Tempering process steel temperature forecasting model of total element driven fusion mechanism and construction and application of tempering process steel temperature forecasting model

Abstract

The invention discloses a tempering process steel temperature forecasting model based on a total element driven fusion mechanism and construction and application of the tempering process steel temperature forecasting model, belongs to the field of tempering furnace optimization control of ferrous metallurgy, is used for enabling the steel temperature forecasting model to have interpretability, physical significance and extrapolation and have higher accuracy, and has the key points that a training data set is obtained; wherein the input data of the tempering process steel temperature forecasting model in the training data set is tempering furnace operation data, and the output data is the total heat absorption rate; and training the tempering process steel temperature forecasting model by the training set. Wherein the total heat absorption rate is obtained through the steps of establishing an identification target function with the total heat absorption rate as the core, identifying the target function through an improved GSK algorithm and the like, the accuracy and robustness in the face of new data are enhanced, and the discomfortability challenge in inverse problem solving in the field of steel temperature forecasting is effectively solved.

Description

Technical Field

[0001] The present invention belongs to the field of tempering furnace optimization control in iron and steel metallurgy, and specifically relates to a tempering process steel temperature prediction model with a full-factor driven fusion mechanism and its construction and application. Background Art

[0002] As a solid foundation for the development of the national economy and industry, the importance of the steel industry is obvious. Faced with multiple severe challenges such as high energy consumption, high emissions and high pollution, steel companies have always adhered to the concept of innovation and are committed to making breakthroughs in the efficient use of raw materials, production process innovation and the adoption of environmental protection technologies. From ore mining to ironmaking, steelmaking and steel rolling, the steel industry chain covers many key links, and each link is highly dependent on the stable operation of thermal equipment. Therefore, fine-tuning and adjusting the operating parameters of thermal equipment and further improving the process model are of great significance for the steel industry to promote energy conservation and emission reduction and improve comprehensive benefits.

[0003] The tempering furnace is an important heat treatment equipment after quenching, and its main function is to heat the steel plate to a specific preset temperature. However, this heating link is accompanied by significant energy consumption problems. Therefore, improving the degree of automation of the tempering furnace is of great significance to achieving the goal of energy conservation and emission reduction. The temperature distribution state of the steel plate inside the tempering furnace is not only a key indicator for evaluating the quality of the heating process, but also an important basis for optimizing the temperature control strategy of the tempering furnace. The heating process of the steel plate is extremely complex, showing the characteristics of nonlinearity, large time lag, strong coupling, and involving multiple heat transfer mechanisms. In addition, the internal environment of the tempering furnace is harsh, with problems of high temperature and dense dust, and directly and accurately measuring the temperature of the steel plate faces many challenges. In view of this, establishing an accurate steel plate temperature prediction model to estimate the temperature of the steel plate in the furnace has become the focus of current research.

[0004] The mathematical models for steel plate temperature estimation are mainly divided into two categories: data-driven models and mechanism models. The data-driven model focuses on the extensive collection and in-depth analysis of industrial data, aiming to discover the implicit relationship between the data and build an accurate temperature prediction model based on this. In contrast, the mechanism model uses the finite element method or the finite difference method to discretize the steel plate, and strictly follows the basic principles of heat transfer to clearly define the heat transfer equations and boundary conditions, thereby constructing a temperature estimation model with theoretical support. In addition, there is also a hybrid model that effectively combines the advantages of the mechanism model and the data-driven model.

[0005] The Chinese patent "CN114015863B A self-correction algorithm for billet heating model" provides an algorithm for automatically correcting the billet heating model. This patent uses the summary absorption rate method to establish a two-dimensional billet heating model, uses a drag couple experiment to determine the preliminary summary absorption rate, and calculates the temperature distribution unevenness coefficient of the billet along the length of the furnace; uses the heat balance method to calculate the heat absorbed by the billet during this period, and uses the billet heating model to calculate the total heat absorption of all billets in the furnace during this period. This method uses the mathematical model of the heating furnace to predict the changes in the internal temperature distribution of the billet during the heating process, uses the optimized furnace temperature setting value to control the heating furnace combustion process, and continuously corrects it in real time according to the on-site operation conditions.

[0006] Chinese patent "CN113849020B A method and device for designing a billet heating curve based on an artificial intelligence algorithm" divides each control section of the heating furnace into multiple sub-segments; obtains historical heating data corresponding to each sub-segment; uses the historical heating data corresponding to each sub-segment to train a preset neural network model to obtain a billet temperature prediction model corresponding to each sub-segment; according to the current value of the temperature influencing factor corresponding to each sub-segment, for each sub-segment, uses the billet temperature prediction model corresponding to the current sub-segment to predict the predicted value of the end outlet temperature corresponding to the sub-segment; fits the predicted values ​​of the end outlet temperature corresponding to each sub-segment to obtain a billet heating curve.

[0007] The Chinese patent "CN113821984B A method for calculating the temperature of a heating furnace billet based on a time-domain convolution model" first establishes a two-dimensional heat transfer mechanism model of the billet based on the basic principles of heat transfer, and calculates the temperature distribution of the billet in real time; then, the billet steel type, size, furnace time, furnace temperature, and billet temperature calculated by the mechanism model are collected, and the billet temperature measured by the infrared thermometer is formed into a feature data set and data preprocessing is performed. Finally, a correction model for the billet temperature is established based on the time-domain convolution model, the model is trained using the data set, and the trained model is used to correct the billet temperature in the furnace.

[0008] However, the technical solution described in the Chinese patent "CN113849020B A method and device for designing a steel billet heating curve based on an artificial intelligence algorithm" completely abandons the knowledge of the mechanism of the heat transfer process. Therefore, the strong nonlinearity, time lag and high dimensionality of the heating process in the furnace have certain adverse effects on the modeling process and model performance, and the data-driven model has low interpretability, physical meaning and extrapolation.

[0009] The technical solution described in the Chinese patent "CN114015863B A self-correction algorithm for billet heating model" requires the unknown heat transfer parameters in the equation to be determined in advance. The heat transfer parameters in the mechanism model of steel plate heating are determined based on the on-site embedded couple experiment. Through the embedded couple experiment, the workload of parameter acquisition is large, the debugging time is long, and it is easy to affect the production of steel plates.

