Rotary kiln temperature field real-time numerical simulation method based on deep learning

Abstract

The invention discloses a rotary kiln temperature field real-time numerical simulation method based on deep learning, and the method comprises the following steps: S1), building a numerical simulation model of a rotary kiln based on an Euler-Lagrange discrete phase model, a coupled radiation model and a chemical reaction model through a computational fluid dynamics method; s2) determining boundary conditions and working condition parameters of the rotary kiln; s3) processing the temperature field sample data set obtained in the step 1) by using an intrinsic orthogonal decomposition method; s4) constructing a deep neural network of the gating circulation unit; s5) in a deployment application stage, obtaining a corresponding intrinsic orthogonal decomposition modal coefficient; s6, the internal temperature field of the rotary kiln is reconstructed, and real-time prediction.The prediction efficiency is greatly improved, computing resource consumption is remarkably reduced, the bottleneck of an existing numerical simulation technology in the aspect of real-time performance is broken through, and rapid and continuous thermal state evaluation and operation optimization decision support can be provided for an industrial site.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of rotary kiln numerical simulation and intelligent prediction, and particularly relates to a rotary kiln temperature field real-time numerical simulation method based on deep learning. BACKGROUND

[0002] A rotary kiln is a typical thermal process equipment, widely used in cement, metallurgy, chemical industry and other industrial fields. Its basic structure is a cylinder-shaped kiln body installed at an inclination and capable of continuous rotation. Through the rolling and forward movement of the material in the kiln, and the full heat exchange with the high-temperature gas flow, a series of thermal treatment processes such as heating, decomposition and calcination of the material are completed. In this process, the internal temperature field of the rotary kiln has a decisive influence on the reaction efficiency and product quality, and maintaining the uniformity and stability of the temperature field is the core of efficient operation.

[0003] However, in actual production process, due to the influence of factors such as changes in material properties, unstable fuel, and fluctuations in operating parameters, the internal thermal behavior of the rotary kiln presents high nonlinearity and instability, which easily leads to under-calcination or over-calcination, and further causes problems such as ring formation, blockage, and local overheating. These problems can seriously affect the heat exchange efficiency and material transmission efficiency, reduce the clinker quality, and increase the energy consumption and equipment operation risk. In addition, the internal temperature of the rotary kiln is often as high as 1800℃ or above, accompanied by high dust-containing environment and intense flow process, making it extremely difficult to directly measure the temperature distribution, and currently relying on manual experience and indirect monitoring, which has the problems of response lag and insufficient control accuracy.

[0004] To overcome the above challenges, computational fluid dynamics (CFD) technology is widely used in rotary kiln process modeling and mechanism analysis. Based on thermodynamics, heat transfer and fluid mechanics principles, CFD can numerically simulate the gas-solid flow and heat transfer process in the kiln in detail, thereby providing theoretical support for operating condition optimization and control strategy. However, CFD simulation has large calculation amount and complex solving process, especially in the case of three-dimensional unstable flow field and intense heat exchange, its high calculation cost limits its real-time application in industrial field, and it is more used in design verification and offline analysis stage.

[0005] In order to improve the real-time performance and engineering usability of CFD simulation, in recent years, researchers have tried to introduce model reduction methods, among which proper orthogonal decomposition (POD) technology is an effective dimension reduction method that can extract dominant modes from high-dimensional CFD simulation data to realize low-dimensional expression of temperature field evolution. By constructing a POD-based reduced model, the calculation cost can be significantly reduced while ensuring the prediction accuracy, which is an important technical path for building real-time digital twin models. However, the traditional POD method has certain limitations in dynamic response modeling, and it is difficult to handle fast nonlinear response under complex operating condition changes. SUMMARY

