A method for predicting the working condition of a cooling bed combining proper orthogonal decomposition and deep learning

Abstract

The application discloses a method for predicting the working condition of a cooling grate by combining intrinsic orthogonal decomposition and deep learning, and belongs to the technical field of digital twinning. The method simulates the flow field of the cooling grate by CFD technology, obtains a large amount of sample data, and then combines the intrinsic orthogonal decomposition method with a multilayer back propagation neural network, so that the flow field of the cooling grate can be quickly predicted, and the problems of high cost and long time of CFD simulation calculation are solved. Meanwhile, a filter correction algorithm is coupled to connect the predicted data with the actual sensor detection values, so that the error between the simulation prediction values and the field detection values can be solved, and the actual deployment of the cooling grate reduction and the digital twinning model in a factory and the guidance of optimizing production can be realized.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of digital twinning, and particularly relates to a grate cooler working condition prediction method combining proper orthogonal decomposition and deep learning. BACKGROUND

[0002] The grate cooler is a large cooling and heat exchange equipment in the cement industry. In the cement production process, the grate cooler is a key equipment for high-temperature clinker cooling and heat recovery. Its stable and efficient operation is of great significance for cement enterprises to improve clinker quality, reduce consumption and save energy. The heat exchange between the clinker and the cooling air inside the grate cooler directly affects the cement quality and the energy consumption of the clinker firing system. In order to improve the step utilization efficiency of heat in the grate cooler, it is necessary to fully understand the temperature field and flow field distribution inside the grate cooler. The grate cooler is essentially a large heat exchanger. High-temperature clinker and low-temperature cooling air exchange heat inside. Its main function is to cool the high-temperature clinker discharged from the rotary kiln, recover heat, and supply secondary air and tertiary air to the rotary kiln and the decomposing furnace respectively to fully utilize the heat in the grate cooler and improve the thermal efficiency of the entire firing system and the clinker quality.

[0003] In the actual application of cement production, the grate cooler faces problems such as equipment wear caused by long-term operation, variability of working conditions, and improper maintenance. The current technical level limits the deployment of thermocouples and other monitoring devices at every key position of the grate cooler to accurately measure the temperature of the clinker in real time. Many domestic cement enterprises still mainly rely on the experience and intuition of workers when adjusting the key operating parameters of the grate cooler. This approach lacks scientific theoretical support and has obvious subjectivity and response lag problems. This unstable control level can easily lead to "undercooling" or "overcooling" of the clinker during the cooling process, affecting the quality of the clinker. The internal structure of the grate cooler and the heat exchange situation are relatively complex, and the clinker temperature is high. Therefore, it is difficult to directly obtain the temperature distribution and flow situation inside the grate cooler.

[0004] In order to improve the heat utilization efficiency of the grate cooler, it is necessary to fully understand the temperature field and flow field distribution inside the grate cooler. In order to improve the calculation speed of numerical simulation and realize real-time prediction of full flow field data, the simulation model of the grate cooler must be reduced in order to establish a digital twinning model with sufficient precision and consideration of calculation efficiency. How to accurately predict and represent the flow field under a specified working condition using the characteristics obtained by POD (Proper Orthogonal Decomposition) is a key technical problem in realizing the reduction of the grate cooler model and digital twinning. SUMMARY

[0005] The purpose of the present application is to provide a grate cooler working condition prediction method combining proper orthogonal decomposition and deep learning, solving the problem that the internal structure and heat exchange of the grate cooler are relatively complex, the clinker temperature is relatively high, and the temperature distribution and flow conditions in the grate cooler are difficult to obtain directly in the prior art.

[0006] The purpose of the present application can be achieved by the following technical solutions:

[0007] A grate cooler working condition prediction method combining proper orthogonal decomposition and deep learning comprises the following steps:

[0008] S1, a numerical simulation model is established. Based on CFD simulation software such as Fluent, the gas-solid two-phase flow behavior of the grate cooler is simulated based on the porous medium model, the boundary conditions, material parameters, particle parameters, etc. are set according to the actual working parameters of the grate cooler, and the calculation is submitted to realize the transient simulation of the grate cooler under different wind speed conditions. The model can achieve: (1) heat exchange simulation between high-temperature clinker particles and cooling air; (2) simulation analysis of internal flow field parameters of the grate cooler.