[0010] The technical solution described in the Chinese patent "CN113821984B A method for calculating the temperature of a heating furnace steel billet based on a time domain convolution model" has insufficient input data, and only includes the measured furnace temperature, the measured steel temperature, and data related to the steel billet, while ignoring important influencing factors such as the burner operation status, fault status, and the flow rate of fuel and combustion air. Therefore, it cannot meet the needs of multiple steel grades and multiple working conditions. In addition, the temperature of the steel billet is only corrected at the location of the infrared thermometer, while ignoring the heating process of the steel billet. Summary of the invention

[0011] In order to solve the above problems, the present invention effectively reflects the process characteristics and laws through a mechanism model, which has good interpretability and extrapolation. The total heat absorption rate output by the full-factor driven recurrent neural network does not require a drag experiment, while overcoming the derivation of empirical formulas, and the resulting total heat absorption rate is more accurate.

[0012] In a first aspect, a method for constructing a steel temperature prediction model for a tempering process based on a full-factor driven fusion mechanism according to some embodiments of the present application includes:

[0013] Obtaining a training data set; wherein the input data of the tempering process steel temperature prediction model in the training data set is the tempering furnace operation data, and the output data is the overall heat absorption rate;

[0014] The tempering process steel temperature prediction model is trained by the training set.

[0015] According to the method for constructing a steel temperature prediction model for the tempering process of the full-factor driven fusion mechanism in some embodiments of the present application, the training data set is obtained based on the following method:

[0016] S3.1. Initialize the population of size NP, and the first individual x1 randomly initializes the generated variables within its specified range;

[0017] S3.2. Obtain the x sequence as the initialization population through Chebyshev mapping;

[0018] S3.3. Sort all individuals in the current population in ascending order according to the objective function value: x1, x2, ..., x NP , update the number of evaluations F E =NP, where x1 is the individual with the best solution to the objective function, x NP The individual with the worst solution to the objective function;

[0019] S3.4.(1) The random number rand is greater than the knowledge ratio K r When the individual is not updated, the fitness value of the objective function is calculated and the number of evaluations F is updated E =F E+ NP, update population; (2) the random number rand is not greater than the knowledge ratio K r When judging:

[0020] (2-1) The random number rand is not less than the number of dimensions N updated by the primary revenue sharing knowledge scheme junior When the primary knowledge sharing stage formula is used to update the individual Calculate the fitness value of the objective function and update the number of evaluations F E =F E+ NP, renewal population;

[0021] (2-2) The random number rand is smaller than the number of dimensions N updated by the primary revenue sharing knowledge scheme junior When the advanced knowledge sharing stage formula is used to update the individual Calculate the fitness value of the objective function and update the number of evaluations F E =F E+ NP, renewal population;

[0022] S3.5. Use the greedy selection method to decide whether to use the old individual or the new individual in the next iteration;

[0023] S3.6. Individuals to be selected The total heat absorption rate of each furnace zone corresponding to each dimension is substituted into the two-dimensional unsteady tempering furnace heat transfer model to calculate the furnace temperature of the jth steel plate measured. and the surface temperature of the jth steel plate measured by the pyrometer in the furnace

[0024] S3.7. Measure the temperature of the jth steel plate out of the furnace and the surface temperature of the jth steel plate measured by the pyrometer in the furnace and Substitute the identification objective function to solve the optimal overall heat absorption rate of the furnace area;

[0025] S3.8. After updating the global optimum using the greedy selection method, determine whether the maximum number of iterations has been reached. If so, output the optimal overall heat absorption rate of each furnace zone and the tempering furnace operation data for constructing the training data set; if not, return to step S3.3.

[0026] According to some embodiments of the present application, a method for constructing a steel temperature prediction model for a tempering process with a full-factor driven fusion mechanism is provided, wherein:

[0027] In step S3.1, a population of size NP is initialized, and the first individual x1 randomly initializes the generated variables within its specified range, as shown in the following formula:

[0028]

[0029] In the formula, x i is the i-th individual, i=1,2,…,NP,x i,j Represents individual x i The jth dimension of and are the lower and upper bounds of the j-th decision variable, respectively, and are random numbers in the range of [0,1];

[0030] In step S3.2, the x sequence is obtained as the initialization population through Chebyshev mapping, as shown in the following formula:

[0031] x i+1 =cos(a*cos -1 (x i )),a=4 (11)

[0032] In the formula, x i+1 is the i+1th individual, a represents the order;

[0033] In step S3.4, the formula of the primary knowledge sharing stage is updated:

[0034]

[0035] In the formula, Represents an individual The jth dimension of Represents an individual The jth dimension of K f represents the knowledge factor; Represents an individual The j-th dimension of , Guass(0,1) represents a Gaussian distribution with a mean of 0 and a variance of 1, represents the jth dimension of a random individual;

[0036] In step S3.4, the advanced knowledge sharing stage formula is updated:

[0037]

[0038] In the formula, The best individual The jth dimension of The worst individual The jth dimension of represents the jth dimension of a random individual;

[0039] In step S3.5, a greedy selection method is used to decide whether to use the old individual or the new individual in the next iteration, as shown in the following formula:

[0040]

[0041] In the formula, represents the selected individual, It is the solution generated by the primary and advanced stages of acquiring shared knowledge. is the i-th solution in the current population.

[0042] According to the method for constructing a steel temperature prediction model for the tempering process of the full-factor driven fusion mechanism in some embodiments of the present application, the tempering furnace operation data includes burner duty cycle, burner fault status, fuel supply and air-fuel ratio, gas preheating temperature, steel plate speed, steel plate position, furnace zone temperature, steel plate measurement length, measurement width, measurement thickness, steel plate number, steel plate type, steel plate entry temperature, steel plate surface temperature measured by a pyrometer in the furnace, steel plate exit temperature, steel plate operation speed, furnace length, heating time, insulation time and any one or combination of the temperatures measured by thermocouples in each furnace zone.

[0043] According to the method for constructing a steel temperature prediction model for the tempering process of the full-factor driven fusion mechanism in some embodiments of the present application, the tempering furnace operation data includes burner duty cycle, burner fault status, fuel supply and air-fuel ratio, gas preheating temperature, steel plate speed, steel plate position and furnace zone temperature.