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: A real-time numerical simulation method for the temperature field of a rotary kiln based on deep learning, characterized by the following steps: S1): Using computational fluid dynamics, based on the Eulerian-Lagrange discrete phase model, coupled with the radiation model and the chemical reaction model, a numerical simulation model of the rotary kiln is established, and the results of the numerical simulation model are verified and the model is adjusted. S2): Determine the boundary conditions and operating parameters of the rotary kiln, and perform simulation calculations on various operating conditions based on the numerical simulation model to obtain a temperature field sample dataset. S3): The temperature field sample dataset obtained in step one is processed using the intrinsic orthogonal decomposition method to extract the first k intrinsic orthogonal decomposition modes and their corresponding eigenvalues ​​that characterize the main features of the rotary kiln temperature field. S4): Construct a gated recurrent unit deep neural network, establish an end-to-end mapping model from the operating parameters to the intrinsic orthogonal decomposition mode coefficients, and train the mapping model; S5): During the deployment and application phase, new operating condition parameters are input into the trained gated recurrent unit deep neural network to obtain the corresponding intrinsic orthogonal decomposition modal coefficients. S6): Based on the intrinsic orthogonal decomposition modes extracted in step two and the mode coefficients of the intrinsic orthogonal decomposition obtained in step three, the internal temperature field of the rotary kiln is reconstructed to achieve real-time prediction.

[0007] Furthermore, in step 1, a transient numerical simulation model of a rotary kiln with pulverized coal combustion as the core process is constructed based on the Eulerian-Lagrange discrete phase model. In the model, the gas phase adopts an incompressible turbulence model, and the solid phase adopts the particle tracking method to simulate the movement and reaction behavior of pulverized coal particles injected into the rotary kiln, coupling the volatilization, combustion, and pyrolysis processes of pulverized coal. The pulverized coal particles mainly include fixed carbon, volatile matter, ash, and moisture components, and their spatial distribution and reaction process in the rotary kiln are simulated by combining particle trajectories.

[0008] Furthermore, the model also needs to incorporate the main chemical reaction mechanisms in the gas phase and the influence of the radiation heat transfer model on the high-temperature environment. The gas phase components include O2, CO2, N2, CO, H2O(g), CH4, and H2. Through preset reaction pathways and kinetic parameters, a mass and energy exchange model between the gas and solid phases is established.

[0009] Furthermore, the intrinsic orthogonal decomposition method in step 3 specifically includes the following steps: S31): Constructing a snapshot matrix: Combining multiple high-dimensional temperature field samples into a snapshot matrix. Wherein m is the total number of grid nodes, and n is the number of samples; S32): singular value decomposition: singular value decomposition is performed on the snapshot matrix to obtain the modal function of the intrinsic orthogonal decomposition ; S33): modal truncation: the first k modes with cumulative energy greater than 99% are selected for retention; S34): state representation: the temperature field sample is represented as a linear combination of these modes, and the modal coefficient is used to represent the reduced state.

[0010] Further, the training process of the gated recurrent unit network in step 4 includes the following steps: S41): divide the sample data into a training set and a validation set; S42): the network is initialized using a preset weight initialization method to initialize the gated recurrent unit network parameters; S43): construct a recurrent neural network containing at least one gated recurrent unit layer; S44): the network input is a working condition variable vector, and the output is the modal coefficient of the intrinsic orthogonal decomposition; S45): use mean square error as the loss function, and use gradient descent algorithm for back propagation; S46): evaluate the prediction accuracy through the validation set, and meet the preset accuracy requirement.

[0011] Further, the step 6 specifically includes the following steps: S61): obtain the predicted modal coefficient of the intrinsic orthogonal decomposition ; S62): obtain the basis function of the intrinsic orthogonal decomposition; S63): linearly combine the modal coefficient and the basis function to obtain the full-field temperature distribution, wherein the linear combination formula is .

[0012] The beneficial effects of the present application are: 1. The present application is based on DPM coal combustion modeling, which combines radiation and reaction kinetics model, has high simulation accuracy, comprehensive working condition coverage, and stronger physical field expression ability.

[0013] 2. The combination of POD reduction and GRU neural network not only significantly reduces the calculation complexity of the original CFD model, but also improves the modeling ability of transient evolution trend, and meets the dynamic prediction demand under complex working conditions.

[0014] 3. The model is realized for industrial deployment, has fast response and strong generalization ability, and can be directly used for the construction of the rotary kiln digital twin system, providing support for intelligent management and control of production processes.