[0009] S2, simulation result verification and adjustment. Based on the actual production data of the factory, the simulation results are verified. By comparing the simulation results and the actual production data, the simulation model is adjusted, the accuracy of the simulation parameters and the model building is verified, and the accuracy and reliability of the model are improved.

[0010] S3, output simulation sample data. After the simulation model is verified, the grate cooler simulation model with different input parameters is submitted in batches to explore the influence of different parameters on the flow field and temperature field of the grate cooler system, so as to obtain different sample spaces.

[0011] S4, proper orthogonal decomposition of sample data. Through POD decomposition, most of the information of the original sample can be represented by the basis mode coefficient matrix and a small number of POD basis modes, so that the high-dimensional flow field data can be reduced and mapped to the low-dimensional orthogonal basis mode space, so that the main features of the flow field temperature field can be analyzed.

[0012] S5, model deployment and flow field prediction. The basis mode eigenvalue coefficient is quickly predicted through the BPNN model, and then the reverse reconstruction is carried out based on POD, so that the flow field temperature field prediction value can be obtained, and the input and output can be realized in milliseconds.

[0013] S6 sensor deployment and prediction data correction. In the actual production process of the factory, due to the interweaving of various complex factors, there is inevitably deviation between the prediction data and the real situation on the spot, the simulation results and the actual measurement data are fused by using a filtering algorithm, the data of the local observation point is combined with the full-field prediction data, the full-field flow field prediction data is updated, and the full-field flow field prediction data is closer to the real situation of the actual physical field. After the measured data is corrected, the prediction flow field and temperature field data of the grate cooler are obtained, the internal flow field of the grate cooler is mastered, and these data are used as important basis for optimization decision support and operation condition adjustment, so that the grate cooler can run in the optimal condition.

[0014] The beneficial effects of the present application are:

[0015] 1. The present application simulates the flow field of the grate cooler by CFD technology, obtains a large amount of sample data, and then combines the intrinsic orthogonal decomposition method with the multilayer back propagation neural network, so as to realize the rapid prediction of the flow field of the grate cooler, solve the problems of high cost and long time of CFD simulation calculation, and couple the filtering correction algorithm to the prediction data and the actual sensor detection value, solve the error problem between the simulation prediction value and the on-site detection value, realize the actual deployment of the grate cooler reduction and digital twin model in the factory, and guide the optimization of production.

[0016] 2. The present application combines the POD data reduction method and the BPNN neural network model to build a digital twin model of the grate cooler, effectively removes the redundant features of the data, greatly improves the efficiency of the calculation of the flow field temperature field, can realize millisecond-level working condition parameter input and flow field data output, and makes the model have sufficient calculation precision and generalization ability.

[0017] 3. The present application integrates and corrects the real-time monitoring data and the prediction data of the BPNN neural network through sensor deployment, can obtain the flow field and temperature field data of the grate cooler corrected based on the measured data, makes the prediction data closer to the real physical field, serves as important basis for optimization decision support and operation condition adjustment, provides optimization suggestions and early warning information. BRIEF DESCRIPTION OF DRAWINGS

[0018] The present application will be further described below in combination with the drawings.

[0019] Figure 1 is a flow diagram of a grate cooler working condition prediction method combining intrinsic orthogonal decomposition and deep learning of the present application;

[0020] Figure 2 is a neural network structure diagram;

[0021] Figure 3 is a flow diagram of model training. DETAILED DESCRIPTION

[0022] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative effort fall within the protection scope of the present application.

[0023] The core of the present application is to use a neural network as a surrogate model to build an approximate model from operating condition variables to flow field data, and to realize the establishment of a data-driven grate cooler reduced-order model and a digital twin method. Through this method, only a few dozen sets of grate cooler flow field temperature field simulations under different operating conditions are needed through CFD software, and then through proper orthogonal decomposition of sample data, all basis modes and corresponding eigenvalues of the grate cooler under the entire operating condition space can be obtained, and a multilayer neural network model of operating condition parameters to basis mode coefficients is established through a back propagation neural network to realize fast prediction of the flow field. The present application simulates the flow field of the grate cooler through CFD technology to obtain a large amount of sample data, and then combines the proper orthogonal decomposition method and the multilayer back propagation neural network to realize fast prediction of the flow field of the grate cooler.