[0044] According to the method for constructing a steel temperature prediction model for the tempering process of the full-factor driven fusion mechanism in some embodiments of the present application, the mathematical expression of the identification objective function is as follows:

[0045]

[0046] In the formula, J represents the objective function, N represents the number of steel plates, Indicates the temperature of the jth steel plate measured out of the furnace; represents the surface temperature of the jth steel plate measured by the pyrometer in the furnace; represents the temperature of the jth steel plate out of the tempering furnace calculated by the tempering furnace heat transfer model; represents the upper surface temperature of the jth steel plate calculated by the tempering furnace heat transfer model when it is transported to the position of the pyrometer in the furnace; W represents the combined vector of the total heat absorption rate at different positions along the length of the furnace, represents the mth total heat absorption rate along the furnace length, and λ1 and λ2 represent the regularization coefficients.

[0047] According to the method for constructing a steel temperature prediction model for the tempering process of the full-factor driven fusion mechanism in some embodiments of the present application, the two-dimensional non-steady-state tempering furnace heat transfer model is established based on the following method:

[0048] S1.2. Establish a two-dimensional non-steady-state heat conduction equation. The specific mathematical description is as follows:

[0049]

[0050] Where, ρ represents the density of steel plate; C p represents the specific heat of the steel plate; T represents the temperature of the steel plate; τ represents the time; λ represents the thermal conductivity of the steel plate; represents the differential along the length direction of the steel plate, represents the differential in the thickness direction of the steel plate;

[0051] S1.3. Establish boundary conditions. The heat flux density on the steel plate surface is mathematically described as:

[0052]

[0053] Where q represents the heat flux density on the steel plate surface; σ represents the Boltzmann constant; Indicates the total heat absorption rate; T f Indicates furnace temperature, T s Indicates the surface temperature of the steel plate;

[0054] S1.4. According to the change of the position of any steel plate in the furnace over time, the boundary conditions of the steel plate are converted into a time-varying temperature field problem, where for any steel plate i, the position x of the tail of the steel plate in the furnace i Determined by the moving speed of the steel plate, expressed as follows:

[0055]

[0056] In the formula, x i (v(t),t) represents the position of the tail of the steel plate in the furnace, t represents the time variable, t f is the time the steel plate is in the furnace, dτ represents the time differential, ν(t) is the speed of the steel plate at time t;

[0057] S1.5. Calculate the boundary conditions, expressed as follows:

[0058]

[0059] In the formula, T(τ,x i ,y) represents the temperature of the steel plate, x i represents the tail position of the steel plate, x represents the coordinate axis along the length direction of the steel plate, y represents the coordinate axis along the thickness direction of the steel plate, b represents the thickness of the steel plate, a represents the length of the steel plate, and q h and q tThey represent the heat flux density on the front side and the heat flux density on the rear side of the steel plate, respectively, and q u and q l They represent the heat flux density on the upper surface and the heat flux density on the lower surface of the steel plate respectively;

[0060] S1.6. The finite difference method is used to discretize the two-dimensional non-steady-state heat conduction equation and boundary conditions, so as to iterate the two-dimensional non-steady-state tempering furnace heat transfer model and output the relevant temperatures at each moment, where the relevant temperatures include the temperature of the jth steel plate measured out of the furnace. and the surface temperature of the jth steel plate measured by the pyrometer in the furnace and the internal temperature of the steel plate.

[0061] According to the method for constructing a steel temperature prediction model for the tempering process of the full-factor driven fusion mechanism in some embodiments of the present application, the model is a recurrent neural network, and the input vector x of the model t =[x 1,t ,x 2,t ,...,x M,t ] represents the state of the tempering furnace operation data at time t, M represents the number of types of tempering furnace operation data, and the output vector y t represents the estimated value of the overall heat absorption rate along the furnace length at time t;

[0062] For any given time step t, the RNN recursively calculates the new hidden state h by the following formula t and the predicted value y t :

[0063] h t =tanh(W hx x t +W hh h t-1 +b h ) (17)

[0064] y t =W hy h t +b y (18)

[0065] Where W hx ∈R m×n ,W hh ∈R m×m represents the weight matrix, W hy represents the weight matrix of the output layer, x t represents the input layer, h t-1 represents the hidden layer, n represents the number of input features, b h ∈R m and b y∈R represents the bias terms for the hidden layer and the output layer respectively, the tanh function is used to introduce nonlinearity, and m represents the number of neurons in the hidden layer;

[0066] The model training includes

[0067] S4.1. Initialize parameters, randomly initialize all weights W hx , W hh , W hy and the bias term b h , b y ;

[0068] S4.2. Forward propagation, given a batch of training samples in For a series of continuous observations, For the corresponding overall heat absorption rate, perform forward propagation to obtain the predicted value In the formula, N represents the number of input data groups. represents the output value of the Tth dimension in the training sample, Represents the predicted value of the Tth dimension during training, where T represents the dimension of the output layer;

[0069] S4.3. Calculate the loss, expressed as follows:

[0070]

[0071] Where L represents the loss function;

[0072] S4.4. Use gradient descent to adjust parameters to minimize the loss function for back propagation, including: calculating the gradient ▽L with respect to W hx , W hh , W hy , b h , b y , update the parameters according to the learning rate η, expressed as follows:

[0073]

[0074]

[0075]

[0076] In the formula, represents the partial derivative with respect to the weight, and L represents the loss function;

[0077] S4.5. Repeat the training until the stopping condition is met.

[0078] In the second aspect, a steel temperature prediction model for the tempering process based on the full-factor driven fusion mechanism of some embodiments of the present application is constructed by any of the methods described in the first aspect.

[0079] In a third aspect, a method for predicting steel temperature in a tempering process according to a full-factor driven fusion mechanism of some embodiments of the present application is characterized in that the tempering process steel temperature prediction model constructed by the method described in the second aspect is implemented;

[0080] The method comprises:

[0081] Inputting the tempering furnace operation data into the tempering process steel temperature prediction model;

[0082] The steel temperature prediction model of the tempering process outputs the summary heat absorption rate of each furnace zone;

[0083] According to the comprehensive heat absorption rate of each furnace zone, the steel temperature during the tempering process is calculated by a two-dimensional non-steady-state tempering furnace heat transfer model.