[0015] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0016] Fig. 1 This is a CFD mesh diagram of the present invention; Fig. 2 This is a flowchart of the prediction method of the present invention; Fig. 3 This is the network structure of the GRU neural network of the present invention. Detailed Implementation

[0017] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby providing a clearer and more explicit definition of the scope of protection of the present invention. Specific implementation examples: like Figs. 1 to 3 The deep learning-based real-time numerical simulation method for the temperature field of a rotary kiln, as shown, includes the following steps: S1): Using computational fluid dynamics, based on the Eulerian-Lagrange discrete phase model, coupled with the radiation model and the chemical reaction model, a numerical simulation model of the rotary kiln is established, and the results of the numerical simulation model are verified and the model is adjusted. S2): Determine the boundary conditions and operating parameters of the rotary kiln, and perform simulation calculations on various operating conditions based on the numerical simulation model to obtain a temperature field sample dataset. In the above steps, the numerical simulation module, based on the Eulerian-Lagrange discrete phase model (DPM), constructs a transient numerical simulation model of a rotary kiln with pulverized coal combustion as the core process. The model employs an incompressible turbulence model for the gas phase and a particle tracking method for the solid phase to simulate the motion and reaction behavior of pulverized coal particles injected into the rotary kiln. It couples the volatilization, combustion, and pyrolysis processes of pulverized coal, and considers the main chemical reaction mechanisms in the gas phase and the influence of the radiative heat transfer model on the high-temperature environment, enabling accurate prediction of the temperature field evolution within the rotary kiln.

[0019] The simulation primarily involves gaseous components such as O2, CO2, N2, CO, H2O(g), CH4, and H2. A mass and energy exchange model between the solid and gaseous phases is established using pre-defined reaction paths and kinetic parameters. Coal powder particles, including fixed carbon, volatile matter, ash, and moisture, are simulated based on particle trajectories to depict their spatial distribution and reaction process within the rotary kiln. To better reflect actual industrial conditions, the numerical model also incorporates the simulation of radiative heat flux, using the P-1 model to approximate high-temperature radiative heat transfer and improve the physical accuracy of temperature prediction.

[0020] This model focuses on the numerical simulation of the pulverized coal combustion process in the rotary kiln, and conducts transient simulation on the temperature field of the high-temperature zone of the rotary kiln. By selecting the industrial field parameters under the typical working conditions for simulation, and comparing and analyzing with part of the field monitoring temperature data, it is verified that the simulation model has good accuracy and physical consistency in combustion behavior and temperature distribution, which provides high-quality sample data support for the construction of subsequent data-driven models.

[0021] Gas phase mass conservation equation:

[0022] Among them: is the gas density, is the gas velocity vector, is the gas mass source term, indicating the mass generated or lost due to phase change, reaction, etc.

[0023] Gas phase momentum conservation equation:

[0024] Among them: is the gas pressure, is the stress tensor, is the resistance between gas and solid phases, is the body force term; Gas phase energy conservation equation:

[0025] Among them is the gas phase enthalpy, is the gas temperature, is the effective thermal conductivity, is the energy source term, including reaction heat release, radiation absorption; The pulverized coal particles are tracked by the Lagrangian method, considering the evolution process of particle mass, velocity and temperature, and its motion control equation is:

[0026] Among them: is the single particle mass, is the particle velocity, is the gas phase drag force on the particle, is the gravity acceleration vector The particle energy conservation equation is:

