CN-121981157-A - Material temperature real-time prediction method based on PINN

CN121981157ACN 121981157 ACN121981157 ACN 121981157ACN-121981157-A

Abstract

The invention provides a PINN-based real-time material temperature prediction method, and relates to the technical field of plate and strip rolling. The method comprises the steps of constructing PINN models, wherein PINN is a physical information neural network, adopting a multi-layer perceptron, determining a total loss function meeting heat conduction physical constraint, initial conditions, boundary conditions and data fitting requirements, generating a plurality of typical working conditions in a working condition space through coverage sampling, inputting historical furnace condition information and slab related information, training PINN models corresponding to the working conditions based on the determined loss function, establishing an offline pre-training model library, calculating similarity between an online working condition and known working conditions in the offline pre-training model library, and carrying out weighted fusion prediction on prediction results of PINN models corresponding to the similar working conditions by adopting a multi-model weighted integration strategy to obtain a real-time prediction temperature field of the slab. Therefore, the real-time calculation cost can be greatly reduced while the physical consistency is ensured, and the high-efficiency temperature field prediction is realized.

Inventors

GUO JIN
LIANG XINYU
Wang Gufan
LIN FENGQIN

Assignees

北京科技大学

Dates

Publication Date: 20260505
Application Date: 20251209

Claims (10)

1. A method for predicting material temperature in real time based on PINN, the method comprising: Constructing PINN a model, wherein PINN is a physical information neural network and adopts a multi-layer perceptron; determining a total loss function meeting thermal conduction physical constraints, initial conditions, boundary conditions and data fitting requirements; Generating a plurality of typical working conditions in a working condition space through coverage sampling, inputting historical furnace condition information and slab related information, training PINN models corresponding to the working conditions based on a determined loss function, and establishing an offline pre-training model library; And calculating the similarity of the known working conditions in the online working condition and the offline pre-training model library, and carrying out weighted fusion prediction on the predicted results of PINN models corresponding to a plurality of similar working conditions by adopting a multi-model weighted integration strategy to obtain a real-time predicted temperature field of the slab.
2. The PINN-based real-time mass temperature prediction method as defined in claim 1, wherein determining a total loss function that meets thermal conduction physical constraints, initial conditions, boundary conditions, and data fitting requirements includes: determining heat conduction equation loss based on heat conduction equation ; Determining initial condition loss based on temperature distribution at initial time ; Determining boundary condition loss based on surface convection heat exchange and radiation heat exchange boundary ; Slab temperature field calculated based on two-dimensional heat conduction finite difference method and used for determining observed data loss Wherein FDM represents a finite difference method; Weighted combination 、、、 Obtaining the total loss : Wherein, the Indicating the total loss of the total of the components, 、、 And Is a weight coefficient.
3. The PINN-based real-time mass temperature prediction method as claimed in claim 2, wherein the heat conduction equation loss is expressed as: Wherein, the In order to be a thermal diffusivity, , In order to be of a thermal conductivity coefficient, In order to achieve a density of the particles, Is specific heat capacity; the number of sampling points; output for PINN models The temperature of the lower plate blank is predicted, In the form of a spatial coordinate system, Is the time of day.
4. The PINN-based real-time mass temperature prediction method as claimed in claim 2, wherein the initial condition loss is expressed as: Wherein, the For the number of points of the initial condition, Output for PINN models The predicted temperature of the slab is set at the moment, Indicating that the slab is in An initial temperature field at that time.
5. The PINN-based real-time mass temperature prediction method as claimed in claim 2, wherein the boundary condition loss is expressed as: Wherein, the Representing convective heat transfer losses; representing radiant heat exchange losses; Is the heat conductivity coefficient; Is the surface emissivity of the slab; is a Stefan-Boltzmann constant; Representation of Convective heat transfer coefficient at the moment; Representation of Furnace wall temperature at time; output for PINN models The temperature of the lower plate blank is predicted, In the form of a spatial coordinate system, The time is the moment; N represents a normal vector, i.e. a direction perpendicular to the surface; indicating the number of sampling points in the radiant heat exchange loss.
6. The PINN-based real-time mass temperature prediction method as claimed in claim 2, wherein the observed data loss is expressed as: Wherein, the The slab temperature field obtained by the two-dimensional heat conduction finite difference method is calculated, In the form of a spatial coordinate system, The time is the moment; Is shown in At the moment of The temperature of the location.
7. The method for predicting material temperature in real time based on PINN as claimed in claim 1, wherein the generating a plurality of typical working conditions in the working condition space through coverage sampling, inputting historical furnace condition information and slab related information, training PINN models corresponding to each working condition based on the determined loss function, and establishing an offline pre-training model library includes: The method comprises the steps of obtaining historical furnace condition information and slab related information and preprocessing, wherein the preprocessing comprises normalization, interpolation and missing value processing, the slab related information comprises chemical element composition of a slab, furnace charging temperature, furnace discharging temperature and geometric information, the geometric information comprises length, width and thickness, and the furnace condition information comprises furnace charging initial temperature field distribution, heat conductivity coefficient, density, specific heat capacity, furnace wall temperature, convection heat exchange coefficient and radiation heat flow density; mapping variable information of different dimensions into a vector with a fixed length, compressing continuous variables through feature sampling or principal component analysis, and adopting independent thermal coding or numerical coding for discrete variables to obtain a working condition feature vector with a uniform format; Selecting a plurality of typical working conditions in a working condition space through coverage sampling, inputting historical furnace condition information and slab related information, and training PINN models corresponding to the working conditions based on the determined loss function Obtaining a model set covering various materials, geometric dimensions and initial conditions : And storing the trained PINN model, the working condition feature vector, the training super-parameters and the verification precision together, and establishing an offline pre-training model library.
8. The PINN-based real-time material temperature prediction method according to claim 7, wherein the method is specific to working conditions Lower slab, feature vector The method comprises the following steps: Wherein, the Respectively represent working conditions Carbon, silicon, manganese, phosphorus, sulfur, nickel, phosphorus, titanium, molybdenum and chromium; the sub-table indicates the length, width and thickness of the slab; the charging temperature and the discharging temperature of the slab are respectively indicated.
9. The method for predicting the material temperature in real time based on PINN of claim 1, wherein the calculating the similarity between the online working condition and the known working condition in the offline pre-training model library, and performing weighted fusion prediction on the predicted results of the PINN model corresponding to a plurality of similar working conditions by adopting a multi-model weighted integration strategy, comprises: collecting furnace condition information and slab related information of a slab to be predicted in real time to generate a characteristic vector of an online working condition ; Calculation of Feature vector of each working condition in off-line pre-training model library Similarity of (2); judging whether the maximum similarity is larger than a preset first threshold value or not; if the model is larger than the preset value, directly calling a corresponding PINN model to obtain the blank at any place Real-time predicted temperature field ; Judging whether the maximum similarity is smaller than a preset second threshold value or not; If the parameters are smaller than the preset parameters, the first PINN model is finely adjusted by utilizing the slab parameters and the furnace condition data of the slab to be predicted, which are acquired in real time, and the slab is output at random by utilizing the finely adjusted first PINN model Real-time predicted temperature field Wherein the first PINN model is an on-line working condition characteristic vector The offline PINN model with highest similarity; otherwise, selecting the closest to the online working condition according to the similarity PINN models corresponding to the working conditions are used as candidate models, and the prediction results of the selected candidate models are weighted and averaged to obtain the slab at random Final real-time predicted temperature field 。
10. The PINN-based real-time material temperature prediction method according to claim 9, wherein a final real-time predicted temperature field is obtained Expressed as: Wherein, the Is the first Predicting results of the candidate models; is the weight; for adjusting parameters, controlling the concentration degree of weight distribution; Is the similarity.

