CN-122020121-A - E-PINN-based sintering ignition temperature prediction method

CN122020121ACN 122020121 ACN122020121 ACN 122020121ACN-122020121-A

Abstract

The invention discloses a sintering ignition temperature prediction method based on E-PINN, and belongs to the technical field of sintering control. The method comprises the steps of firstly measuring operation parameters of an ignition furnace to form input features and output features, processing data of the input features and the output features and dividing the data into a training set and a testing set, secondly pre-training the training set by adopting ridge regression with L2 regularization to obtain regression coefficients and bias coefficients of a linear physical core model, and accordingly establishing the linear physical core model, thirdly, introducing a residual network to compensate the linear physical core model to obtain an ignition temperature prediction model. The invention can unify the data fitting error and physical consistency constraint into a training target, and unifies the prediction precision, stability and physical consistency under complex working conditions.

Inventors

CHEN LIANGJUN
SHAO SHIHUA
BI JING
YU ZHENGWEI
LONG HONGMING
WANG GUANGYING
TAN KANGKANG
LU WEIWEN
Ning Houyin

Assignees

安徽工业大学

Dates

Publication Date: 20260512
Application Date: 20260130

Claims (10)

1. The sintering ignition temperature prediction method based on E-PINN is characterized by comprising the following steps of: 1. measuring operation parameters of the ignition furnace, forming input features and output features, processing data of the input features and the output features, and dividing the data into a training set and a testing set; 2. Pre-training the training set by adopting ridge regression with L2 regularization to obtain regression coefficients and bias coefficients of the linear physical kernel model, and establishing the linear physical kernel model according to the regression coefficients and bias coefficients; 3. And introducing a residual error network to compensate the linear physical core model to obtain an ignition temperature prediction model.
2. The method for predicting sintering ignition temperature based on E-PINN as set forth in claim 1, wherein in the first step, the operating parameters of the ignition furnace are gas flow, combustion air flow, furnace negative pressure, preheated air temperature, air-fuel ratio and ignition furnace outlet temperature, the input features include gas flow, combustion air flow, furnace negative pressure, preheated air temperature and air-fuel ratio, and the output features are ignition furnace outlet temperature.
3. The method of claim 2, wherein in the first step, the data processing includes data cleaning and feature normalization.
4. The E-PINN-based sintering ignition temperature prediction method according to claim 3 is characterized in that the data cleaning process comprises the steps of adopting sliding window median filter suppression for pulse spike, adopting forward filling for random deletion, adopting linear interpolation after reproduction for realizing smooth transition for long-time continuous deletion caused by shutdown or working condition switching.
5. The method for predicting sintering ignition temperature based on E-PINN as set forth in claim 3, wherein the feature normalization process is to normalize features with definite physical upper and lower bounds and distribution bias by Min-Max and to normalize approximately Gaussian distribution features by Z-score.
6. The method for predicting the sintering ignition temperature based on E-PINN as set forth in claim 1, wherein in the second step, the expression of the linear physical core model is: ; Wherein, the The temperature is predicted for the initial ignition and, Is the regression coefficient of the linear physical kernel model, Representation of Is inverted; The bias coefficient is the bias coefficient of the linear physical core model; is a set of input features.
7. The method of claim 6, wherein in the third step, the ignition temperature prediction model is expressed as follows: ; ; Wherein, the Is a multi-layer sensor, which is a multi-layer sensor, As a function of the residual network parameters, Is the temperature compensation quantity; for the final ignition to be predicted the temperature, Is a scaling factor.
8. The method for predicting sintering ignition temperature based on E-PINN as set forth in claim 7, further comprising the step of training and optimizing an ignition prediction model, wherein the method comprises the following specific steps of: (1) Constructing a comprehensive loss function of the ignition temperature: ; Wherein, the In order to integrate the loss function(s), A loss of consistency of the data is achieved, In the event of a loss of physical consistency, For the least intervention regularization term, λ phys and λ reg are trade-off coefficients; (2) Calculating data consistency loss: ; Wherein, the For the ignition temperature predicted by the ignition temperature prediction model, For ignition measurement temperature, N is the number of data measurement points; (3) Calculating physical consistency loss: constructing a physical residual error of the ignition temperature, and then solving the mean square error of the physical residual error: ; ; Wherein, the As a physical residual error, the difference between the two coefficients, To pair(s) An automatic differentiation is performed so that, As an implicit state equation for the firing temperature change rate, Representing a lumped thermophysical parameter set which is difficult to accurately measure; (4) Calculating a minimum intervention regularization term: The following formula is used for calculation: ; ; ; Wherein alpha and beta are weight coefficients; Automatic calculation of loss function versus residual network parameters by back propagation algorithm Partial derivative of (2) Finally, according to gradient descent algorithm, using learning rate Controlling step length, and iteratively updating residual error parameters To the convergence of the integrated loss function.
9. The method for predicting sintering ignition temperature based on E-PINN as set forth in claim 8, wherein the residual network parameters are set before the comprehensive loss function is calculated And initializing a scaling factor gamma to 0 to calculate an initial ignition prediction temperature, and substituting the initial ignition prediction temperature into the comprehensive loss function to train the ignition temperature prediction model.
10. The method for predicting sintering ignition temperature according to claim 9, wherein in the step of calculating the physical consistency loss, ; Wherein, the In order to integrate the energy dissipation ratio, Is an effective energy input rate.

