A Simulation Method for the Temperature Field of a Cement Calciner Based on Digital Twin

Abstract

The present invention belongs to the technical field of cement production, and specifically discloses a simulation method for the temperature field of a cement decomposition furnace based on digital twin. The method includes: firstly, preliminarily selecting input parameters according to the cement clinker burning process, and then establishing a time-series data set from the decomposition furnace. Secondly, perform data preprocessing through mean filtering and normalization, and divide the time-series data set into a training set, a validation set, and a test set. Establish a deep learning model to train and obtain a digital twin virtual model layer, and establish a geometric model of the decomposition furnace to train and obtain a prediction model; use the two trained models to perform anomaly detection and prediction simulation on the input variables of the decomposition furnace. The present invention has the advantages of prediction ability, operation optimization, and decision support by establishing a digital twin model of the decomposition furnace through a mechanism model and a data-driven model of the decomposition furnace, and can reflect the internal temperature distribution and dynamic changes of the decomposition furnace in real time and detect abnormal conditions during operation.

Description

Technical Field

[0001] The present invention belongs to the technical field of cement production, and more specifically, relates to a simulation method for the temperature field of a cement precalciner based on digital twin. Background Art

[0002] As a basic raw material, the production quality of cement directly determines the safety and stability of subsequent products. In cement production, the cement precalciner mainly completes tasks such as pulverized coal combustion and raw meal decomposition.

[0003] Currently, the research on the precalciner process has been relatively mature, but the situation of raw meal decomposition and pulverized coal combustion in the precalciner depends on the experience of operators at the production site.

[0004] Existing technologies such as the rotary kiln condition prediction method that couples simulation and deep learning neural network disclosed in the Chinese patent application with the application number 202411358309.8 obtain sample data through the simulation of the rotary kiln, combine the POD method with neural networks, and develop an efficient prediction model to achieve the rapid prediction of the flow field temperature field of the rotary kiln, significantly reducing the huge overhead of traditional CFD simulation in terms of computing resources and time. At the same time, it solves the limitations of previous simulation technologies in terms of real-time performance, and adjusts the prediction data through a correction algorithm, improving the practicability and reliability of the model, and realizing the reduction of the order of the rotary kiln and the deployment of the digital twin model to the industrial site to provide guidance for the production process in real time, ensuring the improvement of production efficiency and quality.

[0005] Existing technologies such as the digital twin system of a cement precalciner based on POD and deep learning disclosed in the Chinese patent application with the application number 202210256544.9 include a data acquisition and processing module, a database retrieval module, a model order reduction and prediction module, a comprehensive display module, and a remote control module connected in sequence. By combining computational fluid dynamics with POD and deep learning, it realizes the rapid calculation of the entire temperature field and flow field inside the cement precalciner based on a finite number of discrete operating points, and displays them in various ways. The present invention enables the production operators of the cement precalciner to more intuitively and quickly obtain the real-time changes in the temperature field and flow field inside the cement precalciner, so as to timely adjust the pulverized coal combustion and raw meal decomposition conditions in the furnace according to them.

[0006] Regarding the above technical solutions, obviously, there are still the following deficiencies in the current measurement of the temperature field of the cement precalciner: 1. Although it involves calculating the real-time changes in the entire temperature field and flow field inside the cement precalciner, there is still a certain lack of fineness in the description of the dynamic characteristics of the precalciner temperature field, and parameters are not selected from the most basic physical and chemical processes of cement clinker burning, and there are certain limitations in understanding and depicting the internal mechanism of complex physical processes.

[0007] 2. The depth and pertinence of data processing are insufficient, resulting in insufficient support of data quality for model training. Currently, the refined adjustment of data division considering the roles and interrelationships of different data sets has not been carried out, and the strategies for model parameter adjustment and preventing overfitting are not perfect and refined enough. The simulation purpose and application scenarios are single. Currently, it mainly focuses on the prediction and display of the flow field temperature field, and thus there are certain deficiencies in the comprehensiveness and effectiveness of using the simulation results to guide production and solve practical problems. Summary of the Invention

[0008] In view of this, to solve the problems raised in the above background technology, a simulation method for the temperature field of a cement precalciner based on digital twin is proposed.

[0009] The object of the present invention can be achieved by the following technical solutions: The present invention provides a simulation method for the temperature field of a cement precalciner based on digital twin. The method includes: selecting input parameters according to the cement clinker burning process and constructing a time series data set of the precalciner.

[0010] Preprocess and divide the time series data set of the precalciner to obtain a training set, a validation set and a test set, and use them as model inputs.

[0011] Establish a geometric model of the precalciner to train a prediction model, and establish a deep learning model based on CNN-GRU-Attention to train a digital twin virtual model layer.

[0012] Based on the trained digital twin virtual model layer and prediction model, perform anomaly detection and temperature field simulation on the precalciner, and output the predicted values of the temperature field of the precalciner.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) By combining the cement clinker burning process mechanism, data-driven technology and digital twin technology, the present invention realizes high-precision simulation and optimization of the temperature field of the cement precalciner, can reflect the temperature distribution and dynamic changes inside the precalciner in real time, timely discover abnormal situations during operation, optimize operation parameters, and improve production efficiency and clinker quality.

[0014] (2) Starting from the basic physical and chemical processes of cement clinker burning, the present invention ensures that the model can accurately reflect the complex physical processes inside the precalciner. At the same time, input parameters are comprehensively selected, and combined with the mechanism model and data-driven model, the understanding and description of the complex physical processes inside the precalciner are improved.