[0027] Among them: is the particle specific heat capacity, is the particle temperature, is the gas temperature, is the convective heat transfer coefficient, is the particle surface area, is the radiation heat received by the particle, is the heat released by the chemical reaction inside the particle; For the rotary kiln coal combustion process, the process parameters that significantly affect the temperature field are selected, including coal feed rate, primary air flow and temperature, secondary air flow and temperature, etc. as input variables. Based on the established DPM model and the CFD simulation module coupled with chemical reaction and radiation, transient simulation calculation of multiple parameter combinations is carried out under the set boundary conditions. In order to avoid the exponential growth of parameter combinations, the orthogonal experimental design method is used to effectively balance the coverage of sample distribution and the consumption of computing resources, and to construct a representative working condition sample data set. S3): Using the proper orthogonal decomposition method to process the temperature field sample data set obtained in step one, extracting the first k proper orthogonal decomposition modes representing the main characteristics of the rotary kiln temperature field and their corresponding characteristic values; Using the proper orthogonal decomposition (POD) method to process the transient temperature field data obtained by CFD simulation, extracting the main modal characteristics of the rotary kiln under various working conditions, and constructing a reduced-order characteristic space, which includes the following steps: S31): Constructing a snapshot matrix: grouping multiple high-dimensional temperature field samples into a snapshot matrix , where m is the total number of grid nodes and n is the number of samples. S32): Singular value decomposition: singular value decomposition of the snapshot matrix to obtain the modal function of proper orthogonal decomposition ; S33): Modal truncation: selecting the first k modes with cumulative energy greater than 99% for retention. S34): State representation: representing the temperature field samples as a linear combination of these modes, and representing the reduced dimension state with modal coefficients.

[0028] The POD method can project high-dimensional temperature field data into a low-dimensional orthogonal basis space, retaining the main characteristic information of the original data, which is convenient for subsequent neural network modeling. Let the original data be , where , where is the total number of temperature field grids, is the number of simulation samples.

[0029] First, all sample data are centered and standardized to eliminate the influence of different variable dimensions and scales and reduce the interference of local outliers on overall feature extraction.

[0030] Then calculate the covariance matrix of the standardized sample data:

[0031] Covariance matrix Solve eigenvalues and eigenvectors, select the first Characteristic mode And the corresponding eigenvalue Thus, the reduced order expression form is obtained:

[0032] Wherein The first Principal modal basis vector matrix, The first Modal coefficient vector of the sample in the low-dimensional space.

[0033] Through the above POD processing, the high-dimensional temperature field data output by the original CFD simulation can be effectively reduced to a smaller characteristic dimension, providing training data support for subsequent use of GRU neural network to establish the mapping relationship between working condition parameters and temperature field.

[0034] S4): build a gated recurrent unit deep neural network, establish an end-to-end mapping model from the working condition parameters to the proper orthogonal decomposition modal coefficient, and train the mapping model; the training process of the gated recurrent unit network in step 4 includes the following steps: S41): divide the sample data into a training set and a validation set; S42): the network initializes the parameters of the gated recurrent unit network using a preset weight initialization method; S43): construct a recurrent neural network containing at least one gated recurrent unit layer; S44): the network input is the working condition variable vector, and the output is the corresponding proper orthogonal decomposition modal coefficient; S45): use mean square error as the loss function, and use gradient descent algorithm for back propagation; S46): evaluate the prediction accuracy through the validation set, and ensure that the average relative error on the final validation set is less than 3%, meeting the preset accuracy requirement.

[0035] S5): in the deployment application stage, input new working condition parameters to the trained gated recurrent unit deep neural network to obtain the corresponding proper orthogonal decomposition modal coefficient; S6): based on the proper orthogonal decomposition mode extracted in step two and the proper orthogonal decomposition modal coefficient obtained in step three, reconstruct the temperature field inside the rotary kiln to realize real-time prediction. The step 6 specifically includes the following steps: S61): obtain the predicted proper orthogonal decomposition modal coefficient ; S62): obtaining the base function of the eigen-orthogonal decomposition; S63): linearly combining the modal coefficients with the base function to obtain the full-field temperature distribution, wherein the linear combination formula is .

[0036] The above is only an embodiment of the present application, and does not limit the patent scope of the present application, and any equivalent structural transformation made by using the content of the present application specification and drawings, or directly or indirectly applied to other related technical fields, is also included in the patent protection scope of the present application.

