Steel temperature forecast construction method based on data driving and improved particle swarm optimization

Abstract

The invention discloses a steel temperature forecast construction method based on data driving and an improved particle swarm algorithm. The method comprises the following steps that data in the steel billet production process are collected, and the data comprise the material of a steel billet, the charging temperature of the steel billet, the thickness size of the steel billet, the temperature of each section of a heating furnace and the discharging temperature; performing equal-length processing on the data by adopting a mathematical method to obtain data conforming to a model input format; initializing a steel temperature prediction mathematical model and constructing a target function; the objective function is solved through an improved particle swarm algorithm so as to solve the steel temperature prediction model with the optimal parameters, and the improvement of the particle swarm algorithm comprises the steps that a random uniform distribution generation strategy is used for initializing a particle swarm; when positions and speeds of particles in the particle swarm are iteratively updated, a chromosome mutation strategy is introduced to increase diversity of the particle swarm; updating an inertia weight value by adopting a three-section self-adaptive change method; and judging whether the current solution is a locally optimal solution or not by using a locally optimal judgment link and a locally optimal jump-out method.

Description

Technical Field

[0001] The present invention relates to a steel temperature prediction construction method for predicting the temperature of a steel billet discharged from a heating furnace in a metallurgical process, and in particular to a steel temperature prediction construction method based on data drive and improved particle swarm algorithm. Background Art

[0002] In the steel production process, the heating furnace is used to heat the steel billet to a temperature suitable for the next process. Accurately predicting the temperature of the steel billet out of the furnace is crucial to controlling energy consumption and production quality. At present, commonly used prediction methods include heat transfer theory models and data-driven intelligent algorithm models. Although the heat transfer theory model has a clear physical meaning, the model needs to assume multiple parameters due to the complexity of the heat transfer process in the furnace, resulting in low prediction accuracy. The data-driven model uses the actual data collected during the production process and combines machine learning and other technologies to construct a prediction model, avoiding the trouble of parameter assumptions. However, existing data-driven models such as neural networks have certain prediction errors when processing steel temperature prediction, and the accuracy needs to be further improved. Summary of the invention

[0003] In view of the above-mentioned defects existing in the prior art, the present invention provides a steel temperature prediction construction method based on data-driven and improved particle swarm algorithm, which can improve the steel temperature prediction accuracy.

[0004] The present invention is realized by the following technical scheme: a steel temperature prediction construction method based on data drive and improved particle swarm algorithm, which is characterized by comprising the following steps:

[0005] Step S1: collecting data during the production of steel billets, including the material of the steel billets, the temperature of the steel billets entering the furnace, the thickness of the steel billets, the temperature of each section of the heating furnace, and the temperature of the steel billets leaving the furnace; using mathematical methods to perform isometric processing on the above data to obtain data that conforms to the model input format;

[0006] Step S2: Initialize the steel temperature prediction mathematical model using the following formula:

[0007]

[0008] Where, T * is the predicted temperature of the billet out of the furnace, T0 is the temperature of the billet entering the furnace, d is the thickness of the billet, T j is the temperature value of the sampling point in each section of the heating furnace, j is the sampling point number, and n1 is the number of temperature sampling points in the heating furnace; t p is the average temperature value at the burner nozzle of section p in the heating furnace, p is the section number of the heating furnace, m is the number of sections in the heating furnace; θ is the integration coefficient of other factors including heating time, β is the weight parameter of the initial temperature of the steel billet entering the furnace, the weight parameter of the thickness of the steel billet, α is the parameter of the influence of the material factor of the steel billet on the steel temperature, k j is the weight parameter of the temperature at the burner nozzle, a p is the weight parameter of the average value of the temperature at the burner nozzle in the cosine function p section and the time, b p is the weight parameter of the average value of the temperature at the burner nozzle of the sin function section p and the time;

[0009] Step S3: construct the objective function based on the least squares method and regularization idea;

[0010] Step S4: using an improved particle swarm algorithm to solve the objective function to solve the steel temperature prediction model with the optimal parameters. The improvements of the particle swarm algorithm include: a. using a random uniform distribution generation strategy to initialize the particle swarm; b. introducing a chromosome mutation strategy to increase the diversity of the particle swarm when iteratively updating the particle position and velocity in the particle swarm; using a three-stage adaptive change method to update the inertia weight value; c. using a local optimal determination link and a local optimal jump-out method to determine whether the current solution is the local optimal solution;

[0011] Step S5: Use the training data generated by the simulation experiment and the actual measurement data obtained by the actual experimental method to train the prediction model and optimize the parameters to complete the model construction;

[0012] Step S6: Use the constructed model to perform temperature prediction on new billet production data input.