[0015] (3) Through in-depth data preprocessing and refined dataset partitioning, the present invention enhances the support of data quality for model training, ensuring more perfect strategies for model training, parameter adjustment, and overfitting prevention. Meanwhile, the simulation purposes and application scenarios are expanded from single prediction simulation to anomaly detection, operation optimization, and decision support.

[0016] (4) By establishing a digital twin virtual model layer and a decomposition furnace geometric model and combining deep learning technology, the present invention realizes high-precision simulation of the temperature field in the decomposition furnace. And through digital twin technology, it can real-time simulate the temperature field distribution and dynamic changes inside the decomposition furnace, significantly improving the fineness of the description of the dynamic characteristics of the temperature field.

[0017] (5) By using the trained digital twin virtual model and prediction model, the present invention not only realizes the prediction simulation of the temperature field but also supports anomaly detection and operation optimization. Through the simulation results, it can timely detect abnormal conditions in the operation of the decomposition furnace, provide optimization suggestions, and effectively guide actual production. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0019] Figure 1 It is a schematic flowchart of the implementation steps of the method of the present invention.

[0020] Figure 2 It is a flow-assisted sketch for constructing the digital twin layer of the present invention.

[0021] Figure 3 It is a flow-assisted sketch for constructing the virtual model layer of the present invention.

[0022] Figure 4 It is a flow-assisted sketch for constructing the CNN-GRU-Attention of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0023] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0024] Please refer to Figure 1As shown in the figure, the present invention provides a method for simulating the temperature field of a cement precalciner based on digital twin, and the method includes: Step 1, select input parameters according to the cement clinker burning process and construct a time series data set of the precalciner.

[0025] Step 2, preprocess and partition the time series data set of the precalciner to obtain a training set, a validation set and a test set, and use them as model inputs.

[0026] In the embodiment of the present invention, by performing in-depth data preprocessing and refined data set partitioning, the support of data quality for model training is improved, and the strategies for model training, parameter adjustment and preventing overfitting are made more perfect. At the same time, the simulation purpose and application scenarios are extended, from a single prediction simulation to anomaly detection, operation optimization and decision support.

[0027] Step 3, establish a geometric model of the precalciner to train a prediction model, and establish a deep learning model based on CNN-GRU-Attention to train a digital twin virtual model layer.

[0028] Step 4, use the trained digital twin virtual model layer and prediction model to perform anomaly detection and temperature field simulation on the precalciner, and output the predicted value of the temperature field of the precalciner.

[0029] In the embodiment of the present invention, by combining the mechanism of the cement clinker burning process, data-driven technology and digital twin technology, high-precision simulation and optimization of the temperature field of the cement precalciner are realized, which can reflect the temperature distribution and dynamic changes inside the precalciner in real time, timely discover abnormal situations during operation, optimize operation parameters, and improve production efficiency and clinker quality.

[0030] Regarding Step 1, it should be noted that the selected input parameters include offline collected data such as raw meal feeding amount, coal injection amount, temperature of the kiln tail flue gas chamber, tertiary air temperature, temperature at the outlet of the C4A cone, temperature at the outlet of the C4A cone and the outlet temperature of the precalciner, etc. as input parameters.

[0031] It is understandable that temperature sensors are installed at key positions of the precalciner, such as different heights and radial positions inside the furnace body, and each inlet and outlet, such as the raw meal feeding port, coal injection port, tertiary air inlet, kiln tail flue gas chamber outlet, precalciner outlet, etc. These sensors can collect temperature data in real time. For example, thermocouple sensors can accurately obtain temperature information by measuring the thermoelectric potential of different materials, and resistance temperature sensors measure temperature using the characteristic that resistance changes with temperature. And the sensors can communicate and interconnect with the DCS system.

[0032] Understandably, in the cement production scenario, the DCS system can comprehensively collect, transmit and store various key information in the production process, such as temperature data, pressure data, flow data, equipment operation data, operation data and environmental data, etc. Among them, temperature data are like tertiary air temperature, kiln tail smoke chamber temperature, decomposition furnace outlet temperature and each preheater cone outlet temperature, such as C4A cone outlet temperature, etc., and pressure data include decomposition furnace internal pressure and each cone pressure, such as C4A cone pressure, C4B cone pressure, etc. Flow data includes raw material feeding amount, coal injection amount, tertiary air volume and other flow data, equipment data such as decomposition furnace diameter, height, etc., motor start and stop status, valve opening, etc., operation data such as primary air volume, secondary air volume, kiln tail flue gas temperature, decomposition furnace temperature setting value, decomposition furnace pressure setting value and residence time, etc., and environmental data such as external ambient temperature and atmospheric pressure.