[0084] In a third aspect, an embodiment of the present application further provides an electronic device, comprising: one or more processors, a memory, and one or more programs; wherein the one or more programs are stored in the memory, and the one or more programs include instructions, which, when executed by the electronic device, enable the electronic device to execute the first and third aspects and any possible technical solutions thereof.

[0085] In a fourth aspect, an embodiment of the present application further provides a computer-readable storage medium, wherein the computer-readable storage medium includes a computer program. When the computer program runs on an electronic device, the electronic device executes the first aspect and the first and third aspects and any possible technical solutions thereof.

[0086] Beneficial effects of the present invention: Compared with the prior art, the steel temperature prediction model constructed by the technical solution proposed by the present invention integrates the heat transfer formula, has better interpretability, physical meaning and extrapolation than the data-driven model, and has smaller fluctuation of the steel temperature output value. The summary heat absorption rate in the technical solution does not need to be determined in advance by dragging couple experiment, and the summary heat absorption rate is more in line with reality than the value calculated by the empirical formula, so the model accuracy is higher than the mechanism model.

[0087] Specific:

[0088] In a first aspect, the present invention is to train a recurrent neural network using all-factor tempering furnace operation data as input and the identified overall heat absorption rate as output. The overall heat absorption rate of the tempering furnace is usually affected by a series of continuous operating parameters, including the health status of each burner, the operating status (duty cycle), the furnace temperature of each temperature control zone, the fuel flow rate and pressure, the combustion air flow rate and pressure, etc., and these parameters change over time. RNN can effectively capture the time dependency in the input data by storing historical information through hidden layers, which is very important for understanding and predicting the heat absorption rate.

[0089] In the second aspect, the present invention proposes a GSK algorithm combined with a Gaussian mutation strategy to solve the problem of identifying the overall heat absorption rate of a tempering furnace. The initial population distribution is crucial to the convergence speed and accuracy. In order to maintain population diversity, the Chebyshev mapping is used to initialize the population. Compared with the random method, this makes the individual distribution more uniform, reduces overlap, and enhances the global search ability of the algorithm. In addition, by introducing Gaussian mutation, the population diversity is further improved, which helps explore a wider solution space and avoid falling into the local optimum, thereby increasing the probability of finding a dominant solution.

[0090] In the third aspect, the objective function designed by the present invention takes the overall heat absorption rate as the core, and models the steel plate heating process based on the two-dimensional heat transfer model, which ensures high accuracy and computational efficiency. The regularization strategy is applied in the objective function to optimize the structure, which not only prevents overfitting and improves the generalization ability of the model, but also enhances the accuracy and robustness when facing new data, effectively solving the challenge of ill-posedness in solving inverse problems. BRIEF DESCRIPTION OF THE DRAWINGS

[0091] Figure 1 Technical solution flow chart.

[0092] Figure 2 Schematic diagram of heating of a two-dimensional steel plate cross section.

[0093] Figure 3 Flow chart of the improved GSK algorithm for identifying the overall heat absorption rate.

[0094] Figure 4 Schematic diagram of a recurrent neural network. DETAILED DESCRIPTION

[0095] The following describes the embodiments of the present application in detail with reference to the accompanying drawings, examples of which are shown in the accompanying drawings. The present application provides a method and an electronic device. The method and the device are based on the same technical concept. Since the principles of solving problems by the method and the device are similar, the implementation of the device and the method can refer to each other, and the repeated parts will not be repeated.

[0096] like Figure 1 As shown, the method for constructing a steel temperature prediction model for the tempering process with a full-factor driven fusion mechanism proposed in the present invention includes the following steps:

[0097] S1: Construct a two-dimensional unsteady-state tempering furnace heat transfer model and boundary conditions.

[0098] S2: Establish an identification objective function with the overall heat absorption rate as the core.

[0099] S3: Use the improved GSK algorithm to identify the target function.

[0100] S4: The tempering furnace operation data is used as input and the corresponding summary heat absorption rate is used as output to train the recurrent neural network.

[0101] S5: The dynamically changing summary heat absorption rate output by the recurrent neural network is brought into the tempering furnace heat transfer model to obtain the steel plate temperature prediction model.

[0102] Figure 1 The main flow chart of the technical solution of the present invention is shown, taking the tempering heat treatment production line of a steel plant as an example:

[0103] The tempering furnace operation data includes: burner duty cycle, burner fault status, fuel supply and air-fuel ratio, gas preheating temperature, steel plate speed, steel plate position, furnace zone temperature, steel plate measurement length, measurement width, measurement thickness, steel plate number, steel plate type, steel plate entry temperature, steel plate surface temperature measured by the furnace pyrometer, steel plate exit temperature, steel plate operation speed, furnace length, heating time, insulation time, and thermocouple measurement temperature of each furnace zone

[0104] The steel plate specifications are 20mm×2000mm×12000mm, the roller speed is 0.2-20m / min, the steel plate entering the furnace is room temperature 20℃, and the process requires the surface temperature of the steel plate to reach 180℃ when it leaves the furnace. The effective length of the tempering furnace is 106m, with a total of 34 upper and lower furnace zones, each furnace zone is equipped with 6-12 burners, the burner type is a high-speed burner, and the middle section and the tail of the tempering furnace are equipped with furnace pyrometers to measure the surface temperature of the steel plate. The algorithm termination condition is to reach the maximum number of iterations.

[0105] The two-dimensional non-steady-state tempering furnace heat transfer model in step S1 is shown in Figure 2 , the build steps are as follows:

[0106] S1.1: Construct a two-dimensional unsteady-state tempering furnace heat transfer model and boundary conditions. First, make the following assumptions to simplify the model.

[0107] a) Ignore the effects of steel plate oxide scale and surface chemical changes on heat transfer conditions.

[0108] b) It is approximately assumed that the ambient temperature of the steel plate is uniform in each temperature control zone;

[0109] c) Ignore the influence of heat conduction and radiation between the steel plate and the furnace roller;

[0110] d) The steel plate moves at a constant speed in the furnace, and the heat conduction along the length of the steel plate is ignored;

[0111] e) For the convenience of calculation, the convective heat transfer heat flux density is expressed in the form of radiation heat transfer heat flux density and converted into a radiation processing method, that is, the overall heat absorption rate method is used to calculate the steel temperature.

[0112] f) It is assumed that the temperature of the medium in the furnace is evenly distributed along the width of the furnace

[0113] g) Considering the thermocouple installation position and insertion depth, it is approximately assumed that the furnace gas temperature is equal to the furnace temperature.