[0013] In the present invention, by using a random uniform distribution generation strategy to initialize a particle swarm, it is possible to ensure that particles are evenly distributed in a search space, which is helpful to increase the search efficiency and global convergence of the algorithm; by introducing a chromosome mutation strategy, the diversity of the particle swarm can be increased, the probability of the particle swarm falling into a local optimum can be reduced, and it is helpful to jump out of the local optimal solution, thereby better searching for a global optimal solution and improving the convergence speed of the algorithm; by using a three-stage adaptive change method to update the inertia weight value, the flexibility of particle flight speed change can be improved; by using a local optimal determination link and a local optimal jump-out method, the problem that a particle swarm optimization algorithm is prone to falling into a local optimum during a search process, resulting in inaccurate final calculation results, can be effectively overcome.

[0014] Furthermore, in step S3, the objective function is the error between the temperature calculation prediction value and the measured value. The objective function is optimized using the regularization idea, and the formula used is:

[0015]

[0016] Where n is the number of arrays in the learning dataset, m is the number of parameters to be calculated, and T i is the actual temperature of the billet, T * i Predict the temperature T of the billet out of the furnace * The discretized value of , λ1 and λ2 are regularization parameters, is the unknown weight parameter value to be determined, and minJ is the minimized theoretical calculated value.

[0017] Furthermore, in step S4, when the particle swarm is initialized using the random uniform distribution generation strategy, a dimension in the particle swarm position matrix is ​​randomly selected to determine the search space range of the problem, that is, to determine the upper and lower limits of the dimension in the randomly selected particle swarm position matrix, and then the values ​​of the corresponding dimensions of each particle are generated in the form of uniform distribution during initialization. The numerical generation formula used is:

[0018]

[0019] In the formula, j is the randomly selected dimension, i is the particle number, N is the size of the particle group, and x ij is the position information after uniform distribution, O max With O min They are the upper and lower limits of the value range of different dimensions respectively.

[0020] Furthermore, in step S4, the following formula (5) is used for judgment when performing the local optimum determination link. If it is determined to be trapped in the local optimum, the new particle position update formula (6) is executed to realize the local optimum jump out:

[0021]

[0022] In the formula, Slope is the slope value of the global optimal fitness function value curve within the range of m consecutive iterations, and m is the autonomous setting value; k (g) is the fitness function value corresponding to the global optimal solution when the number of iterations is k; k-5 is the fitness function value corresponding to the global optimal solution when the number of iterations is k-5;

[0023] The particle position jump calculation formula is:

[0024] x ij (k) = x ij (k)-r1×c1×(gx ij (k))+r2×c2×((bad-x ij (k))) (6)

[0025] In the formula, bad is the location information of the worst point in the world, xij (k) is the position information when the number of iterations is k, r1 and r2 are two uniformly distributed random numbers in the range [0,1), c1 and c2 are learning factor coefficients, k is the number of iterations, and g is the historical optimal position information of the population.

[0026] Furthermore, in step S4, when the inertia weight value is updated by the three-stage adaptive change method, first, in each stage of the update iteration process, the inertia weight value is switched by determining the size of the fitness function value in the algorithm, and the "two-step" is changed to "three-step", and the step switching formula is as follows:

[0027]

[0028] In the formula, f(g) is the fitness function value corresponding to the global optimal solution, Fitness1 and Fitness2 are autonomously set values, and [w s1 , w e1 ],[w s2 , w e2 ][w s3 , w e3 ] are three groups of inertia weight values ​​respectively;

[0029] The switching formula of inertia weight value is:

[0030]

[0031] Where r is a uniformly distributed random number in the range of [0,4), K is the number of iterations, G is the maximum value, ωs and ωe are the initial value and the final value of the inertia weight, respectively, Fitenss3 is the autonomous setting value; ω is the inertia weight value after the inertia weight value is set and adaptively changed in each stage.