[0033] Among them, temperature data can reflect the thermal state in the decomposition furnace. If the temperature of the smoke chamber at the tail of the kiln rises abnormally, it may be due to poor ventilation in the kiln or incomplete combustion of coal powder. If the outlet temperature of the decomposition furnace is too low, it may lead to insufficient decomposition of raw materials and affect the quality of clinker. The pressure change can intuitively reflect the ventilation and material flow state of the system. If the pressure inside the decomposition furnace fluctuates greatly, it may indicate that there is a risk of blockage in the system, affecting the continuity of production. Flow data is directly related to the chemical reaction and heat exchange efficiency in the decomposition furnace. Unstable raw material feeding will affect the decomposition effect of raw materials. Inappropriate coal injection will cause temperature fluctuations, which will affect the burning quality of clinker. The operating state of the motor determines whether the equipment is operating normally, and the valve opening affects the flow and direction of materials and gases. These data are crucial to ensure the smooth operation of the entire production process. Therefore, comprehensive collection is carried out through the DCS system to ensure the operability and effectiveness of subsequent simulation. The primary air volume is used to transport pulverized coal and provide the oxygen required for combustion, which affects the combustion efficiency and temperature distribution. The secondary air volume is used to supplement the oxygen required for combustion and further adjust the combustion state in the furnace. The flue gas temperature at the kiln tail refers to the flue gas temperature entering the decomposition furnace from the rotary kiln, which affects the heat input of the decomposition furnace. The decomposition furnace temperature setting value refers to the target temperature of the decomposition furnace. The residence time refers to the residence time of the raw material in the decomposition furnace.

[0034] It should be noted that step 2 includes preprocessing including mean filtering and normalization, and the normalization can be performed by minimum-maximum normalization.

[0035] What needs to be explained about step 3 is that the geometric model of the decomposition furnace is established, including: R1. According to the design drawing of the decomposition furnace of the cement plant, a 3D model of the decomposition furnace is established using three-dimensional modeling software.

[0036] R2. Set the 3D model of the precalciner to have 4 raw meal feeding inlets, 4 tertiary coal injection inlets, 2 tertiary air inlets, 1 kiln tail gas chamber outlet, 1 precalciner outlet, and the inner wall inlet and walls of the precalciner according to the cement burning process.

[0037] R3. Adopt the hybrid mesh generation technology to mesh the internal area, boundary layer, and pre-set analysis area of the precalciner, and detect the mesh quality by calculating the orthogonal quality index of the mesh; among them, the hybrid mesh includes layered polyhedral mesh, pure polyhedral mesh, and hexahedral mesh.

[0038] Furthermore, the hexahedral mesh is used for the internal area of the precalciner. The number of hexahedral meshes is less under the same size, which is suitable for simple geometric bodies and can reduce the calculation amount. The layered polyhedral mesh is used for the transition area from the outer wall to the inside of the precalciner, that is, the boundary layer. The layered polyhedral mesh can better capture the physical phenomena in the boundary layer. The pure polyhedral mesh is used for the mixing area of raw meal, pulverized coal, and air at the bottom of the precalciner. The pure polyhedral mesh is suitable for complex geometric bodies and can improve the accuracy of the simulation.

[0039] Evaluate the mesh generation quality by calculating the orthogonal quality index of the mesh. The orthogonal quality formula is ; is the mesh orthogonal quality index, which is used to evaluate the relationship between the shape of the mesh element and its actual volume. The closer the orthogonal quality is to 1, the closer the shape of the mesh element is to orthogonal, and the higher the mesh quality. represents the determinant of the Jacobian matrix. The Jacobian matrix is usually expressed as , and the Jacobian determinant is used to describe the change ratio relationship of the volume element from the original coordinate system to the new coordinate system . 、 、 respectively represent the partial derivatives of the coordinates of the mesh element in the reference space. 、 and are the coordinates of the mesh element in the physical space, is the volume of the mesh element. This formula is mainly used for the quality evaluation of mesh generation, especially in finite element analysis (FEA) or computational fluid dynamics (CFD) simulations. By calculating the orthogonal quality index of the mesh, the quality and rationality of the mesh generation can be evaluated, providing a reliable basis for subsequent numerical calculations. Table 1 shows the orthogonal quality of the meshes.

[0040] Table 1 Mesh Orthogonal Quality

[0041]

[0042] R4. Set the boundary conditions for the raw meal feeding inlet, tertiary coal injection inlet, tertiary air inlet, kiln tail gas chamber outlet, calciner outlet, inner wall and wall of the calciner.

[0043] R5. Map the nearest neighbor algorithm based on Euclidean distance into a data model, and conduct physical simulation inside the calciner. According to the simulation results, adjust the model parameters, and thus output the geometric model of the calciner.

[0044] In the embodiment of the present invention, starting from the basic physical and chemical processes of cement clinker burning, it is ensured that the model can accurately reflect the complex physical processes inside the calciner. At the same time, input parameters are comprehensively selected, and by combining the mechanism model and the data-driven model, the understanding and characterization of the complex physical processes inside the calciner are improved.

[0045] Further, the physical simulation inside the calciner can be carried out by combining the mass conservation equation, momentum conservation equation and energy conservation equation. According to the simulation results, adjust the model parameters to achieve the best combustion state simulation. Specifically, it includes adjusting operation parameters such as raw meal feeding amount, coal injection amount, and tertiary air temperature, as well as optimizing grid division and boundary condition settings to improve the simulation accuracy.

[0046] Specifically, adjusting the model parameters can be carried out by combining optimization algorithms such as genetic algorithm and particle swarm optimization algorithm for automatic parameter search. At the same time, it combines grid optimization such as grid encryption and type selection, boundary condition adjustment, physical model selection, initial condition setting, solver configuration such as time step and convergence criterion, parameter sensitivity analysis such as single-factor and multi-factor analysis, and experimental data comparison and verification such as error analysis and model calibration and other methods to comprehensively optimize the model parameters, ensuring the accuracy and reliability of the simulation model, and thus providing a solid foundation for the temperature field simulation of the calciner. Existing common technologies are used in the process and will not be elaborated.