[0024] Specifically, as shown in Figure 1 The grate cooler operating condition prediction method combining proper orthogonal decomposition and deep learning provided by the present application comprises the following steps:

[0025] S1, CFD simulation of the grate cooler. The numerical simulation module is based on a porous medium model to establish a transient simulation model of the grate cooler to simulate the gas-solid two-phase flow behavior of the grate cooler. The simulation model takes into account factors such as the structural characteristics and material characteristics of the grate cooler. Through averaging and statistical processing, the accumulated cement clinker layer can be approximately processed as a large-porosity porous medium, the flow and heat exchange of cooling air between high-temperature cement clinker is similar to the flow and heat transfer of gas flowing between porous media, the non-thermodynamic equilibrium heat exchange between cooling air and clinker is realized, the gas-solid flow in a large-scale industrial equipment is better predicted, the internal flow field law is reflected, and the accuracy is verified by comparing with engineering data.

[0026] Based on the Representative Elementary Volume (REV) assumption and combined with the volume average method, the expression of the macroscopic continuity equation of the porous medium is as follows

[0027]

[0028] The porous medium model in Fluent is to add an empirical assumption-based flow direction resistance to the defined porous medium area, which is equivalent to adding a momentum source term to the standard momentum equation, which can be expressed as

[0029]

[0030] where S i represents the source term of momentum equation; i represents x, y, z three directions; the first and second terms on the right side of the formula represent the viscous resistance term and the inertial resistance term respectively; D, C are the viscous resistance coefficient matrix and the inertial resistance coefficient matrix respectively; μ, ρ represent the viscosity and density of the fluid respectively; is the velocity vector of the fluid in j direction and the velocity size.

[0031] The fluid energy equation of the heat transfer model inside the porous medium is as follows:

[0032]

[0033] The solid energy equation is as follows:

[0034]

[0035] where T f represents the average temperature of the fluid, T s represents the average temperature of the solid, h v represents the unit volume convective heat transfer coefficient between the solid phase and the fluid of the porous medium, ε represents the porosity, λ s , λ f represent the thermal conductivity of the solid and the fluid respectively, q f , q s represent the heat source in the fluid and the heat source in the solid respectively.

[0036] S2, set up a multi-parameter input sample space. Use the Design point function in ANSYS to submit the grate cooler simulation model with different input parameters in batches, explore the influence of different parameters such as cooling air flow rate, cooling air temperature, different particle sizes of clinker particles and porosity of clinker particles on the flow field of the grate cooler system, and obtain different sample spaces.

[0037] S3, intrinsic orthogonal decomposition of sample data. Through the POD orthogonal decomposition method, based on the sample variance maximization theory, all the base mode flow fields of the grate cooler under the whole working condition space and the corresponding characteristic values are obtained. Intrinsic orthogonal decomposition or principal component analysis is a method of vector data statistical analysis, which can reduce the order of high-dimensional flow field data, map it to a low-dimensional orthogonal base mode space, so as to analyze the main characteristics of the flow field and the corresponding base mode coefficients. In essence, it is to maximize the sample variance in each dimension after reducing the flow field to low dimension. First, the sample (flow field data) calculated by simulation needs to be standardized. Let the original data be x i , i = 1, 2, 3, …, r, and the sample data x iis the n-dimensional vector (n is related to the number of mesh divided in CFD calculation), r is the number of samples.

[0038]

[0039] Equation (1) is used to obtain the mean value of the sample, and equation (2) is used to subtract the mean value from the original data. Data standardization is beneficial to eliminate the differences between the characteristics of sample data and avoid the influence of individual discrete values on the analysis results. Thus, the covariance matrix of the standardized data can be obtained:

[0040]

[0041] By solving the eigenvalues of the nXn order covariance matrix, the first m order eigenvalues can be denoted as λ 1 , λ 2 , …, λ m , and the corresponding basis modal eigenvectors can be denoted as ξ 1 , ξ 2 , …, ξ m . The value of m is determined according to the proportion of the variance value of different basis modes in the total variance value, so as to ensure that the characteristic components contained in the basis modes account for more than 95% of the entire sample space. Then, the original sample data can be approximately represented as:

[0042] X=U λ ξ(8)

[0043] In this way, the basis modal coefficient matrix and a small number of POD basis modes can be used to represent most of the information of the original sample.