[0032] Furthermore, in step S4, when the chromosome mutation strategy is introduced to increase the diversity of the particle swarm, the method adopted is: compare the size between the individual fitness values ​​of the particles with the average fitness function value of the current particle swarm. If the individual fitness value is higher than the average fitness function value, the particle position is updated according to formula (9):

[0033] x * ij (k) = x ij (k)-ω×v ij (k)-ω×(gp i ) (9)

[0034] In the formula, x* ij is the particle position after the mutation of the i-th particle with dimension j, x ijis the position information of the i-th particle with dimension j before mutation, k is the number of iterations of the current algorithm, v ij is the velocity information of the i-th particle with dimension j before the mutation, p i is the individual's historical optimal position information at the current moment, g is the population's historical optimal position information, ω is the inertia weight coefficient, j is the randomly selected dimension, and i is the particle number.

[0035] The beneficial effects of the present invention are as follows: the present invention adopts a multi-strategy collaborative particle swarm optimization algorithm to construct a steel temperature prediction model, which improves the accuracy of the prediction and the generalization ability of the model. Moreover, the method of the present invention has a faster convergence speed while ensuring the prediction accuracy, and can find the global optimal solution more quickly. Moreover, the method of the present invention does not need to model the heat transfer process in the heating furnace and make parameter assumptions during the entire prediction process, thus avoiding related problems. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a schematic diagram of the structure of the heating furnace;

[0037] Figure 2 It is a schematic diagram of heat exchange in a heating furnace;

[0038] Figure 3 It is a schematic diagram of the relationship between the particle group position and velocity update formula;

[0039] Figure 4 It is the algorithm flow chart of the present invention;

[0040] Figure 5 It is a fitness value comparison diagram of the optimal position of the population history of the present invention, in which the horizontal axis is the number of iterations and the vertical axis is the fitness value;

[0041] In the figure, 1. steel billet, 2. burner nozzle. DETAILED DESCRIPTION

[0042] The present invention will be further described below by way of non-limiting embodiments and in conjunction with the accompanying drawings:

[0043] The present invention provides a method for constructing a steel temperature prediction model based on data drive and improved particle swarm optimization (XPSO), aiming to solve the problem of soft measurement of billet out-of-furnace temperature.

[0044] like Figure 1 The figure shows the schematic diagram of the heating furnace structure. When the billet is heated in the heating furnace, it passes through the preheating section, heating section and soaking section in sequence from the furnace entrance to the furnace exit. The heating section can be divided into multiple sections for furnace temperature control, and each section is divided into upper and lower parts for heating.

[0045] like Figure 2The figure shows a schematic diagram of heat exchange in the heating furnace. The heating furnace wall radiates heat to the surface of the steel billet inside it, and the furnace gas in the heating furnace also radiates heat to the steel billet. After the steel billet absorbs heat, its temperature rises and it also begins to transfer some heat to the heating furnace wall and the heating furnace. This is a two-way heat exchange process. The heating furnace wall continuously transfers heat to the steel billet, and the steel billet also radiates and transfers part of the heat to the shell and the external environment.

[0046] A steel temperature prediction construction method based on data-driven and improved particle swarm algorithm comprises the following steps:

[0047] Step S1: collecting data during the production of steel billets, including the material of the steel billets, the temperature of the steel billets entering the furnace, the thickness of the steel billets, the temperature of each section of the heating furnace, and the temperature of the steel billets leaving the furnace; using mathematical methods to perform isometric processing on the above data to obtain data that conforms to the model input format;

[0048] The specific processing method is: use the calculation normalization formula to perform equalization processing on the collected data, and change its length to meet the standard of the input item of the objective function. The normalization formula is as follows:

[0049]

[0050] Among them, y is the collected sample data, y n is the normalized sample data, y max is the maximum value of the sample data, y min is the minimum value of the sample data.