[0047] It can be understood that the boundary condition setting is a key step in the construction of the calciner geometric model, which is used to define the physical quantity distribution in areas such as inlets, outlets and walls during the simulation process. Specifically, it includes setting the mass flow rate, temperature and chemical composition of the raw meal feeding inlet, tertiary coal injection inlet and tertiary air inlet. Setting the temperature, pressure and composition of the kiln tail gas chamber outlet and the calciner outlet. And setting the temperature, thermal conductivity and roughness of the inner wall and wall of the calciner. In addition, it is also necessary to set the initial temperature field, pressure field and flow velocity field at the beginning of the simulation, and set boundary conditions such as thermal radiation and chemical reactions according to the simulation requirements. The setting of these boundary conditions directly affects the accuracy and reliability of the simulation, ensuring that the model can truly reflect the actual working conditions of the calciner.

[0048] Regarding step 3, it should be noted that the specific training process of the prediction model is as follows: G1. After completing the construction of the decomposition furnace geometric model, set its initial parameters. Based on the mass conservation equation, momentum conservation equation, and energy conservation equation, in the geometric model, first, without considering the inlet and outlet temperatures, perform iterative operations according to the set number of iterations until the outlet wind speed of the decomposition furnace reaches the set maximum value.

[0049] G2. Import the chemical analysis of pulverized coal and the chemical analysis of raw meal, and start simulating the relevant chemical reaction behaviors of pulverized coal combustion and raw meal decomposition.

[0050] G3. After completing the chemical reaction simulation, analyze the simulation results and compare them with the actual data. If the difference between the simulation results and the corresponding actual data is greater than the set difference, adjust the model parameters and output the final prediction model.

[0051] Specifically, the initial parameters of the decomposition furnace geometric model are comprehensively set by combining professional knowledge in the cement production field, the statistical analysis results of a large amount of historical production data, and common experience values in the industry. These parameters cover chemical reaction kinetics parameters, such as the rate constants of different pulverized coal combustion reactions and raw meal decomposition reactions, which are initially determined based on the reaction characteristics of common pulverized coal and raw meal components, heat transfer and mass transfer parameters such as the heat transfer coefficient, estimated by referring to the heat transfer research data of similar industrial equipment and combining the actual structure and material of the decomposition furnace. The mass transfer coefficient is estimated according to the principles of fluid mechanics and the material flow characteristics in the decomposition furnace, and the initial values of fluid physical property parameters, such as the density and viscosity of the fluid, are determined according to the physical property manuals of pulverized coal, raw meal, and gas at different temperatures and pressures.

[0052] It is understandable that the data of the chemical analysis of pulverized coal and the chemical analysis of raw meal detail the content of volatile matter, fixed carbon, ash, moisture, and various trace elements in the pulverized coal, the proportions of the main components such as calcium carbonate, silica, alumina, and iron oxide in the raw meal, and the information on impurity components that may affect chemical reactions.

[0053] It is understandable that the data of the chemical analysis of pulverized coal and the chemical analysis of raw meal detail the content of volatile matter, fixed carbon, ash, moisture, and various trace elements in the pulverized coal, the proportions of the main components such as calcium carbonate, silica, alumina, and iron oxide in the raw meal, and the information on impurity components that may affect chemical reactions.

[0054] It is also understandable that the physical simulation results based on the three major conservation laws provide a large amount of valuable data for model training. These data reflect the physical and chemical changes of the precalciner under different working conditions, and the model improves its prediction ability by learning the laws in these data. During the training process, the results under different parameter settings are simulated according to the conservation laws, and the differences between the model prediction values and the actual simulation values calculated based on the conservation laws are observed. These differences are used to adjust model parameters such as the weights and thresholds of the neural network, making the model prediction more accurate. For example, if there is a deviation between the temperature field predicted by the model and the temperature field calculated according to the law of conservation of energy, the model parameters can be adjusted through an optimization algorithm to make the model prediction closer to the real situation, thereby continuously optimizing the prediction model and improving its prediction accuracy and reliability for the operating state of the precalciner.

[0055] Regarding step 3, it should be noted that the digital twin virtual model layer consists of a digital twin layer and a virtual model layer. Among them, please refer to Figure 2 As shown, the specific acquisition process of the digital twin layer is as follows: C1. Screen out the tertiary air temperature, kiln tail gas chamber temperature, tail coal chemical analysis, and cement raw material chemical analysis from the input parameters.

[0056] C2. Take the precalciner geometric model as the physical model and generate temperature field data through physical model simulation.

[0057] C3. Perform preprocessing such as mean filtering and normalization on the tertiary air temperature, kiln tail gas chamber temperature, tail coal chemical analysis, and cement raw material chemical analysis to obtain the preprocessed data.

[0058] C4. Use the preprocessed data as input and build a temperature field prediction model based on the K-nearest neighbor algorithm.

[0059] C5. Integrate the temperature field data generated by the physical model and the temperature field prediction data based on the K-nearest neighbor algorithm to build a data model.

[0060] C6. Through the data model, map the collected and predicted temperature data to the actual spatial structure of the precalciner to form a temperature field distribution.

[0061] C7. Based on the mapped temperature field, combine the real-time operation data and historical data of the precalciner to build the digital twin layer model.

