Multi-stage preheater working condition prediction method based on eigenvalue orthogonal decomposition and neural network

Abstract

The application discloses a multistage preheater working condition prediction method based on eigen-orthogonal decomposition and a neural network, and belongs to the technical field of multistage preheater digital twinning, wherein the method establishes a reduced-order model for quickly performing multistage preheater smooth working condition prediction through eigen-orthogonal decomposition and a back propagation neural network technology, and couples a filtering correction algorithm to correct the prediction result through sensor measured values, so that the efficiency of flow field solving calculation is greatly improved, the time cost is reduced, the generalization ability is quite high, the method can be actually deployed in a cement plant, and on-site personnel can adjust and optimize working condition parameters according to the preheater flow field.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of multi-stage preheater digital twinning, in particular to a multi-stage preheater working condition prediction method based on proper orthogonal decomposition and neural network. BACKGROUND

[0002] CFD (Computational Fluid Dynamics) simulation is currently widely used in modern engineering design and scientific research, such as aerospace, automobile design, biomedical, environmental engineering and many other fields. Through high-precision numerical simulation, CFD can help researchers deeply understand the complex behavior of fluid, optimize product design, reduce the number of experiments, and thus accelerate the research and development process. However, due to the complexity of fluid dynamics itself and the large number of grids divided to achieve simulation accuracy, CFD simulation needs to rely on high-performance computers. This not only means high investment cost of hardware equipment, but also significantly increases the overall cost of the project due to long computing time. Especially in dealing with multi-stage preheater simulation tasks of large scale, high precision and complex turbulent flow characteristics, CFD simulation calculation time may be as long as several days or even weeks, which cannot quickly calculate and respond to the complex and variable working conditions of actual projects.

[0003] In order to improve the calculation speed of numerical simulation and realize real-time prediction of full flow field data, the multi-stage preheater simulation model must be reduced, and a digital twin model with sufficient accuracy and consideration of computational efficiency must be established. The POD (Proper Orthogonal Decomposition) method can capture the characteristics of the flow field, greatly reducing the complexity of the flow field. By combining POD and RBF (Radial Basis Function), Kriging and other surrogate models, the prediction of steady flow field and unsteady flow field can be realized, but these surrogate models are difficult to handle multiple input and output parameters, and the modeling efficiency is low. The method of directly establishing a reduced-order model based on data-driven deep learning of flow field data has the disadvantages of large consumption of computing resources, complex modeling and poor generalization ability. How to accurately predict and represent the flow field under a specified working condition by using the characteristics obtained by POD analysis is a key technical problem in realizing the reduction and digital twinning of the multi-stage preheater model. SUMMARY

[0004] The purpose of the present application is to provide a multi-stage preheater working condition prediction method based on proper orthogonal decomposition and neural network, which can realize real-time and rapid response of multi-stage preheater working condition data, greatly improve the calculation speed and reduce the time cost.

[0005] The purpose of the present application can be achieved by the following technical solutions:

[0006] A multi-stage preheater working condition prediction method based on proper orthogonal decomposition and neural network, comprising the following steps:

[0007] SS1, numerical simulation: Euler multiphase flow model is used to simulate the gas-solid two-phase flow behavior of the multi-stage preheater, and the accuracy is verified by comparing with the engineering data;

[0008] SS2, obtain sample space: considering the preheater flow field law under different working condition parameters as the sample space;

[0009] The working condition parameters include the inclination angle of the material distribution box, the system wind speed and the particle size distribution;

[0010] SS3, eigenvalue decomposition: through the POD orthogonal decomposition method, based on the sample variance maximization theory, all the base mode flow fields of the multi-stage preheater under the whole working condition space and the corresponding eigenvalues are obtained;

[0011] SS4, train neural network: build a neural network, train the multi-input and output mapping model between the working condition parameter space and the flow field base mode eigenvalue coefficient;

[0012] SS5, model deployment and flow field prediction;

[0013] SS6, sensor deployment and data correction: through the sensor detection at the local point of the preheater, the local observation data are obtained, and the predicted data are updated in the whole field based on the filtering correction algorithm.