[0114] Based on the above assumptions, a mathematical model of steel plate heating process is established.

[0115] S1.2: Establish a two-dimensional non-steady-state heat conduction equation. The specific mathematical description is as follows:

[0116]

[0117] Where, ρ represents the density of steel plate, Kg / m 3 ; C p represents the specific heat of the steel plate, J / (kg●℃); T represents the temperature of the steel plate, ℃; τ represents the time, s; λ represents the thermal conductivity of the steel plate, W / (m●℃); represents the differential along the length direction of the steel plate, It represents the differential in the thickness direction of the steel plate.

[0118] S1.3: Establish boundary conditions. The heat flux density on the steel plate surface is mathematically described as:

[0119]

[0120] Where: q represents the heat flux density on the steel plate surface, W / m 2 ; σ is the Boltzmann constant, 5.67●10 -8 W / (m 2 ●K 4 ); Indicates the total heat absorption rate; T f With T s Represent furnace temperature and steel plate surface temperature, K respectively.

[0121] S1.4: Since the position of a particular steel plate in the furnace changes with time, a moving coordinate system is used, so that the boundary conditions of the steel plate are also converted into a time-varying temperature field problem. Consider any steel plate i in the furnace, the position of the tail of the steel plate in the furnace is x i Determined by the moving speed of the steel plate:

[0122]

[0123] Where: x i (v(t),t) represents the position of the tail of the steel plate in the furnace; t represents the time variable; t f is the time the steel plate is in the furnace; dτ represents the time differential; ν(t) is the speed of the steel plate at time t, m / min; t fis the time the steel plate is in the furnace, s.

[0124] S1.5: The boundary conditions corresponding to the steel plate tracking model are:

[0125]

[0126] Where: T(τ,x i ,y) means, x i represents the tail position of the steel plate, x represents the coordinate axis along the length direction of the steel plate, y represents the coordinate axis along the thickness direction of the steel plate, b represents the thickness of the steel plate, a represents the length of the steel plate, and q h and q t are the heat flux density of the front side and the heat flux density of the rear side of the steel plate, W / m 2 .q u and q l are the upper surface heat flux density and the lower surface heat flux density, W / m 2 .

[0127] S1.6: The finite difference method is used to discretize the heat conduction equation and boundary conditions of the steel plate. The calculation format of the numerical equation established by the finite difference method is an explicit difference format, and each node equation can be solved independently. This allows the model to be iterated to generate relevant temperatures at each moment, including the temperature of the jth steel plate measured out of the furnace. and the surface temperature of the jth steel plate measured by the pyrometer in the furnace And the temperature inside the steel plate.

[0128] In step S2, an identification objective function with the overall heat absorption rate as the core is established, requiring the model calculation value and the steel temperature measured value to have the smallest error. In order to solve the ill-posed problem, the regularization strategy is applied to the modification of the objective function of this optimization problem. The purpose of this operation is to reduce the overfitting of the model to the data while improving the generalization ability of the model. The mathematical expression of the identification objective function is as follows:

[0129]

[0130] Wherein, J represents the objective function. In this embodiment, 20 steel plates are used. It represents the upper surface temperature of the jth steel plate measured out of the furnace, °C; represents the surface temperature of the jth steel plate measured by the pyrometer in the furnace, °C; represents the temperature of the jth steel plate out of the tempering furnace calculated by the tempering furnace heat transfer model, °C; represents the upper surface temperature of the jth steel plate when it is transported to the position of the pyrometer in the furnace calculated by the tempering furnace heat transfer model, °C. W represents the combined vector of the total heat absorption rate at different positions along the length of the furnace, represents the mth total heat absorption rate along the length of the furnace. λ1 and λ2 are both regularization coefficients.

[0131] The improved GSK algorithm proposed in step S3 performs the identification and optimization process as follows: Figure 3 As shown, the running steps are as follows:

[0132] S3.1: The algorithm first initializes a population of size NP. The first individual x1 randomly initializes the generated variables within its specified range. The variable generation formula is as follows:

[0133]

[0134] In the formula, x i is the ith individual, x i,j Represents individual x i The jth dimension of , i = 1, 2, ..., NP, and are the lower and upper bounds of the jth decision variable, and take a random number in the range [0,1]; rand() represents a random function

[0135] S3.2: Chebyshev mapping has good traversal uniformity and can improve the optimization speed of the algorithm. At the same time, it can generate initial values ​​with a relatively uniform distribution between [-1,1]. Therefore, the Chebyshev mapping will be used to initialize the population. The mapping expression is as follows:

[0136] x i+1 =cos(a*cos -1 (x i )),a=4 (11)

[0137] In the formula, x i+1 is the i+1th individual, and a represents the order.

[0138] Step S3.1 randomly generates the initial value of individual x1, and iterates according to the above formula to generate an x ​​sequence, i=i+1; this x sequence is used as the initial population.

[0139] S3.3: Note that the number of dimensions updated or changed using the primary and advanced shared knowledge acquisition scheme will depend on the knowledge rate K. The knowledge rate is used to control the amount of knowledge transferred between generations using other solutions using the primary and advanced shared knowledge acquisition scheme. Therefore, the number of dimensions that will be changed using the primary phase and the number of dimensions that will be changed using the advanced phase must be calculated at the beginning of each generation.

[0140] Let N junior is the number of dimensions updated using the primary revenue sharing knowledge scheme, calculated as

[0141]

[0142] In the formula, d is the problem size; K is a positive integer greater than zero, indicating the knowledge rate; g is the number of iterations; g max is the maximum number of iterations. The number of dimensions N that will be updated in the advanced revenue sharing phase senior Calculated by the following formula:

[0143] N senior =dN junior (13)

[0144] Make the following judgment:

[0145] (1) The random number rand is greater than the knowledge ratio K r When the individual is not updated, the fitness value of the objective function is calculated and the number of evaluations F is updated E =F E+ NP, update population; (2) the random number rand is not greater than the knowledge ratio K r When judging:

[0146] (2-1) The random number rand is not less than the number of dimensions N updated by the primary revenue sharing knowledge scheme junior When the primary knowledge sharing stage formula is used to update the individual Calculate the fitness value of the objective function and update the number of evaluations F E =F E+ NP, renewal population;

[0147] (2-2) The random number rand is smaller than the number of dimensions N updated by the primary revenue sharing knowledge scheme junior When the advanced knowledge sharing stage formula is used to update the individual Calculate the fitness value of the objective function and update the number of evaluations F E =F E+ NP, renewal population;

[0148] S3.4: The following steps should be followed during the initial stage of harvesting and sharing knowledge:

[0149] Sort all individuals in the current population in ascending order according to the objective function value: x1, x2, ..., x NP . Where x1 is the solution with the minimum objective function value (best individual), x NP is the solution with the worst objective function value (worst individual). Then, for each solution x i , the following operations should be selected: (a) A better solution x than the current solution i-1 and the nearest worse solution x i+1 (Individual x i+1 The fitness of individual x i (b) Randomly select another individual xr As a source of shared knowledge, individuals need to determine whether they have the desire to learn before updating.