Description

Material temperature real-time prediction method based on PINN Technical Field The invention relates to the technical field of plate and strip rolling, in particular to a PINN-based real-time material temperature prediction method. Background In the modern steel manufacturing process, billet heating is an important process before rolling, and the temperature control precision of the billet heating is directly related to the structural performance and mechanical properties of the product. The temperature distribution in the heating furnace not only affects the heating uniformity of the billet, but also determines the stability and energy efficiency level of the subsequent rolling process. Along with the development of green manufacturing and intelligent manufacturing, the industrial site has higher requirements on temperature prediction and temperature rise control of the heating process, namely, the heating quality is ensured, and the energy consumption control and the service life of equipment are considered. The existing research on slab temperature prediction has a certain result in theoretical methods and local application, but the temperature prediction research depends on the traditional numerical simulation or a single data driving model, and the whole has obvious defects. Firstly, the existing numerical simulation method based on the heat transfer equation has high calculation complexity and insufficient real-time performance, and is difficult to meet the requirements of quick decision and on-line control of an industrial field. Secondly, a prediction model which is purely dependent on data driving has certain self-adaptive capability, but is easily influenced by noise and data fluctuation, shows overfitting risks, and causes the prediction result to have defects in stability and physical consistency. And the fusion method combining the mechanism and the data can obviously improve the precision in an experimental environment, but under the actual production condition, the model performance and the popularization capability are obviously reduced due to sparse temperature measurement data, complex boundary conditions and large uncertainty. Finally, the problem of optimizing the heating curve essentially belongs to a typical multi-objective global optimization task, the uniformity of the internal temperature of the slab is considered, the energy consumption and the running stability of equipment are considered, the conventional method generally has limitations on the optimization efficiency and the feasibility of the result, and an optimization scheme which has both precision and operability is difficult to provide for actual production. Disclosure of Invention The embodiment of the invention provides a material temperature real-time prediction method based on PINN, which can greatly reduce real-time calculation cost and realize efficient temperature field prediction while ensuring physical consistency. The technical scheme is as follows: In one aspect, a method for predicting material temperature in real time based on PINN is provided, the method comprising: Constructing PINN a model, wherein PINN is a physical information neural network and adopts a multi-layer perceptron; determining a total loss function meeting thermal conduction physical constraints, initial conditions, boundary conditions and data fitting requirements; Generating a plurality of typical working conditions in a working condition space through coverage sampling, inputting historical furnace condition information and slab related information, training PINN models corresponding to the working conditions based on a determined loss function, and establishing an offline pre-training model library; And calculating the similarity of the known working conditions in the online working condition and the offline pre-training model library, and carrying out weighted fusion prediction on the predicted results of PINN models corresponding to a plurality of similar working conditions by adopting a multi-model weighted integration strategy to obtain a real-time predicted temperature field of the slab. Further, determining the total loss function that satisfies the thermal conduction physical constraint, the initial condition, the boundary condition, and the data fitting requirement includes: determining heat conduction equation loss based on heat conduction equation ; Determining initial condition loss based on temperature distribution at initial time; Determining boundary condition loss based on surface convection heat exchange and radiation heat exchange boundary; Slab temperature field calculated based on two-dimensional heat conduction finite difference method and used for determining observed data lossWherein FDM represents a finite difference method; Weighted combination 、、、Obtaining the total loss: Wherein, the Indicating the total loss of the total of the components,、、AndIs a weight coefficient. Further, the heat conduction equation loss i