Claims

1. A deep learning-based real-time numerical simulation method for a rotary kiln temperature field, characterized in that: Includes the following steps: S1): Using computational fluid dynamics, based on the Eulerian-Lagrange discrete phase model, coupled with the radiation model and the chemical reaction model, a numerical simulation model of the rotary kiln is established, and the results of the numerical simulation model are verified and the model is adjusted. S2): Determine the boundary conditions and operating parameters of the rotary kiln, and perform simulation calculations on various operating conditions based on the numerical simulation model to obtain a temperature field sample dataset. S3): The temperature field sample dataset obtained in step one is processed using the intrinsic orthogonal decomposition method to extract the first k intrinsic orthogonal decomposition modes and their corresponding eigenvalues ​​that characterize the main features of the rotary kiln temperature field. S4): Construct a gated recurrent unit deep neural network, establish an end-to-end mapping model from the operating parameters to the intrinsic orthogonal decomposition mode coefficients, and train the mapping model; S5): During the deployment and application phase, new operating condition parameters are input into the trained gated recurrent unit deep neural network to obtain the corresponding intrinsic orthogonal decomposition modal coefficients. S6): Based on the intrinsic orthogonal decomposition modes extracted in step two and the mode coefficients of the intrinsic orthogonal decomposition obtained in step three, the internal temperature field of the rotary kiln is reconstructed to achieve real-time prediction.

2. The deep learning-based real-time numerical simulation method of a rotary kiln temperature field according to claim 1, characterized in that: In step 1, a transient numerical simulation model of a rotary kiln with pulverized coal combustion as the core process is constructed based on the Eulerian-Lagrange discrete phase model. In the model, the gas phase adopts an incompressible turbulence model, and the solid phase adopts the particle tracking method to simulate the movement and reaction behavior of pulverized coal particles injected into the rotary kiln. The volatilization, combustion, and pyrolysis processes of pulverized coal are coupled. The pulverized coal particles mainly include fixed carbon, volatile matter, ash, and moisture components. The spatial distribution and reaction process of the particles in the rotary kiln are simulated by combining the particle trajectory.

3. The deep learning-based real-time numerical simulation method of a rotary kiln temperature field according to claim 2, characterized in that: The model also needs to incorporate the main chemical reaction mechanisms in the gas phase and the influence of the radiation heat transfer model on the high-temperature environment. The gas phase components include O2, CO2, N2, CO, H2O(g), CH4, and H2. Through the preset reaction pathways and kinetic parameters, a mass and energy exchange model between the gas and solid phases is established.

4. The deep learning-based real-time numerical simulation method of a rotary kiln temperature field according to claim 1, characterized in that: The intrinsic orthogonal decomposition method in step 3 specifically includes the following steps: S31): Constructing snapshot matrix: multiple high-dimensional temperature field samples are composed into a snapshot matrix where m is the total number of grid nodes, and n is the number of samples. S32): singular value decomposition: singular value decomposition is performed on the snapshot matrix to obtain the modal function of the eigen-orthogonal decomposition ; S33): Modal truncation: Select the top k modes with accumulated energy greater than 99% and retain them; S34): State representation: The temperature field sample is represented as a linear combination of these modes, and the state after dimensionality reduction is represented by the mode coefficients.

5. The deep learning-based real-time numerical simulation method of a rotary kiln temperature field according to claim 1, characterized in that: The training process of the gated recurrent unit network in step 4 includes the following steps: S41): Divide the sample data into a training set and a validation set; S42): The network initializes the parameters of the gated recurrent unit network using a preset weight initialization method; S43): Construct a recurrent neural network containing at least one gated recurrent unit layer; S44): The network input is a vector of operating condition variables, and the output is the modal coefficients of the corresponding intrinsic orthogonal decomposition; S45): Using mean squared error as the loss function, the gradient descent algorithm is used for backpropagation; S46): Evaluate the prediction accuracy using the validation set to ensure it meets the preset accuracy requirements.

6. The deep learning-based real-time numerical simulation method of a rotary kiln temperature field according to claim 4, characterized in that: Step 6 specifically includes the following steps: S61): obtaining predicted modal coefficients of the eigen-orthogonal decomposition ; S62): Obtain the basis functions of the eigenorthogonal decomposition; S63): linearly combine the modal coefficients with the basis functions to obtain the full-field temperature distribution, where the linear combination formula is .

Images