[0150] The formula for the primary harvesting and sharing knowledge stage is as follows:

[0151]

[0152] Indicates individual The jth dimension of Represents an individual The j-th dimension, K f represents the knowledge factor; Represents an individual The j-th dimension of , Guass(0,1) represents a Gaussian distribution with a mean of 0 and a variance of 1, represents the j-th dimension of a random individual.

[0153] The knowledge factor K f is a real number greater than 0 that controls the total amount of knowledge acquired and shared that will be added to the current individual from other individuals over generations.

[0154] Two individuals (closest to the best and closest to the worst) need to update the third individual. When updating the best and worst solutions, the two closest to the best and worst solutions are selected respectively. i is the global optimal solution, then the following two solutions should be used to update the global optimal solution: (x i ,x i+1 ,x i+2 ). If x i is the global worst value, then the two closest previous solutions should be used to update the global worst value: (x i ,x i-1 ,x i-2 ).

[0155] S3.5: The following steps should be followed during the Advanced Harvesting and Sharing Knowledge phase:

[0156] Sort all individuals in the current population in ascending order according to the objective function value: x1, x2, …, x NP . Where x1 is the solution with the minimum objective function value (best individual), x NPis the solution with the worst objective function value (worst individual). Then, the classified population is divided into three groups. The first group contains the best solutions, the second group contains medium or better solutions, and the last group contains the worst solutions. Finally, in order to update each solution x1, two solutions are randomly selected from the best and worst 100p% solutions in the entire population to form the obtained part, while the third solution is randomly selected from the middle NP-2×100p% solution to form the shared part.

[0157] The formula for the advanced harvesting and sharing knowledge stage is as follows:

[0158]

[0159] express, The best individual The jth dimension of The worst individual The jth dimension of represents the j-th dimension of a random individual.

[0160] S3.6: After the initialization and primary and high-level knowledge acquisition and sharing phases, GSK adopts a greedy selection method to decide whether to use the old or new solution in the next iteration, which can be expressed as:

[0161]

[0162] in It is the solution generated in the primary or advanced stage of acquiring shared knowledge. is the i-th solution in the current population. Among them, the selected individuals are express.

[0163] S3.7: The individuals to be selected are The total heat absorption rate of each furnace zone corresponding to each dimension is substituted into the two-dimensional non-steady-state tempering furnace heat transfer model of step S1 to calculate the discharge temperature of the jth billet to be measured. and the upper surface temperature of the jth steel billet measured by the pyrometer in the furnace

[0164] S3.8: Then measure the temperature of the jth billet out of the furnace and the upper surface temperature of the jth steel billet measured by the pyrometer in the furnace and Substitute into the identification objective function in step S2 Find

[0165] S3.9: After updating the global optimum using greedy selection, determine whether the maximum number of iterations has been reached. If so, output the optimal overall heat absorption rate of each furnace zone; if not, return to step S3.3.

[0166] Step S4 uses the tempering furnace operation data as input and the corresponding total heat absorption rate as output to train the recurrent neural network. The recurrent neural network is as follows: Figure 4 As shown, the running steps are as follows:

[0167] S4.1: Model definition. Input vector x t =[x 1,t ,x 2,t ,...,x M,t ] represents all relevant variables at time t, which here represent the tempering furnace operation data such as burner duty cycle, burner fault status, fuel supply and air-fuel ratio, gas preheating temperature, steel plate speed, steel plate position, furnace zone temperature, etc. The size of M is determined by the type of operation data. Output vector y t , represents the estimated value of the total heat absorption rate along the length of the furnace at time t. Hidden state h t ∈R m , represents the model internal memory, where m is the number of neurons in the hidden layer.

[0168] For any given time step t, the RNN recursively calculates the new hidden state h by the following formula t and the predicted value y t :

[0169] h t =tanh(W hx x t +W hh h t-1 +b h ) (17)

[0170] y t =W hy h t +b y (18)

[0171] Here, W hx ∈R m×n ,W hh ∈R m×m is the weight matrix, and n is the number of input features. h ∈R m and b y ∈R are the bias terms for the hidden layer and the output layer respectively. The tanh function is used to introduce nonlinearity to help the model learn more complex patterns. ,W hy represents the weight matrix of the output layer, x trepresents the input layer, h t-1 represents the hidden layer.

[0172] S4.2: Initialize parameters, randomly initialize all weights W hx , W hh , W hy and the bias term b h , b y These initial values ​​are usually drawn from a normal distribution within a small range.

[0173] S4.3: Forward propagation, given a batch of training samples (in For a series of continuous observations, is the corresponding total heat absorption rate), and forward propagation is performed to obtain the predicted value Where N represents the number of input data groups. represents the output value of the Tth dimension in the training sample, It represents the predicted value of the Tth dimension during training, and T represents the dimension of the output layer.

[0174] S4.4: Calculate the loss and choose a suitable loss function to measure the difference between the predicted value and the true value. For regression problems, common choices are mean square error:

[0175]

[0176] S4.5: Back propagation, using the gradient descent method to adjust the parameters to minimize the above loss function. First, we need to calculate the gradient About W hx , W hh , W hy , b h , b y , and then update the parameters according to the selected learning rate η:

[0177]

[0178]

[0179] In the formula, represents the partial derivative with respect to the weight, and L represents the loss function.