Description

E-PINN-based sintering ignition temperature prediction method Technical Field The invention belongs to the technical field of sintering control, and particularly relates to a sintering ignition temperature prediction method based on E-PINN. Background In the long-flow steel production, the sintering process bears the important task of converting the powdery iron ore raw material into the sintering ore which can be used by the blast furnace, and the thermal state directly influences the strength of the sintering ore, the return rate and the stable operation of blast furnace smelting. Ignition is used as an initial link of sintering thermal cycle, initial heat flux is provided for a material surface through gas combustion in an ignition furnace, and a forming mode and a propelling condition of a combustion zone are determined, so that the method is a key process step for realizing the cooperative optimization of quality, energy consumption and yield in the sintering process. Engineering practice shows that when the ignition temperature is lower, the ignition is easy to cause insufficient ignition of coke powder and poor development of a sintered layer, so that the yield and strength are reduced, and when the ignition temperature is too high, local excessive melting is caused, the fuel consumption is obviously increased, the air permeability of a material layer is deteriorated, and the process fluctuation is further amplified. Therefore, stable control of the firing temperature is one of the core problems in the fine operation of the sintering process. The ignition process at the industrial site has significant non-linear and non-steady state characteristics. The gas flow and heat value fluctuation, the air-fuel ratio adjustment and the combustion air preheating temperature change directly affect the combustion heat release intensity, and the granularity composition and the water fluctuation of the material layer can change the air permeability of the material layer and indirectly disturb the combustion environment through the ways of hearth negative pressure, air leakage and the like. In addition, the ignition furnace and the structure thereof have larger thermal inertia, and obvious response delay exists in the temperature measurement signal, so that the effect of operation adjustment is difficult to embody in time. Under the condition, the traditional control mode which depends on experience setting or single feedback is easy to generate adjustment lag and temperature oscillation, and is difficult to adapt to complex working condition changes. The construction of the ignition temperature prediction model with foresight provides a reliable basis for prediction control, and becomes an important way for reducing energy consumption and stabilizing quality. Around modeling and prediction of sintering ignition and process, the prior art mainly comprises three types of mechanism modeling, data-driven modeling and mechanism-data fusion. The sintering process is generally described by a mechanism modeling method based on the physicochemical processes such as gas-solid heat transfer, fuel combustion, substance migration and the like. For example, numerical simulation of heat and mass transfer behavior in iron ore sintering process-reviewed in [ D ]. Zhejiang university, 2024 ] and numerical simulation of gas-solid heat and mass transfer behavior in iron ore sintering process [ J ]. Steel, 2015, 50 (9): 1-7 ] established and perfected numerical models of heat and mass transfer in sintering process and analyzed the influence of fuel distribution on temperature field and sintering behavior, numerical simulation of heat and mass transfer in iron ore sintering process based on gas-solid convective heat transfer coefficient calculation model [ J ]. Sintered pellets, 2025, 50 (1): 38-44 ] proposed a simulation method of heat and mass transfer in sintering process based on gas-solid convective heat transfer coefficient calculation model, numerical simulation of mass and heat transfer in iron ore sintering process [ J ]. North university (natural science edition), 2025, 46 (1): 35-43 ] studied numerical simulation of mass transfer in iron ore sintering process. The method has strong physical interpretability, but has high dependence on heat exchange coefficient, boundary condition and physical parameter, and has larger calculation complexity, and key parameters are difficult to acquire in real time in industrial sites, thus limiting the online application capability. The data driving method relies on industrial operation data, and a nonlinear mapping relation between technological parameters and temperature is characterized through a machine learning or deep learning model. For example ,"A prediction system of burn-through point based on gradient boosting decision tree and decision rules[J]. ISIJ International, 2019, 59(12): 2156–2164", a sintering end point (BTP) prediction system is built based on a gradient lifting tree, a