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

[0015] 1. The present application establishes a reduced-order model for quickly predicting the smooth working condition of the multi-stage preheater through proper orthogonal decomposition and back propagation neural network technology, and couples a filtering correction algorithm to correct the prediction results through the measured values of the sensors, which not only greatly improves the efficiency of flow field solving calculation and reduces the time cost, but also has considerable generalization ability, so that it can be actually deployed in a cement plant, and the on-site personnel can adjust and optimize the working condition parameters according to the preheater flow field situation.

[0016] 2. The present application can realize rapid prediction and millisecond-level response: unlike the traditional CFD simulation which needs to calculate a large number of mathematical and physical equations, the POD-BPNN model greatly improves the efficiency of flow field calculation by reducing the order of the model, and can realize millisecond-level working condition parameter input-flow field data output;

[0017] 3. The present application can realize accurate prediction under complex working condition parameters: POD can map high-dimensional flow field data to low-dimensional orthogonal base mode space, effectively remove the redundant features of the data, and thus effectively improve the model generalization ability and have the accurate prediction ability under complex input parameters;

[0018] 4、The training and prediction of the model are completely based on data driving, without migration cost, and the model has small calculation cost, is convenient to deploy on an industrial control platform, and provides a necessary condition for system-level digital twinning;

[0019] 5、The algorithm correction based on measured data: the prediction data is updated by a filtering correction algorithm, so that the prediction data is closer to the real physical field. BRIEF DESCRIPTION OF DRAWINGS

[0020] The application will be further described below with reference to the accompanying drawings.

[0021] Figure 1 is a multi-stage preheater simulation cloud picture;

[0022] Figure 2 is a BPNN neural network architecture diagram;

[0023] Figure 3 is a POD-BPNN preheater flow field prediction flowchart;

[0024] Figure 4 is a prediction value filtering correction algorithm. DETAILED DESCRIPTION

[0025] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of the application.

[0026] The core of the application is to use a neural network as a proxy model to build an approximate model from operating condition variables to flow field data, and to realize the establishment of a data-driven multi-stage preheater reduced-order model. Through this method, we only need to perform steady-state flow field simulation of the multi-stage preheater under dozens of different operating conditions through CFD software, and then perform proper orthogonal decomposition on the flow field data to obtain all the modal flow fields of the multi-stage preheater in the entire operating condition space and the corresponding characteristic values. Then, a multi-layer neural network model of operating condition parameters to modal coefficients is established through a back propagation neural network (BPNN), and fast flow field prediction is realized.

[0027] Specifically, as shown in Figure 3 The multi-stage preheater operating condition prediction method based on proper orthogonal decomposition and neural network includes the following steps:

[0028] SS1, multi-stage preheater CFD simulation

[0029] As Figure 1 shown, the cyclone is an important gas-solid separation equipment in the cement industry, where the gas flow is turbulent and exhibits strong non-uniformity and anisotropy, resulting in extremely complex flow process. The multi-stage preheater is a series of cyclones, and the upper cyclone is connected with the lower cyclone by a weight valve, which presents a special boundary that only transmits material but not gas, and the simulation is relatively more difficult. The method uses the Eulerian multiphase flow model in FLUENT, and the particle motion behavior is considered as a fluid. The key point of this method is to describe the viscosity and pressure of the solid phase motion, which is generally modeled by introducing the Kinetic Theory of Granular Flow (KTGF) theory. This method only needs one set of calculation grid, and each grid is about several or tens of times the particle size, which can greatly reduce the calculation amount, better predict the gas-solid flow in large-scale industrial equipment, and reflect the internal flow field law.

[0030] SS2, set up multi-parameter input to obtain sample space

[0031] Using the Designpoint function in ANSYS, the simulation model of the 6-stage preheater with different input parameters is submitted in batches to explore the influence of different parameters such as the inclination angle of the distribution box, the system wind speed, and the particle size on the flow field of the preheater system.

[0032] SS3, intrinsic orthogonal decomposition of sample data

[0033] 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 and map it to a low-dimensional orthogonal basis mode space, so as to analyze the main characteristics of the flow field and the corresponding basis mode coefficients. In essence, it is to maximize the sample variance in each dimension after reducing the flow field to low dimension.

