points to the center line is calculated, and the weight is calculated according to the distance and mapped to the 48x48x48 cubic data. Finally, the 48x48x48 cubic data with weight information is further compressed to 36x36x36 cubic data through the down-sampling method, so as to facilitate the subsequent model input and training.

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

[0120] The input data is historical 5-time regular cubic data with a shape of 36x36x36x5, and the working condition parameter data needs to be input into the Transformer layer with a shape of 5x1. Through splicing in the Transformer layer, the data is finally transmitted to the 3D convolutional neural network (3DCNN) and the autoencoder for processing, and the predicted output of the spatial temperature field data at the future time is 36x36x36x1. In the training process, the above saved Euclidean distance weight function is used to assign the spatial points as a weighted item of the loss function. This weight function can effectively balance the training focus of the boundary area and the center area, avoiding overfitting of the boundary area.

[0121] Since the special Token expansion position of the Qwen large model is limited to 205, i.e. 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 serve as identifiers for numerical data, and by mapping the output data of each prediction small model with the corresponding real temperature field data, the internal network of the large model can be fine-tuned and corrected, allowing it to 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 step of the 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 the prompt word engineering, and the adjusted results can be saved to the historical data, providing users with more accurate and demand-oriented prediction results.

[0122] By applying the iterative prediction mechanism, the next state space temperature field data generated by the prediction small model and the correction large model is adjusted as the training set, and as the sliding window advances, the next state is continuously trained and predicted. During the training process, the prediction results are restored from 36x36x36 cubic data to 48x48x48 grid data using layer-by-layer mapping and upsampling techniques to conform to the grid structure and provide higher spatial resolution.

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

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

[0125] In order to accurately monitor the running state and combustion process of the boiler furnace, the embodiment of the present application adopts a distributed sensor network to collect key physical quantity data in real time inside the boiler and external input. These sensors cover various areas of the boiler furnace, collecting multi-dimensional data such as furnace temperature, flue gas velocity, pressure, etc., as well as key parameters such as primary air, secondary air, overfire air speed and coal supply. Each sensor module is individually configured according to its physical location and measurement requirements, and transmits the collected data to the central processing unit through a stable communication protocol. For large-scale data generated by the boiler furnace combustion process, the system designs an efficient distributed storage architecture to ensure efficient storage of historical data and real-time collected data. The system combines relational databases and non-relational databases, the former is used to store structured numerical data generated during the operation of the boiler (such as temperature, pressure, etc.), and the latter is used to store sensor data, combustion process logs, and boiler combustion images, etc. Unstructured data.

[0126] In order to support real-time inference and prediction of large-scale models, the 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 the task, ensuring efficient processing of different types of computing tasks. For example, small prediction models (such as preliminary prediction of combustion temperature field) can be deployed on low-power CPUs for operation, while correction large models involving complex operations (such as correcting grids with large errors) can be accelerated by GPUs. In addition, edge computing devices play a crucial role in the system. By deploying edge computing nodes on the boiler site, edge computing devices can reduce data transmission delay and ensure immediate processing and real-time feedback of data. Edge computing devices not only perform local data preprocessing, but also perform model inference, adjust control parameters and other tasks, and quickly respond to changes in boiler operation. Through the cooperative work of edge computing and central computing units, the system realizes low delay and high accuracy of data acquisition and processing.

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

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

[0129] 2. The method of introducing a large language model into the prediction model proposed in the present application fully utilizes the emergent ability 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 early prediction small model, and further optimize the model performance through the prompt word guidance in the training process. In this way, the model not only can adaptively improve the accuracy, but also can effectively deal with complex multi-modal data input, and enhance the robustness and generalization ability of the overall prediction system.

[0130] 3. The prediction small model and the correction large model proposed in the present application have high expansibility and strong generalization ability. After learning the spatio-temporal characteristics of the space field historical data in the encoder part of the autoencoder, the 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 way makes the model adapt to different working conditions and different types of space field learning tasks, thereby improving its applicability and accuracy in various application scenarios.

[0131] The technical features of the above embodiments can be combined in any manner. To make the description concise, not all possible combinations of the technical features in the above embodiments are described, but as long as the combinations of the technical features do not exist contradictions, they should be considered as the scope of the present application.

[0132] The principles and implementation modes of the present application are described by specific examples in this paper, and the above embodiment descriptions are only used to help understand the method and its core idea of the present application; at the same time, for those skilled in the art, according to the idea of the present application, the specific implementation mode and application range will be changed. In conclusion, the content of the specification should not be understood as a limitation of the present application.

Claims

1. A large model-based furnace internal fluid space field control system, characterized by, The system comprises a distributed sensor network, a data collection device, a distributed storage device, and an edge computing device. The distributed sensor network is arranged inside a furnace. The data collection device is connected to the distributed sensor network, and the distributed storage device is connected to the data collection device, and the distributed storage device is configured to store fluid space field data measured by the distributed sensor network. The edge computing device is configured to predict the fluid space field in the furnace based on the measured fluid space field data. The edge computing device is arranged with a small prediction model and a large correction model. The small prediction model is configured to preliminarily predict the fluid space field in the furnace based on the measured fluid space field data, and obtain preliminary prediction fluid space field data. The large correction model is configured to correct the preliminary prediction fluid space field data, and obtain prediction fluid space field data, and control the fluid space field in the furnace based on the prediction fluid space field data. The furnace fluid space field control system based on a large model further comprises a central computing unit. The central computing unit is configured to train an encoding neural network model to obtain a small prediction model, and construct a large correction model based on a Qwen large model. In the aspect of training the encoding neural network model to obtain the small prediction model, the central computing unit is specifically configured to: simulate the fluid space field of the furnace using Ansys software to obtain fluid space field data at each time step from the initial simulation time to the fluid stable state; preprocess the fluid space field data at each time step from the initial simulation time to the fluid stable state to obtain preprocessed fluid space field data at each time step from the initial simulation time to the fluid stable state, and construct a fluid space field data sequence; divide the fluid space field data sequence in a sliding window manner to construct a training set; construct a loss function; 2. The large model based furnace inner fluid space field control system according to claim 1, characterized in that, train the encoding neural network model using the training set and the loss function to obtain a trained encoding neural network model as the small prediction model. The edge computing device comprises 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.