[0062] It is understandable that based on the K-nearest neighbor algorithm, by calculating the distance between the target data and the historical data, K nearest neighbor data are screened out. Based on the temperature field information of these neighboring data, the temperature field under the current working condition is predicted, providing more timely prediction data for the subsequent model.

[0063] Understandably, the temperature field data generated by integrating the physical model and the temperature field prediction data based on the K-nearest neighbor algorithm are integrated to construct a data model. In this process, data mining and machine learning technologies are used to deeply analyze the correlations and potential laws between the data, enabling the data model to more accurately describe the characteristics of the decomposition furnace temperature field.

[0064] It is also understandable that through the data model, the temperature data collected and predicted are mapped onto the actual spatial structure of the decomposition furnace to form a temperature field distribution. During the mapping process, factors such as the geometric shape and internal components of the decomposition furnace are fully considered to ensure the accuracy of the temperature field mapping.

[0065] Please refer to Figure 3 and Figure 4 As shown, further, the specific acquisition process of the virtual model layer is as follows: H1. Generate model files through the digital twins in the digital twin layer and organize them according to the time series. According to the requirements of model training, the continuous time series data is divided into windows of a fixed length, and the data within each window is used as an independent input sample.

[0066] H2. Perform feature enhancement and normalization processing on the divided time series data, and input the preprocessed input sequence into the CNN layer of the deep learning model based on CNN-GRU-Attention. In the convolutional layer, convolutional kernels of different sizes and strides are set.

[0067] Preferably, the input sequence passes through the convolutional layer of the CNN layer, and convolution operations are performed by sliding the convolutional kernel over the data to extract features in the input sequence. The convolution operation formula is: , where is the data in the input sequence, is the convolutional kernel weight, is the bias, and are the convolutional kernel sizes, is the coordinate of the output feature map. The data processed by the convolutional layer is then subjected to a dimensionality reduction operation through the pooling layer. The maximum pooling method is used, and the pooling window size is 2×2. The maximum value within each window is taken as the pooling output.

[0068] Understandably, the data within each window is used as an independent input sample for subsequent model training and prediction. The selection of the window length needs to comprehensively consider the change frequency of the temperature field data and the model's ability to capture time series information, and the optimal value is generally determined through experiments and verification. For example, for temperature field data with relatively frequent changes, a shorter window length such as 5 - 10 time steps can be selected. For data with relatively slow changes, a longer window length such as 20 - 30 time steps can be selected.

[0069] Among them, the GRU layer contains a reset gate and an update gate. The reset gate controls the degree of forgetting or resetting of the previous hidden state before calculating the candidate hidden state, and the update gate is used to determine the amount of previous information that should be passed to the future.

[0070] The reset gate controls the degree of forgetting or resetting of the previous hidden state before calculating the candidate hidden state, and the update gate is used to determine the amount of previous information that should be passed to the future. The calculation formula for the reset gate is and the calculation formula for the update gate is The calculation formula for the candidate hidden state is and the calculation formula for the final hidden state is , is the Sigmoid activation function. The Sigmoid activation function compresses the input value between 0 and 1 and is commonly used in the gating mechanism to control the flow of information. , and are weight matrices. and respectively represent the weight matrices used for linearly transforming the input vector and the previous hidden state in the reset gate calculation and the update gate calculation. represents the weight matrix used for linearly transforming the input vector and the previous hidden state after being processed by the reset gate in the candidate hidden state calculation. , and respectively represent the bias vectors in the reset gate calculation, the update gate calculation, and the candidate hidden state calculation. represents the concatenation of the input vector and the previous hidden state to form a new vector as the input for subsequent calculations. is the hyperbolic tangent activation function, which compresses the input value between -1 and 1 and is used to calculate the candidate hidden state . represents the element-wise multiplication (Hadamard product) operator, that is, multiplying the elements at the corresponding positions of two vectors.

[0071] H3. Through the sliding convolution operation of the convolution kernel on the input data, local features in the data are extracted. After being processed by the convolutional layer, the data is then reduced in dimension through the pooling layer using the max pooling or average pooling method.

[0072] H4. Convert the input sequence processed by the CNN layer into the form of feature vectors, send them to the GRU layer, and input the feature sequence output by the GRU layer into the Attention layer for attention mechanism calculation.

[0073] H5. The output of the Attention layer passes through an output layer for the final result output. The output layer adopts a fully connected layer structure. The calculation formula of the output layer is , is the weight matrix of the output layer, represents the predicted output result of the model, represents the feature vector output by the Attention layer, Introduce a learnable constant term for the calculation result to increase the expressive power of the model. The overall function of this formula is to map the feature vector obtained by the previous layer processing through the operations of the weight matrix and the bias vector to the required output dimension to obtain the final predicted value.

[0074] Preferably, the Attention layer is divided into paths, paths and paths. For the input matrix , each path first maps it through a linear layer to obtain the corresponding query vector , key vector and value vector , where the formulas are respectively , , , where , and are respectively path, path and path weight matrices. Calculate the dot product of and , and then perform a scaling operation. The formula is , is the dimension of the key vector , represents the transpose symbol, represents the score in Attention. Perform normalization processing on using the Softmax function to obtain the attention weight . Finally, multiply the attention weight by the value vector and sum to obtain the output of the Attention layer. Through this process, the Attention layer can calculate the similarity or correlation between the feature sequence and the feature value, highlight the weights of important elements, and screen out the information more valuable for the temperature field prediction.​

[0075] Specifically, Figure 3 in represents the absolute value of the difference between the predicted value and the actual value, represents the threshold of the difference between the predicted value with the set permission and the actual value.