[0051] Step S2: Based on the inspiration of the activation function in the neural network fitting data, a method combining polynomial and Fourier formula is used to initialize the steel temperature prediction mathematical model;

[0052] Step S2.1: Based on the inspiration of the activation function in the neural network fitting data, a mathematical formula combining polynomial function and Fourier function is used as the mathematical expression of the prediction model. The specific formula is shown in formula (2):

[0053]

[0054] Where, T * is the predicted temperature of the billet out of the furnace, T0 is the temperature of the billet entering the furnace, d is the thickness of the billet, T j is the temperature value of the sampling point in each section of the heating furnace, j is the sampling point number, and n1 is the number of temperature sampling points in the heating furnace; t p is the average temperature value at the burner nozzle of section p in the heating furnace, p is the section number of the heating furnace, m is the number of sections in the heating furnace; θ is the integration coefficient of other factors including heating time, β is the weight parameter of the initial temperature of the steel billet entering the furnace, the weight parameter of the thickness of the steel billet, α is the parameter of the influence of the material factor of the steel billet on the steel temperature, k j is the weight parameter of the temperature at the burner nozzle, a p is the weight parameter of the average value of the temperature at the burner nozzle in the cosine function p section and the time, b p It is the weight parameter of the average value of the temperature at the burner nozzle of the sin function p section and the time.

[0055] Step S3: construct an objective function based on the least squares method and regularization idea. The objective function ensures the tracking ability of the model temperature calculation value to the actual value, providing a reliable basis for billet temperature prediction;

[0056] Step S3.1: Determine the objective function as the error between the temperature calculation prediction value and the measured value; use the regularization idea to optimize the objective function, and the formula (3) used is:

[0057]

[0058] Where n is the number of arrays in the learning dataset, m is the number of parameters to be calculated, and T i is the actual temperature of the billet, T * i Predict the temperature T of the billet out of the furnace * The discretized value of , λ1 and λ2 are regularization parameters, is the unknown weight parameter value to be determined, minJ is the minimum theoretical calculated value;

[0059] Regularization processing is added to reduce the possibility of model overfitting and remove the impact of noise in the learning data set, so that the error between the model-calculated predicted value and the measured value is reduced to an allowable range.

[0060] Step S4: Using the improved particle swarm algorithm (XPSO) to solve the objective function, so as to solve the steel temperature prediction model with the optimal parameters. The improvements of the particle swarm algorithm include: a. using the random uniform distribution generation strategy to initialize the particle swarm; b. introducing the chromosome mutation strategy to increase the diversity of the particle swarm when iteratively updating the particle position and speed in the particle swarm; using the three-stage adaptive change method to update the inertia weight value; c. using the local optimal determination link and the local optimal jump-out method to determine whether the current solution is the local optimal solution. The algorithm flow chart of the present invention is shown in the attached figure. Figure 4 As shown, the specific process is:

[0061] Step S4.1: Initialize the particle swarm population size N, dimension D, maximum number of iterations IterM and fitness function f=J;

[0062] Step S4.2: Initialize the particle swarm using a random uniform distribution generation strategy Specifically, a dimension in the particle swarm position matrix is ​​randomly selected to determine the search space range of the problem, that is, to determine the upper and lower limits of this dimension in the randomly selected particle swarm position matrix. Then, during initialization, the values ​​of the corresponding dimensions of each particle are generated in the form of uniform distribution. The numerical generation formula (4) is:

[0063]

[0064] In the formula, j is the randomly selected dimension, i is the particle number, N is the size of the particle group, and x ij is the position information after uniform distribution, O max With O min They are the upper and lower limits of the value range of different dimensions respectively;

[0065] Step S4.3: Calculate the fitness objective function value f(X) of X; record the current position of each particle as the individual historical optimal position information pi at the current moment, and record the position corresponding to the particle with the minimum fitness value selected as the population historical optimal position information g of the particle swarm;

[0066] Step S4.4: Perform local optimal determination. The determination rule is shown in formula (5). During the optimization process, when the slope value Slope of the global optimal fitness function value curve within the range of m consecutive iterations is less than the specified value, it is determined that the particle swarm has the possibility of falling into the local optimal state. The parameter value m is an autonomously set value. If it is determined to be trapped in the local optimal state, the new particle position update formula (6) is executed to realize the local optimal state jump out and go to step S4.3; otherwise, execute step S4.5;

[0067]