[0180] S4.6: Cycle training, repeat the above steps for multiple rounds until the stopping condition is met (such as reaching the maximum number of iterations or the performance on the validation set no longer improves). During this period, a series of operations such as forward propagation, loss calculation, back propagation and parameter update are performed on the entire dataset.

[0181] In step S5, the dynamically changing overall heat absorption rate output by the recurrent neural network is introduced into the tempering furnace heat transfer model, thereby obtaining a more accurate steel plate temperature prediction model.

[0182] Compared with the prior art, the steel temperature prediction model constructed by the technical solution proposed in the present invention integrates the heat transfer formula, has better interpretability, physical meaning and extrapolation than the data-driven model, and has smaller fluctuation of the steel temperature output value. The summary heat absorption rate in the technical solution does not need to be determined in advance by dragging couple experiment, and the summary heat absorption rate is more in line with reality than the value calculated by the empirical formula, so the model accuracy is higher than the mechanism model.

[0183]

[0184] Based on the above embodiments, the embodiments of the present application further provide a computer program, which, when executed on a computer, enables the computer to execute the methods provided in the above embodiments.

[0185] Based on the above embodiments, the embodiments of the present application further provide a computer storage medium, in which a computer program is stored. When the computer program is executed by a computer, the computer executes the method provided in the above embodiments.

[0186] The storage medium may be any available medium that can be accessed by a computer. For example, but not limited to, a computer-readable medium may include RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage medium or other magnetic storage device, or any other medium that can be used to carry or store the desired program code in the form of instructions or data structures and can be accessed by a computer.

[0187] Based on the above embodiments, an embodiment of the present application further provides a chip, which is used to read a computer program stored in a memory to implement the method provided in the above embodiments.

[0188] Based on the above embodiments, an embodiment of the present application provides a computer program product. When the computer program product is run on an electronic device, the method provided in the above embodiments is implemented.

[0189] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0190] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0191] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0192] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0193] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A method for constructing a steel temperature prediction model for a tempering process with a full-factor driven fusion mechanism, characterized in that: include Obtaining a training data set; wherein the input data of the tempering process steel temperature prediction model in the training data set is the tempering furnace operation data, and the output data is the overall heat absorption rate; The tempering process steel temperature prediction model is trained by the training set.

2. The method for constructing a steel temperature prediction model for the tempering process with a full-factor driven fusion mechanism according to claim 1 is characterized in that: The training data set is obtained based on the following method: S3.

1. Initialize the population of size NP, and the first individual x1 randomly initializes the generated variables within its specified range; S3.

2. Obtain the x sequence as the initialization population through Chebyshev mapping; S3.

3. Sort all individuals in the current population in ascending order according to the objective function value: x1, x2, ..., x NP , update the number of evaluations F E =NP, where x1 is the individual with the best solution to the objective function, x NP The individual with the worst solution to the objective function; S3.4.(1) The random number rand is greater than the knowledge ratio K r When the individual is not updated, the fitness value of the objective function is calculated and the number of evaluations F is updated E =F E+ NP, update population; (2) the random number rand is not greater than the knowledge ratio K r When judging: (2-1) The random number rand is not less than the number of dimensions N updated by the primary revenue sharing knowledge scheme junior When the primary knowledge sharing stage formula is used to update the individual Calculate the fitness value of the objective function and update the number of evaluations F E =F E+ NP, renewal population; (2-2) The random number rand is smaller than the number of dimensions N updated by the primary revenue sharing knowledge scheme junior When the advanced knowledge sharing stage formula is used to update the individual Calculate the fitness value of the objective function and update the number of evaluations F E =F E+ NP, renewal population; S3.

5. Use the greedy selection method to decide whether to use the old individual or the new individual in the next iteration; S3.

6. Individuals to be selected The total heat absorption rate of each furnace zone corresponding to each dimension is substituted into the two-dimensional unsteady tempering furnace heat transfer model to calculate the furnace temperature of the jth steel plate measured. and the surface temperature of the jth steel plate measured by the pyrometer in the furnace S3.

7. Measure the temperature of the jth steel plate out of the furnace and the surface temperature T of the jth steel plate measured by the pyrometer in the furnace j,g2 and Substitute the identification objective function to solve the optimal overall heat absorption rate of the furnace area; S3.

8. After updating the global optimum using the greedy selection method, determine whether the maximum number of iterations has been reached. If so, output the optimal overall heat absorption rate of each furnace zone and the tempering furnace operation data for constructing the training data set; if not, return to step S3.

3.

3. The method for constructing a steel temperature prediction model for the tempering process of the full-factor driven fusion mechanism according to claim 2 is characterized in that: in: In step S3.1, a population of size NP is initialized, and the first individual x1 randomly initializes the generated variables within its specified range, as shown in the following formula: In the formula, x i is the i-th individual, i=1,2,…,NP,x i,j Represents individual x i The jth dimension of and are the lower and upper bounds of the j-th decision variable, respectively, and are random numbers in the range of [0,1]; In step S3.2, the x sequence is obtained as the initialization population through Chebyshev mapping, as shown in the following formula: x i+1 =cos(a*cos -1 (x i )),a=4 (11) In the formula, x i+1 is the i+1th individual, a represents the order; In step S3.4, the formula of the primary knowledge sharing stage is updated: In the formula, Represents an individual The jth dimension of Represents an individual The jth dimension of K f represents the knowledge factor; Represents an individual The jth dimension of , Guass(0,1) represents a Gaussian distribution with a mean of 0 and a variance of 1. represents the jth dimension of a random individual; In step S3.4, the advanced knowledge sharing stage formula is updated: In the formula, The best individual The jth dimension of The worst individual The jth dimension of represents the jth dimension of a random individual; In step S3.5, a greedy selection method is used to decide whether to use the old individual or the new individual in the next iteration, as shown in the following formula: In the formula, represents the selected individual, It is the solution generated by the primary and advanced stages of acquiring shared knowledge. is the i-th solution in the current population.

4. The method for constructing a steel temperature prediction model for the tempering process with a full-factor driven fusion mechanism according to any one of claims 2 to 3, characterized in that: The tempering furnace operation data includes burner duty cycle, burner fault status, fuel supply and air-fuel ratio, gas preheating temperature, steel plate speed, steel plate position, furnace zone temperature, steel plate measurement length, measurement width, measurement thickness, steel plate number, steel plate type, steel plate entry temperature, steel plate surface temperature measured by the pyrometer in the furnace, steel plate exit temperature, steel plate operation speed, furnace length, heating time, insulation time and any one or combination of the temperatures measured by thermocouples in each furnace zone.