[0076] Furthermore, the training process of the digital twin virtual model layer also includes: 1) Using the training set to train the constructed model based on CNN-GRU-Attention, calculating the difference between the model prediction result and the true value using a pre-set loss function, calculating the gradient of the loss function with respect to the model parameters using the backpropagation algorithm during the training process, and using an optimizer to update the model parameters according to the gradient to minimize the loss function.

[0077] 2) Adopting a learning rate decay strategy to gradually reduce the learning rate as the number of training rounds increases, and applying regularization techniques to constrain the virtual model layer.

[0078] Understandably, the loss function can specifically select the mean squared error loss function MSE, the mean absolute error loss function MAE, etc. to measure the difference between the model prediction result and the true value. The optimizer can select Adam, SGD, etc. to update the model parameters according to the gradient. Regularization techniques can select L1 regularization, L2 regularization, Dropout, etc. to constrain the model to prevent overfitting and improve the generalization ability of the model. By continuously adjusting the model parameters and training strategies, the model can more accurately learn the rules in the temperature field data, improve the prediction ability of the decomposition furnace temperature field, and finally train the digital twin virtual model layer.

[0079] It should be noted that in order to further improve the performance of the Attention layer, the basic attention mechanism can be improved and extended. For example, adopting the multi-head attention mechanism, obtaining multiple groups of query vectors, key vectors, and value vectors through multiple different linear transformations, calculating the attention weights respectively and fusing them, so as to be able to simultaneously focus on different aspects of the input features and enhance the model's ability to capture complex features. Or introducing position encoding information, integrating the position information of the input sequence into the attention calculation, enabling the model to better understand the relative position relationship of the elements in the sequence and improving the processing ability of time series data.

[0080] It is also understandable that according to the actual application requirements, the prediction results need to be adjusted and post-processed, such as restricting the predicted temperature value within a reasonable range, or smoothing the prediction results to improve the stability and reliability of the prediction.

[0081] In the embodiments of the present invention, by establishing a digital twin virtual model layer and a decomposition furnace geometric model, and combining deep learning technology, high-precision simulation of the temperature field of the decomposition furnace is achieved. And through digital twin technology, the temperature field distribution and dynamic changes inside the decomposition furnace can be simulated in real time, significantly improving the fineness of the description of the dynamic characteristics of the temperature field.

[0082] Regarding what needs to be supplemented in step 4, anomaly detection and temperature field simulation of the decomposition furnace are carried out, including: obtaining real-time data from the actual operation of the decomposition furnace as the current input data, and performing mean filtering and normalization processing on the current input data.

[0083] Input the preprocessed data into the trained digital twin virtual model layer for temperature field simulation, and output a set of temperature field data inside the decomposition furnace through the digital twin virtual model layer, including temperature distribution, pressure distribution, flow rate distribution, etc.

[0084] Determine the steady-state operating point of the virtual model layer, establish a hybrid model using a series-parallel connection method, and judge whether the predicted value of the input parameters of the decomposition furnace deviates from the normal value through the hybrid model. If it deviates, it is determined as an abnormal working condition, and a warning message is output, and the next step is executed. If it does not deviate, the next step is directly executed.

[0085] Perform temperature field simulation based on the hybrid model and output the predicted value of the decomposition furnace temperature field.

[0086] Furthermore, temperature field simulation is carried out, including: geometric model simulation: Based on the 3D geometric model and mesh division of the decomposition furnace, simulations of flow, heat transfer, and chemical reactions are carried out.

[0087] Physical model simulation: Based on the laws of conservation of mass, momentum, and energy, calculate the temperature field distribution inside the decomposition furnace.

[0088] Chemical reaction simulation: Import coal powder chemical analysis and raw meal chemical analysis, and simulate the chemical reaction behaviors of coal powder combustion and raw meal decomposition.

[0089] Even further, a hybrid model is established, including: connecting a data-driven model based on CNN-GRU-Attention in series behind the temperature field data model of the digital twin layer.

[0090] Parallelly connect the serially connected model with another data-driven model based on CNN-GRU-Attention to form a series-parallel hybrid model.

[0091] In a specific embodiment, if the predicted value deviates from the normal value, the system will output an abnormal warning message and store it in the database of the digital twin layer to provide reference data for subsequent analysis. At the same time, it will prompt the operator to make corresponding adjustments. At this time, the temperature field simulation results can help the operator better understand the causes and impacts of the abnormality. Furthermore, based on the abnormal detection results, the operator can adjust parameters such as the raw meal feeding amount, coal injection amount, and tertiary air temperature to optimize the operation status of the precalciner.

[0092] By using the trained digital twin virtual model and prediction model, the embodiments of the present invention not only achieve the prediction simulation of the temperature field, but also support abnormal detection and operation optimization. Through the simulation results, abnormal situations in the operation of the precalciner can be detected in a timely manner, and optimization suggestions can be provided to effectively guide actual production.

[0093] In another embodiment of the present invention, it also includes verifying the effectiveness of the cement precalciner temperature field simulation method described in the present invention.