[0068] In the formula, Slope is the slope value of the global optimal fitness function value curve within the range of m consecutive iterations, and m is the autonomous setting value; k (g) is the fitness function value corresponding to the global optimal solution when the number of iterations is k; k-5 is the fitness function value corresponding to the global optimal solution when the number of iterations is k-5;

[0069] The particle position jump calculation formula is:

[0070] x ij (k) = x ij (k)-r1×c1×(gx ij (k))+r2×c2×((bad-x ij (k))) (6)

[0071] In the formula, bad is the location information of the worst point in the world, x ij (k) is the position information when the number of iterations is k, r1 and r2 are two uniformly distributed random numbers in the range [0,1), c1 and c2 are learning factor coefficients, k is the number of iterations, and g is the historical optimal position information of the population;

[0072] The local optimal determination link and the local optimal jump-out method can effectively overcome the problem that the particle swarm optimization algorithm is prone to fall into the local optimal state during the search process, resulting in inaccurate final calculation results.

[0073] Step S4.5: Use the three-stage adaptive change method to update the inertia weight ω. The "stage" adaptive idea is mainly used in the process of updating and iterating the inertia weight. Its purpose is to improve the flexibility of the particle flight speed change. The main implementation process is: First, in the update and iteration process of each stage, the inertia weight value is switched by determining the size of the fitness function value in the algorithm. At the same time, in order to solve the problem of poor flexibility of the "two-step" transition and conversion conditions, the "two-step" is changed to "three-step". The formula for step switching is as follows:

[0074]

[0075] In the formula, f(g) is the fitness function value corresponding to the global optimal solution, Fitness1 and Fitness2 are autonomously set values, and [w s1 , w e1 ][w s2 , w e2 ][w s3 , w e3 ] are three sets of inertia weight values, [w s , w e ] is the starting value and ending value of each step. The specific w is determined by formula (8). Each set of values ​​contains the maximum and minimum inertia weight values ​​in its corresponding stage, which are adjusted according to the algorithm solution status in different application backgrounds.

[0076] In order to solve the problem that the algorithm will lose a certain degree of flexibility if the inertia weight value does not change in each stage after being set, further adjustments are made on the basis of the "step-by-step" improvement, so that the inertia weight value can change adaptively in each "step". The adaptive transformation formula is also "step-by-step", and the switching condition is related to the value of the fitness function in the algorithm. The switching formula of the inertia weight value is:

[0077]

[0078] Where r is a uniformly distributed random number in the range [0,4), K is the number of iterations, G is the maximum value, and ω s With ω e are the initial value and final value of the inertia weight of each step, respectively; Fitenss3 is the self-set value; ω is the inertia weight value after the inertia weight value is set and adaptively changed in each stage. Taking "Step 1" as an example, the switching formula of the inertia weight value is:

[0079]

[0080] In the formula, ω s1 With ω e1 are the initial and final values ​​of the inertia weight of “step 1”, respectively, obtained by formula (7).

[0081] Step S4.6: Add chromosome mutation strategy to maintain the diversity of particle swarm, improve the convergence speed of the algorithm and reduce the probability of particle swarm falling into local optimum. Specifically: compare the size of the fitness value of individual particles with the average fitness function value of the current particle swarm. If the individual value is higher than the average value, update the particle position according to formula (9):

[0082] x * ij (k) = x ij (k)-ω×v ij (k)-ω×(gp i ) (9)

[0083] In the formula, x* ij is the particle position after the mutation of the i-th particle with dimension j, x ij is the position information of the i-th particle with dimension j before mutation, k is the number of iterations of the current algorithm, v ij is the velocity information of the i-th particle with dimension j before the mutation, p i is the individual's historical optimal position information at the current moment, g is the population's historical optimal position information, ω is the inertia weight coefficient, j is the randomly selected dimension, and i is the particle number;

[0084] Step S4.7: Update the particle swarm position. The particle swarm position update formula and velocity update formula are as follows: ij (k+1)=ω×v ij (k)+r1×c1×(p i -x ij (k))+r2×c2×(gx ij (k)) (10)

[0085] x ij (k+1)=x ij (k)+v ij(k+1) (11)