[0034] First, the simulation calculated sample (flow field data) needs to be standardized. Let the original data be x i , i = 1, 2, 3, … r, the sample data x i is an n-dimensional vector (n is related to the number of grids divided in CFD calculation), and r is the number of samples.

[0035]

[0036] 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 sample data characteristics and avoid the influence of individual discrete values on the analysis results.

[0037] From this, the covariance matrix of the standardized data can be obtained:

[0038]

[0039] 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 of different basis modes in the total variance, so as to ensure that the characteristic components contained in the basis modes account for more than 95% of the entire sample space.

[0040] Then the original sample data can be approximately expressed as:

[0041] X = U λ ξ (4)

[0042] 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.

[0043] SS4, constructing a neural network (BPNN) model

[0044] The reduced eigenvalue matrix corresponding to the preheater flow field under different input parameters cannot be estimated by humans, and the back propagation neural network with multiple input parameters and multiple output results has the characteristics of strong nonlinear mapping ability, self-learning weight, and good generalization ability, and is suitable as a proxy model to train the mapping relationship between the input parameters and the POD basis coefficient matrix.

[0045] Since the sample data obtained by CFD simulation of the multi-stage preheater is relatively small (usually tens or hundreds of groups, which belongs to small sample data), a shallow neural network has sufficient nonlinear mapping ability and generalization ability, and a BPNN model with two hidden layers can be selected, which takes into account stability and feature description ability. The neural network structure is as shown in Figure 2 .

[0046] The neural network establishes a mapping from k-dimensional input to m-dimensional output, where the input is x = [x1, x2, …, x k ] T , k corresponds to the number of training sample input parameters, the number of hidden layer neurons is p and q respectively, the output is y = [y1, y2, …, y m ] T , and m corresponds to the number of POD basis modes after dimension reduction. ω 1 , ω 2 , ω 3 are the weight matrices between layers, and their sizes are p x k, q x p and m x q respectively.

[0047] The BPNN model updates the weight matrix by backpropagating the error and increases the non-linear mapping ability and stability of the model by different activation functions, thereby completing the training of the model. The specific training steps are as follows:

[0048] 1) Divide the samples into training set and validation set

[0049] 2) Initialize the weight matrix with normal distribution to avoid gradient disappearance or gradient explosion in the initial training stage;

[0050] 3) Calculate the input and output of the first hidden layer, and use the "Swish" activation function to increase the non-linear expression ability of the model;

[0051] 4) Calculate the input and output of the second hidden layer, and use the "Leaky Relu" activation function to accelerate the training convergence, while avoiding the problems of gradient disappearance and neuron saturation;

[0052] 5) Calculate the backpropagation error of each layer and update the weight matrix;

[0053] 6) Calculate the prediction error of the validation set to evaluate the reasonable training rounds until the model has sufficient calculation accuracy and generalization ability.

[0054] SS5, Model deployment and flow field prediction

[0055] The reduced order model generated by the 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, it can be deployed on factory industrial control equipment. For a new set of factory actual parameter combination j , the BPNN model can quickly predict the base modal characteristic value coefficient Based on POD, the predicted flow field prediction value can be obtained, realizing millisecond-level input and output.

[0056] SS6, Sensor deployment and prediction data correction

[0057] The actual production of the factory is different, and the prediction data inevitably deviates from the true value. Therefore, a filtering algorithm is used to optimally estimate the system state based on the system input and output observation data and prediction data, so that the prediction data is closer to the true physical field.

[0058] As shown in Figure 4 , the method uses a filtering algorithm based on distance correction: suppose the factory is equipped with n sensors, the measured value of each sensor is i (i=1, 2, 3…n), and the flow field prediction value at the same position is Then for any data point xk The modified value is:

[0059]

[0060] wherein D i is the distance from each sensor position, and R is the maximum radius of each cyclone of the preheater. k The algorithm systematically considers the correction of the prediction data by all measurement points, can effectively improve the accuracy of the flow field prediction value on the basis of preserving the true characteristics of the flow field, and is consistent with the real working condition.