[0094] When the raw meal feeding amount and the tail coal data are stable and when the raw meal feeding amount and the coal injection amount are relatively stable, the input variables are simplified. The temperature of the kiln tail flue gas chamber, the tertiary air temperature, the outlet temperature of the C4A cone, and the outlet temperature of the C4A cone are selected as the inputs of the CNN-GRU-Attention model. The input of the geometric model in the data twin layer and other inputs in the above inputs are the actual operation data of the factory.

[0095] Thereby, 103 groups of cement precalciner temperature field samples are generated. The temperature of the kiln tail flue gas chamber, the tertiary air temperature, the outlet temperature of the C4A cone, and the outlet temperature of the C4B cone in the 103 groups of data are the actual operation data of the factory. The outlet temperature of the precalciner in the first 103 groups of samples is replaced by the calculation result of the data twin layer. The 103 groups of temperature field datasets in the data twin layer are used as the training set to train the steady-state working point of the virtual model layer. The outlet temperature of the precalciner in 27 groups of data is the actual operation data of the factory. Finally, the outlet temperature of the precalciner predicted by the steady-state working point of the virtual model layer is compared with the actual operation data of the factory. As the number of iterations increases, the error curve of the training set gradually tends to be stable, indicating that the training result is good and has good prediction ability for the outlet temperature of the precalciner. For example, after 250 iterations, the error curve of the test set gradually decreases and finally drops below 0.1. And when the iteration of the model is completed, the normalized test set of the outlet temperature of the precalciner is input into the established CNN-GRU-Attention model. The predicted results of the test set and the true comparison results are shown in Table 2.

[0096] Table 2 Comparison of the predicted values and actual values in the data twin layer

[0097]

[0098] As shown in Table 2, except for the normal state of the raw material feeding amount and the tail coal, the abnormal variables that appeared abnormal in the test set among the rest are the internal pressure of the precalciner, the pressure of the C5A cone, the pressure of the C5B cone, the temperature of the kiln tail flue gas chamber, the temperature at the outlet of the precalciner, the internal temperature of the precalciner, and the temperature at the outlet of the precalciner. The data of the above variables in the test set exceeded the threshold given by the validation set, and the model judged them as abnormal. The deviation of the predicted values of each variable is shown in Table 3.

[0099] Comparison between Predicted Values and Actual Values in Table 3

[0100]

[0101] As can be seen from the above table, when an abnormal working condition occurs, CNN-GRU-Attention can judge that the predicted value has deviated from the normal value. Since the temperature of the kiln tail flue gas chamber drops most significantly, and the other temperatures all drop to varying degrees, the internal pressure of the precalciner, the pressure of the C5A cone, and the pressure of the C5B cone deviate from the normal threshold first. It is thus judged that the data abnormality is caused by the on-site operator in the cement plant opening the kiln tail door to clean the kiln skin.

[0102] The above content is only an example and explanation of the concept of the present invention. Those skilled in the art of this technology can make various modifications or supplements to the described specific embodiments or use similar methods to replace them. As long as they do not deviate from the concept of the invention or exceed the scope defined by the present invention, they should fall within the protection scope of the present invention.

Claims

1. A simulation method for the temperature field of a cement decomposition furnace based on digital twin, characterized in that The method includes: Select input parameters according to the cement clinker burning process and construct a time-series data set of the precalciner; Preprocess and partition the time-series data set of the precalciner to obtain a training set, a validation set, and a test set, which are used as model inputs; Establish a geometric model of the precalciner to train a prediction model, and establish a deep learning model based on CNN-GRU-Attention to train a digital twin virtual model layer; Perform anomaly detection and temperature field simulation on the precalciner based on the trained digital twin virtual model layer and prediction model, and output the predicted values of the temperature field of the precalciner; The digital twin virtual model layer consists of a digital twin layer and a virtual model layer. Among them, the specific acquisition process of the digital twin layer is as follows: Screen out the tertiary air temperature, kiln tail gas chamber temperature, pulverized coal chemical analysis of the tail coal, and cement raw material chemical analysis from the input parameters; Use the geometric model of the precalciner as a physical model to simulate and generate temperature field data through the physical model; Preprocess the tertiary air temperature, kiln tail gas chamber temperature, pulverized coal chemical analysis of the tail coal, and cement raw material chemical analysis to obtain preprocessed data; Use the preprocessed data as input and construct a temperature field prediction model based on the K-nearest neighbor algorithm; Integrate the temperature field data generated by the physical model and the temperature field prediction data based on the K-nearest neighbor algorithm to construct a data model; Through the data model, map the collected and predicted temperature data to the actual spatial structure of the precalciner to form a temperature field distribution; Based on the mapped temperature field, combine the real-time operation data and historical data of the precalciner to construct a digital twin layer model; The specific acquisition process of the virtual model layer is as follows: Generate model files through the digital twins in the digital twin layer, organize them according to the time series, and divide the continuous time series data into windows of a fixed length according to the requirements of model training. Take the data within each window as an independent input sample; Perform feature enhancement and normalization on the divided time series data, and input the preprocessed input sequence into the CNN layer of the deep learning model based on CNN-GRU-Attention. In the convolutional layer, set convolutional kernels of different sizes and strides; Extract local features in the data through the sliding convolution operation of the convolutional kernel on the input data. The data processed by the convolutional layer is then reduced in dimension through the pooling layer using the max pooling or average pooling method; Convert the input sequence processed by the CNN layer into the form of feature vectors, send it to the GRU layer, and input the feature sequence output by the GRU layer into the Attention layer for attention mechanism calculation; The output of the Attention layer passes through an output layer for the final result output. The output layer adopts a fully connected layer structure. The calculation formula of the output layer is , is the weight matrix of the output layer, represents the predicted output result of the model, represents the feature vector output by the Attention layer, is a constant term.