[0086] Where ω is the inertia weight coefficient, r1 and r2 are two uniformly distributed random numbers in the range [0,1), c1 and c2 are learning factor coefficients, which are non-negative constants, k is the number of iterations of the current algorithm, and k = 1, 2, 3, ..., v ij (k+1) is the velocity information of the i-th particle with dimension j for the k+1th iteration, x ij (k+1) is the position information of the i-th particle with dimension j after iteration k+1, v ij (k) is the velocity information of the i-th particle with dimension j for k iterations, x ij (k) is the position information of the i-th particle with dimension j for k iterations, pi is the historical optimal position information of the individual at the current moment, and g is the historical optimal position information of the population;

[0087] Step S4.8: Determine whether the loop termination condition is met. If so, construct a steel temperature prediction model based on the parameters corresponding to the particle position. If not, repeat steps S4.3 to S4.7.

[0088] Step S5: using the training data generated by the simulation experiment (according to the existing furnace billet temperature rise mechanism model combined with Matlab software to write a program to generate 1280 sets of simulation data sets, of which 1200 sets are used as learning data sets and the remaining 80 sets are used as test data sets), combined with the improved particle swarm algorithm, the prediction model is trained and parameter optimization is performed, specifically: the parameters of each algorithm are set, and each particle swarm algorithm is repeatedly run independently 10 times, so as to obtain the average value, maximum value, median, variance and standard deviation of the error between the fitting model and the test data set, and the size of the above index data is comprehensively considered to test and evaluate the prediction performance of the model;

[0089] The simulation results are shown in Table 1:

[0090] Table 1:

[0091]

[0092] According to the experimental results, the average value of the model prediction error solved by the improved particle swarm algorithm (XPSO) of the present invention reaches 0.47°C and the median is 0.37°C, which is the smallest compared with other algorithms;

[0093] As attached Figure 5 As shown, compared with the fitness value of the population historical optimal position using other algorithms, the improved particle swarm algorithm (XPSO) of the present invention has a faster search speed in the early stage of iteration and can quickly enter a convergence state.

[0094] Step S6: The actual measurement data of the steel billet is obtained by the actual thermocouple experimental method (a total of 48 sets of simulation data sets, of which 38 sets are used as learning data sets and the remaining 10 sets are used as detection data sets), and the improved particle swarm algorithm is combined for training and parameter optimization to obtain the steel billet temperature prediction result; the actual collected data are shown in Table 2:

[0095] Table 2:

[0096]

[0097]

[0098] Step S7: Compare the predicted result of the steel billet temperature with the actually detected steel billet temperature, and judge the accuracy of the constructed steel temperature prediction model according to the comparison result. When the accuracy requirement is met, the model construction is completed; the error value between the predicted result and the actually detected steel billet temperature is shown in Table 3:

[0099] Table 3:

[0100]

[0101] As can be seen from the table above, the minimum value of the prediction deviation obtained by solving the prediction model by this algorithm is 1.04℃, and the average value is 3.46℃. Since the actual measured temperature of the billet when it comes out of the furnace is between 1150-1250℃, the relative average error rate of the prediction result of this model is only 0.30%. Therefore, this model can accurately predict the temperature.

[0102] Using the constructed model, temperature prediction can be performed by inputting new billet production data.

[0103] The other parts of this embodiment are all prior art and will not be described in detail here.