[0061] After the corrected prediction value is calculated, the on-site personnel can evaluate the overall situation of the preheater in real time, and adjust the process parameters according to experience and in a targeted manner to optimize resource utilization.

[0062] The present application simulates the flow field of the multi-stage preheater 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 multi-stage preheater, solve the problems of high cost and long time of CFD simulation calculation, and couple the filtering correction algorithm to connect 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 multi-stage preheater reduced-order model in the factory and guide the optimization of production.

[0063] 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 the modifications or supplements do not deviate from the invention or exceed the scope defined by the present claims, and should belong to the protection scope of the present application.

Claims

1. A method for predicting the operating conditions of a multi-stage preheater based on intrinsic orthogonal decomposition and neural networks, characterized in that, Includes the following steps: SS1. Numerical simulation: Simulate the gas-solid two-phase flow behavior of a multi-stage preheater and compare the results with engineering data to verify the accuracy. SS2. Obtaining the sample space: The flow field characteristics of the preheater under different operating parameters are considered as the sample space. The operating parameters include the tilt angle of the spreading box, the system wind speed, and the particle size distribution. SS3, Eigenvalue Decomposition: Using the POD orthogonal decomposition method, based on the sample variance maximization theory, we obtain all the fundamental modes of flow field and their corresponding eigenvalues ​​of the multi-stage preheater under the entire operating space. SS4, Training Neural Network: 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; SS5, Model Deployment and Flow Field Prediction; SS6, Sensor Deployment and Data Correction: Local observation data is obtained by performing sensor detection at local points in the preheater, and the predicted data is updated across the entire field based on a filtering correction algorithm; Step SS6 employs a distance-based correction filtering algorithm: Assume the factory has n sensors deployed, and the measured value of each sensor is T. i The predicted flow field value at the same location is If i = 1, 2, 3…n, then for any data point x in the predicted flow field… k Its corrected value is: Among them, D i For data point x k The distance from each sensor location, R, is the maximum radius of each stage of the preheater's cyclone separator.

2. The method for predicting the operating conditions of a multi-stage preheater based on intrinsic orthogonal decomposition and neural networks according to claim 1, characterized in that, Step SS3 includes: Standardize the flow field data samples calculated from the simulation; The covariance matrix of the standardized data can be calculated: To obtain the sample mean, To subtract the mean from the original data, we solve for the eigenvalues ​​of the n×n covariance matrix. The first m eigenvalues ​​can be denoted as λ. 1 , λ 2 、…、λ m The corresponding fundamental mode eigenvectors can be denoted as ξ. 1 ξ 2 、…、ξ m The value of m is determined based on the proportion of the variance of different basic modes to the total variance, ensuring that the feature components contained in the basic modes account for more than 95% of the entire sample space. Therefore, the original sample data can be approximately represented as: X = U λ ξ.

3. The method for predicting the operating conditions of a multi-stage preheater based on intrinsic orthogonal decomposition and neural networks according to claim 1, characterized in that, In step SS4, the neural network uses a BPNN model with two hidden layers to establish a mapping from k-dimensional input to m-dimensional output, where the input is x = [x1, x2, ..., xm]. k ] T k corresponds to the number of input parameters in the training samples, p and q are the number of neurons in the hidden layer, and the output is y = [y1, y2, ..., y]. m ] T m corresponds to the number of POD basic modes after dimensionality reduction; ω 1 ω 2 ω 3 These are the weight matrices between the layers, with sizes of p×k, q×p, and m×q, respectively.

4. The method for predicting the operating conditions of a multi-stage preheater based on intrinsic orthogonal decomposition and neural networks according to claim 1, characterized in that, The specific training steps for the neural network described in step SS4 are as follows: The samples are divided into a training set and a validation set; The weight matrix is ​​initialized using a normal distribution; The input and output of the first hidden layer are calculated, and the "Swish" 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.

5. The method for predicting the operating conditions of a multi-stage preheater based on intrinsic orthogonal decomposition and neural networks according to claim 2, characterized in that, In step SS5, for a new set of actual plant parameter combinations Z j The eigenvalue coefficients of the fundamental modes are predicted using the BPNN model. By performing reverse reconstruction based on the POD, the predicted flow field values ​​can be obtained.

Images