2. The method for simulating the temperature field of a cement decomposing furnace based on digital twin according to claim 1, characterized in that: The establishment of the geometric model of the precalciner includes: According to the design drawing of the precalciner in the cement plant, establish a 3D model of the precalciner using 3D modeling software; Set the 3D model of the precalciner to have 4 raw material feeding inlets, 4 tail coal coal injection inlets, 2 tertiary air inlets, 1 kiln tail gas chamber outlet, 1 precalciner outlet, and the inner wall inlet and walls of the precalciner according to the cement burning process; The hybrid mesh generation technique is adopted to generate meshes for the internal region, boundary layer, and pre-set analysis region of the precalciner, and the mesh quality is detected by calculating the orthogonal quality index of the meshes. Among them, the hybrid meshes include layered polyhedral meshes, pure polyhedral meshes, and hexahedral meshes; Set the boundary conditions for the raw meal feeding inlet, tertiary coal injection inlet, tertiary air inlet, kiln tail smoke chamber outlet, precalciner outlet, inner wall and wall of the precalciner; Map it to a data model based on the nearest neighbor algorithm of Euclidean distance, and conduct physical simulation inside the precalciner. According to the simulation results, adjust the model parameters, and thus output the geometric model of the precalciner.

3. A method for simulating the temperature field of a cement decomposition furnace based on digital twin according to claim 1, characterized in that: The specific training process of the prediction model is as follows: After completing the construction of the precalciner geometric model, set its initial parameters. Based on the mass conservation equation, momentum conservation equation, and energy conservation equation, in the geometric model, first do not consider the inlet and outlet temperatures, and perform iterative operation according to the set number of iterations until the outlet wind speed of the precalciner reaches the set maximum value; Import the chemical analysis of pulverized coal and the chemical analysis of raw meal, and start to simulate the relevant chemical reaction behaviors of pulverized coal combustion and raw meal decomposition; After completing the chemical reaction simulation, analyze the simulation results and compare them with the actual data. If the difference between the simulation results and the corresponding actual data is greater than the set difference, adjust the model parameters and output the final prediction model.

4. A method for simulating the temperature field of a cement decomposition furnace based on digital twin according to claim 3, characterized in that: The input of the Attention layer has three paths, namely path, path, and path; For the input matrix , each path first maps it through a linear layer to obtain the corresponding query vector , key vector and value vector ; Calculation Take the dot product with , perform a scaling operation, and obtain the result . Then, normalize it using the Softmax function to obtain the attention weights, multiply the attention weights with the value vector and sum them to obtain the output of the Attention layer.

5. A method for simulating the temperature field of a cement precalciner based on digital twin according to claim 3, characterized in that: The training process of the digital twin virtual model layer also includes: Use the training set to train the constructed model based on CNN-GRU-Attention, and calculate the difference between the model prediction result and the true value using the pre-set loss function. During the training process, use the backpropagation algorithm to calculate the gradient of the loss function with respect to the model parameters, and use the optimizer to update the model parameters according to the gradient to make the loss function reach the minimum value; Adopt the learning rate decay strategy to gradually reduce the learning rate as the number of training rounds increases, and use regularization technology to constrain the virtual model layer.

6. The temperature field simulation method of a cement decomposition furnace based on digital twin according to claim 1, wherein: The abnormal detection and temperature field simulation of the precalciner include: Obtain real-time data from the actual operation of the precalciner as the current input data, and perform mean filtering and normalization processing on the current input data; Input the preprocessed data into the trained digital twin virtual model layer for temperature field simulation, and output the temperature field data set inside the precalciner through the digital twin virtual model layer; Determine the steady-state operating point of the virtual model layer, establish a hybrid model using a series-parallel hybrid connection method, and judge whether the predicted value of the input parameters of the precalciner deviates from the normal value through the hybrid model. If it deviates, it is determined as an abnormal condition, and an early warning message is output, and the next step is executed. If it does not deviate, directly execute the next step; Based on the hybrid model, perform temperature field simulation and output the predicted value of the precalciner temperature field.

7. A method for simulating the temperature field of a cement decomposition furnace based on digital twin according to claim 6, characterized in that: The temperature field simulation includes: Geometric model simulation: Based on the 3D geometric model and mesh generation of the precalciner, perform simulations of fluid flow, heat transfer, and chemical reactions; Physical model simulation: Based on the laws of mass conservation, momentum conservation, and energy conservation, calculate the temperature field distribution inside the precalciner; Chemical reaction simulation: Import the chemical analysis of pulverized coal and the chemical analysis of raw meal, and simulate the chemical reaction behavior of pulverized coal combustion and raw meal decomposition.

8. A method for simulating the temperature field of a cement decomposition furnace based on digital twin according to claim 6, characterized in that: The establishment of the hybrid model includes: Connect a data-driven model based on CNN-GRU-Attention behind the temperature field data model of the digital twin layer; Parallelly connect the cascaded model with another data-driven model based on CNN-GRU-Attention to form a series-parallel hybrid model.

Images