3. The large model based furnace inner fluid space field control system according to claim 1, wherein, The CPU unit is configured to process the measured fluid space field data. One of the GPU units is arranged with the small prediction model, and the other GPU unit is arranged with the large correction model. The distributed storage device comprises a relational database and a non-relational database. The relational database is configured to store structured numerical data. The non-relational database is configured to store unstructured or semi-structured graphical data and / or log data. The furnace fluid space field control system based on a large model further comprises a power module and a heat dissipation system. The power module is connected to the data collection device, the distributed storage device, and the edge computing device. The power module is configured to supply power to the data collection device, the distributed storage device, and the edge computing device. The heat dissipation system is used for dissipating heat of the power module and the edge computing device.

4. The large model-based furnace inner fluid space field control system according to claim 1, wherein, The training sample in the training set is: , and the label of the training sample is ; is the i-th training sample; is the label of the i-th training sample, , and are the preprocessed fluid space field data of the i-th, i+1-th and i+k-1-th time steps, respectively, is the preprocessed fluid space field data of the i+k-th time step; the loss function is: ； ； ； wherein, is a loss function, is a loss for the mth grid point of the fluid space field, is a linear weight for the mth grid point of the fluid space field; is the Euclidean distance between the mth grid point of the fluid space field and the center point of the fluid space field, and a and β are coefficients of the linear relationship, , and are the x-axis, y-axis and z-axis coordinates of the mth grid point of the fluid space field, respectively, , and are the x-axis, y-axis and z-axis coordinates of the center point of the fluid space field, respectively.

5. The large model based furnace inner fluid space field control system according to claim 1, wherein, The fluid space field data from the simulation initial time to the fluid stable state is preprocessed to obtain preprocessed fluid space field data from the simulation initial time to the fluid stable state, specifically including: The target fluid space field data of the irregular grid is mapped into a cubic grid in a point mapping manner to obtain mapped target fluid space field data; the target fluid space field data is fluid space field data at any time step from the simulation initial time to the fluid stable state; The mapped target fluid space field data is down-sampled in a down-sampling manner to obtain preprocessed target fluid space field data.

6. The large model based furnace inner fluid space field control system according to claim 1, wherein, The encoding neural network model comprises 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 sequentially connected; The encoder and the decoder are also connected in a skip connection manner; The working condition parameter input layer is connected with the decoder; The formula of the encoder is: ； The formula of the Transformer layer is: ； The formula of the decoder is: ； ； wherein, is a low-dimensional space of the mapping, is an encoding function of the encoder, is an output of the data input layer, is a compressed representation of the Transformer layer, is a compression function of the Transformer layer, is a prediction result, is a decoding and prediction function, is a fusion feature of the working condition parameters and the compressed representation of the Transformer layer, is a concatenation function, is a working condition parameter.

7. The large model based furnace inner fluid space field control system according to claim 6, wherein, The encoder comprises a first 3D convolution layer, a first residual block, a down-sampling layer, and a second 3D convolution layer which are sequentially connected; The decoder comprises a transmission layer, a third 3D convolution layer, a second residual block, and a deconvolution layer which are sequentially connected.

8. The large model based furnace inner fluid space field control system according to claim 1, wherein, In the aspect 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: initializing the value of t as 0; acquiring fluid space field data of the furnace at the current time step and k-1 time steps before the current time step to form input data of the tth time step; k represents a time sequence dependence degree; inputting the input data of the tth time step into the prediction small model to obtain preliminary predicted fluid space field data of the t+1th time step; correcting the preliminary predicted fluid space field data of the t+1th time step by using the correction large model to obtain fluid space field data of the t+1th time step; updating the input data of the tth time step by using the fluid space field data of the t+1th time step to obtain input data of the t+1th time step; increasing the value of t by 1 and returning to the step of inputting the input data of the tth time step into the prediction small model to obtain preliminary predicted fluid space field data of the t+1th time step until the fluid of the furnace reaches a stable state.

9. The large model based furnace inner fluid space field control system according to claim 8, wherein, The time sequence dependence degree is acquired in the following manner: Let the value of the number is 1 ; computing a temporal mutual information quantity between the fluid space field data at the first time step and the fluid space field data at the second time step; judgment formula is established; wherein, is a threshold of timing mutual information quantity; If true, then let the value of the number increase by 1 and return to the step of "calculating the temporal mutual information quantity between the fluid spatial field data at the time step and the fluid spatial field data at the time step"; If not, output the value of as the timing dependency degree; wherein, the calculation formula of the time sequence mutual information quantity is: ； wherein is the fluid space field data for the th time step is the fluid space field data for the th time step is the temporal mutual information between the fluid space field data for the th time step is the entropy of the fluid space field data for the th time step is the entropy of the fluid space field data for the th time step is the cross entropy between the fluid space field data for the th time step is the fluid space field data for the th time step is the fluid space field data for the th time step

Images