[0044] S4, training BPNN neural network. The neural network (BPNN) is built, and the multi-input and output mapping model between the working condition parameter space and the flow field basis modal eigenvalue coefficient is trained to achieve a good fitting effect. The reduced eigenvalue matrix corresponding to the flow field of the cooling grate under different input parameters cannot be estimated artificially, and the back propagation neural network with multiple input parameters and multiple output results has the characteristics of strong non-linear mapping ability, self-learning weight and good generalization ability, and is suitable for training the mapping relationship between the input parameters and the POD basis coefficient matrix. For the sample data obtained by CFD simulation of the cooling grate, the sample data is relatively small (generally several tens or hundreds, which belongs to small sample data), and the shallow neural network has enough non-linear mapping ability and generalization ability. The BPNN model with two hidden layers can be selected, and the stability and feature description ability are considered. The neural network structure is shown in Figure 2 .

[0045] The neural network establishes a mapping from r-dimensional input to m-dimensional output, where the input is, k corresponds to the number of training sample input parameters, the number of hidden layer neurons is p and q respectively, and the output is y=[y1,y2,…,ym ] T , m is the number of POD basis modes after dimension reduction. ω1, ω2, ω3 are the weight matrices between layers, whose sizes are ×k, q × p and m × q respectively. BPNN model updates the weight matrix by backpropagating the error, and uses different activation functions to increase the non-linear mapping ability and stability of the model, so as to complete the training of the model, as shown in FIG. 3, and the specific training steps are as follows: Figure 3

[0046] 1) Divide the samples into training set and validation set for training and result verification;

[0047] 2) Use orthogonal initialization weight matrix, and weight initialization affects the training effect and convergence speed of neural network. By initializing the weight matrix as an orthogonal matrix, the information transmission is maintained and the gradient vanishing problem in training is reduced. The weight matrix W is orthogonal, that is, T W = I;

[0048] 3) Calculate the input and output of the first hidden layer, and use the "Sigmoid" activation function to increase the non-linear expression ability of the model, smooth the gradient, and avoid the jumping output value;

[0049] 4) Calculate the input and output of the second hidden layer, and use the "Leaky Relu" activation function to accelerate the training convergence, solve the problem that if the input is negative, the gradient is completely zero in the backpropagation process, that is, the gradient vanishing problem, and also avoid the problem of neuron saturation;

[0050] 5) Calculate the reverse error of each layer and update the weight matrix;

[0051] 6) Calculate the prediction error of the validation set to evaluate the reasonable training rounds until the model has sufficient calculation precision and generalization ability.

[0052] S5, model deployment and flow field prediction. The reduced order model generated by this method has a relatively small volume, fast calculation, low memory occupation, and does not depend on foreign CFD software environment. After configuring a small amount of open source Python library, the digital twin model can be deployed on the factory industrial control device. For a new set of factory actual parameter combination Z j , the BPNN model can quickly predict the basis mode eigenvalue coefficient and then based on POD, the predicted flow field prediction value can be obtained, which realizes millisecond-level input and output.

[0053] ​S6, sensor deployment and data correction. By detecting the sensor at the local point of the grate cooler, the local observation data is obtained, and the prediction data is updated by the correction based on the distance and the filtering correction algorithm based on the Gaussian function. Assuming that n sensors are arranged, the measured value of each sensor is T i (i = 1, 2, 3…n), and the predicted value of the same position flow field is Then for any data point x of the predicted flow field k , the corrected value is:

[0054]

[0055] Where d i is the distance between the data point x k and each sensor position; σ is the correction coefficient, which is related to the specific structural parameters of the grate cooler, and is fitted according to the specific working conditions of different grate coolers. This algorithm systematically considers the correction of all sensor measurement points to the predicted data, which can effectively improve the accuracy of the predicted flow field value on the basis of preserving the true characteristics of the flow field, so that the predicted data is closer to the real physical field data. After the measured data is corrected, the predicted flow field and temperature field data of the grate cooler are obtained, the internal flow field of the grate cooler is mastered, and these data are used as important basis for optimization decision support and operation condition adjustment, so that the grate cooler can be operated under the optimal condition.

[0056] The above is only an example and description of the present application, and those skilled in the art can make various modifications or supplements to the described specific embodiments or use similar ways to replace, as long as they do not deviate from the invention or exceed the scope defined by the present claims, which shall belong to the protection scope of the present application.