5. The method for constructing a steel temperature prediction model for the tempering process with a full-factor driven fusion mechanism according to claim 4 is characterized in that: The tempering furnace operation data includes burner duty cycle, burner fault status, fuel supply and air-fuel ratio, gas preheating temperature, steel plate speed, steel plate position and furnace zone temperature.

6. The method for constructing a steel temperature prediction model for the tempering process of the full-factor driven fusion mechanism according to any one of claims 2-3, characterized in that: The mathematical expression of the identification objective function is as follows: In the formula, J represents the objective function, N represents the number of steel plates, Indicates the temperature of the jth steel plate measured out of the furnace; represents the surface temperature of the jth steel plate measured by the pyrometer in the furnace; represents the temperature of the jth steel plate out of the tempering furnace calculated by the tempering furnace heat transfer model; represents the upper surface temperature of the jth steel plate calculated by the tempering furnace heat transfer model when it is transported to the position of the pyrometer in the furnace; W represents the combined vector of the total heat absorption rate at different positions along the length of the furnace, represents the mth total heat absorption rate along the furnace length, and λ1 and λ2 represent the regularization coefficients.

7. The method for constructing a steel temperature prediction model for the tempering process with a full-factor driven fusion mechanism according to any one of claims 2 to 3, characterized in that: The two-dimensional unsteady tempering furnace heat transfer model is established based on the following method: S1.

2. Establish a two-dimensional non-steady-state heat conduction equation. The specific mathematical description is as follows: Where, ρ represents the density of steel plate; C p represents the specific heat of the steel plate; T represents the temperature of the steel plate; τ represents the time; λ represents the thermal conductivity of the steel plate; represents the differential along the length direction of the steel plate, represents the differential in the thickness direction of the steel plate; S1.

3. Establish boundary conditions. The heat flux density on the steel plate surface is mathematically described as: Where q represents the heat flux density on the steel plate surface; σ represents the Boltzmann constant; Indicates the total heat absorption rate; T f Indicates furnace temperature, T s Indicates the surface temperature of the steel plate; S1.

4. According to the change of the position of any steel plate in the furnace over time, the boundary conditions of the steel plate are converted into a time-varying temperature field problem, where for any steel plate i, the position x of the tail of the steel plate in the furnace i Determined by the moving speed of the steel plate, expressed as follows: In the formula, x i (v(t),t) represents the position of the tail of the steel plate in the furnace, t represents the time variable, t f is the time the steel plate is in the furnace, dτ represents the time differential, ν(t) is the speed of the steel plate at time t; S1.

5. Calculate the boundary conditions, expressed as follows: In the formula, T(τ,x i ,y) represents the temperature of the steel plate, x i represents the tail position of the steel plate, x represents the coordinate axis along the length direction of the steel plate, y represents the coordinate axis along the thickness direction of the steel plate, b represents the thickness of the steel plate, a represents the length of the steel plate, and q h and q t represent the heat flux density on the front side and the heat flux density on the back side of the steel plate, respectively, u and q l They represent the heat flux density on the upper surface and the heat flux density on the lower surface of the steel plate respectively; S1.

6. The finite difference method is used to discretize the two-dimensional non-steady-state heat conduction equation and boundary conditions, so as to iterate the two-dimensional non-steady-state tempering furnace heat transfer model and output the relevant temperatures at each moment, where the relevant temperatures include the furnace temperature of the jth steel plate measured after it is taken out of the furnace. and the surface temperature of the jth steel plate measured by the pyrometer in the furnace and the internal temperature of the steel plate.

8. The method for constructing a steel temperature prediction model for the tempering process with a full-factor driven fusion mechanism according to any one of claims 2 to 3, characterized in that: The model is a recurrent neural network, and the input vector x of the model is t =[x 1,t ,x 2,t ,...,x M,t ] represents the state of the tempering furnace operation data at time t, M represents the number of types of tempering furnace operation data, and the output vector y t represents the estimated value of the overall heat absorption rate along the furnace length at time t; For any given time step t, the RNN recursively calculates the new hidden state h by the following formula t and the predicted value y t : h t =tanh(W hx x t +W hh h t-1 +b h ) (17) y t =W hy h t +b y (18) Where W hx ∈R m×n ,W hh ∈R m×m represents the weight matrix, W hy represents the weight matrix of the output layer, x t represents the input layer, h t-1 represents the hidden layer, n represents the number of input features, b h ∈R m and b y ∈R represents the bias terms for the hidden layer and the output layer respectively, the tanh function is used to introduce nonlinearity, and m represents the number of neurons in the hidden layer; The model training includes S4.

1. Initialize parameters, randomly initialize all weights W hx , W hh , W hy and the bias term b h , b y ; S4.

2. Forward propagation, given a batch of training samples in For a series of continuous observations, For the corresponding overall heat absorption rate, perform forward propagation to obtain the predicted value In the formula, N represents the number of input data groups. represents the output value of the Tth dimension in the training sample, Represents the predicted value of the Tth dimension during training, where T represents the dimension of the output layer; S4.

3. Calculate the loss, expressed as follows: Where L represents the loss function; S4.

4. Use gradient descent to adjust parameters to minimize the loss function for back propagation, including: calculating the gradient About W hx , W hh , W hy , b h , b y , update the parameters according to the learning rate η, expressed as follows: In the formula, represents the partial derivative with respect to the weight, and L represents the loss function; S4.

5. Repeat the training until the stopping condition is met.

9. A steel temperature prediction model for the tempering process with a full-factor driven fusion mechanism, constructed by the method described in any one of claims 1 to 9.

10. A method for predicting steel temperature during tempering process with all-factor driven fusion mechanism, characterized in that: The tempering process steel temperature prediction model constructed by the method of claim 9 is implemented; The method comprises: Inputting the tempering furnace operation data into the tempering process steel temperature prediction model; The steel temperature prediction model of the tempering process outputs the summary heat absorption rate of each furnace zone; According to the comprehensive heat absorption rate of each furnace zone, the steel temperature during the tempering process is calculated by a two-dimensional non-steady-state tempering furnace heat transfer model.

Images