Claims

1. A method for predicting the operating conditions of a grate cooler by combining intrinsic orthogonal decomposition and deep learning, characterized in that, Includes the following steps: S1. Establish a numerical simulation model: Simulate the gas-solid two-phase flow behavior of the grate cooler, submit calculations according to the actual working parameters of the grate cooler, and realize the transient simulation of the grate cooler under different wind speed conditions. S2. Simulation Result Verification and Adjustment: By comparing simulation results with actual production data, adjust the simulation model and verify the accuracy of simulation parameters and model construction. S3. Output simulation sample data: Batch submission of grate cooler simulation models with different input parameters to explore the influence of different parameters on the flow field and temperature field of the grate cooler system in order to obtain different sample spaces; S4. Intrinsic orthogonal decomposition of sample data: Using the POD orthogonal decomposition method, based on the sample variance maximization theory, we obtain all the basic modes of flow field and their corresponding eigenvalues ​​of the grate cooler in the entire operating space. S5. Build a neural network and train a multi-input output mapping model between the working condition parameter space and the eigenvalue coefficients of the flow field basic modes to achieve a better fitting effect. S6. Model Deployment and Flow Field Prediction; S7. Sensor Deployment and Data Correction: By performing sensor detection at local points in the grate cooler, local observation data is obtained, and the predicted data is updated across the entire field using distance-based correction and Gaussian function-based filtering correction algorithms. In step S7, it is assumed that n sensors are deployed, and the measured value of each sensor is... (i=1,2,3…n), the predicted flow field value at the same location is Then for any data point in the predicted flow field... Its corrected value is: in, For data points Distance from each sensor location; This is a correction factor.

2. The method for predicting the operating conditions of a grate cooler combining intrinsic orthogonal decomposition and deep learning according to claim 1, characterized in that, In step S1, a numerical simulation model is established based on a porous medium model.

3. The method for predicting the operating conditions of a grate cooler combining intrinsic orthogonal decomposition and deep learning according to claim 2, characterized in that, Based on the characterization of the volume element assumption and combined with the volume averaging method, the expression for the macroscopic continuity equation of the porous medium is as follows: in, Indicates the density of the fluid. Indicates porosity; The porous media model in Fluent adds a flow-direction drag, based primarily on empirical assumptions, to the defined porous media region. This is equivalent to adding a momentum source term to the standard momentum equation, expressed as: In the formula, The source term of the momentum equation is represented by ; i represents the three directions x, y, and z; the first and second terms on the right-hand side of the equation represent the viscous drag term and the inertial drag term, respectively; D and C are the matrices of the viscous drag coefficient and the inertial drag coefficient, respectively. Indicates the viscosity of the fluid; Let j be the velocity vector and magnitude of the fluid in the j-direction; The heat transfer model inside a porous medium has the following fluid energy equation: ; The solid-state energy equation is as follows: In the formula, Indicates the average temperature of the fluid. Indicates the average temperature of a solid. This represents the convective heat transfer coefficient per unit volume between the solid phase and the fluid in a porous medium. These represent the thermal conductivity of the solid and the fluid, respectively. They represent internal heat sources in fluids and internal heat sources in solids, respectively.

4. The method for predicting the operating conditions of a grate cooler combining intrinsic orthogonal decomposition and deep learning according to claim 1, characterized in that, The neural network selected is a BPNN model with two hidden layers. This neural network establishes a mapping from k-dimensional input to m-dimensional output, where the input is... , k corresponds to the number of input parameters in the training samples, the number of hidden layer neurons are p and q respectively, and the output is m corresponds to the number of POD basic modes after dimensionality reduction; These are the weight matrices between the layers, with sizes of p×k, q×p, and m×q, respectively.

5. The method for predicting the operating conditions of a grate cooler combining intrinsic orthogonal decomposition and deep learning according to claim 4, characterized in that, The specific training steps for the neural network model are as follows: The samples are divided into a training set and a validation set; Information transmission is maintained by initializing the weight matrix as an orthogonal matrix; The input and output of the first hidden layer are calculated, and the "Sigmoid" activation function is used to increase the non-linear expressive power of the model; Calculate the input and output of the second hidden layer, and use the "Leaky ReLU" activation function to accelerate training convergence; Calculate the back-inversion error of each layer and update the weight matrix; Calculate the prediction error on the validation set to evaluate the appropriate number of training epochs until the model has sufficient computational accuracy and generalization ability.

Images