Claims

1. A steel temperature prediction construction method based on data-driven and improved particle swarm algorithm, characterized by: The steps include: Step S1: collecting data during the production of steel billets, including the material of the steel billets, the temperature of the steel billets entering the furnace, the thickness of the steel billets, the temperature of each section of the heating furnace, and the temperature of the steel billets leaving the furnace; using mathematical methods to perform isometric processing on the above data to obtain data that conforms to the model input format; Step S2: Initialize the steel temperature prediction mathematical model using the following formula: Where, T * is the predicted temperature of the billet out of the furnace, T0 is the temperature of the billet entering the furnace, d is the thickness of the billet, T j is the temperature value of the sampling point in each section of the heating furnace, j is the sampling point number, and n1 is the number of temperature sampling points in the heating furnace; t p is the average temperature value at the burner nozzle of section p in the heating furnace, p is the section number of the heating furnace, and m is the number of sections in the heating furnace; θ is the integration coefficient of other factors including heating time, β is the weight parameter of the initial temperature of the steel billet entering the furnace, the weight parameter of the thickness of the steel billet, α is the parameter of the influence of the material factor of the steel billet on the steel temperature, k j is the weight parameter of the temperature at the burner nozzle, a p is the weight parameter of the average value of the temperature at the burner nozzle in the cosine function p section and the time, b p is the weight parameter of the average value of the temperature at the burner nozzle of the sin function section p and the time; Step S3: construct the objective function based on the least squares method and regularization idea; Step S4: using an improved particle swarm algorithm to solve the objective function to solve the steel temperature prediction model with the optimal parameters. The improvements of the particle swarm algorithm include: a. using a random uniform distribution generation strategy to initialize the particle swarm; b. introducing a chromosome mutation strategy to increase the diversity of the particle swarm when iteratively updating the particle position and velocity in the particle swarm; using a three-stage adaptive change method to update the inertia weight value; c. using a local optimal determination link and a local optimal jump-out method to determine whether the current solution is the local optimal solution; Step S5: Use the training data generated by the simulation experiment and the actual measurement data obtained by the actual experimental method to train the prediction model and optimize the parameters to complete the model construction; Step S6: Use the constructed model to perform temperature prediction on new billet production data input.

2. The method for constructing a steel temperature prediction model based on data-driven and improved particle swarm algorithm according to claim 1 is characterized in that: In step S3, the objective function is the error between the temperature calculation prediction value and the measured value. The objective function is optimized using the regularization idea, and the formula used is: Where n is the number of arrays in the learning dataset, m is the number of parameters to be calculated, and T i is the actual temperature of the billet, T * i Predict the temperature T of the billet out of the furnace * The discretized value of , λ1 and λ2 are regularization parameters, is the unknown weight parameter value to be determined, and minJ is the minimized theoretical calculated value.

3. The method for constructing a steel temperature prediction model based on data-driven and improved particle swarm algorithm according to claim 1 or 2, characterized in that: In step S4, when the particle swarm is initialized using the random uniform distribution generation strategy, a dimension in the particle swarm position matrix is ​​randomly selected to determine the search space range of the problem, that is, to determine the upper and lower limits of the dimension in the randomly selected particle swarm position matrix. Then, during initialization, the values ​​of the corresponding dimensions of each particle are generated in the form of uniform distribution. The numerical generation formula used is: In the formula, j is the randomly selected dimension, i is the particle number, N is the size of the particle group, and x ij is the position information after uniform distribution, O max With O min They are the upper and lower limits of the value range of different dimensions respectively.

4. The method for constructing a steel temperature prediction model based on data drive and improved particle swarm algorithm according to claim 3 is characterized by: In step S4, the following formula (5) is used to judge when performing the local optimum determination link. If it is determined to be trapped in the local optimum, the new particle position update formula (6) is executed to achieve local optimum jump: In the formula, Slope is the slope value of the global optimal fitness function value curve within the range of m consecutive iterations, and m is the autonomous setting value; k (g) is the fitness function value corresponding to the global optimal solution when the number of iterations is k; k-5 is the fitness function value corresponding to the global optimal solution when the number of iterations is k-5; The particle position jump calculation formula is: x ij (k)=x ij (k)-r1×c1×(g-x ij (k))+r2×c2×((bad-x ij (k))) (6) In the formula, bad is the location information of the worst point in the world, x ij (k) is the position information when the number of iterations is k, r1 and r2 are two uniformly distributed random numbers in the range [0,1), c1 and c2 are learning factor coefficients, k is the number of iterations, and g is the historical optimal position information of the population.

5. The method for constructing a steel temperature prediction model based on data drive and improved particle swarm algorithm according to claim 3 is characterized by: In step S4, when the inertia weight value is updated by the three-stage adaptive change method, first, in each stage of the update iteration process, the inertia weight value is switched by determining the size of the fitness function value in the algorithm, and the "two-step" is changed to "three-step". The formula for step switching is as follows: In the formula, f(g) is the fitness function value corresponding to the global optimal solution, Fitness1 and Fitness2 are autonomously set values, and [ω s1 ,ω e1 ][ω s2 ,ω e2 ][ω s3 ,ω e3 ] are three groups of inertia weight values ​​respectively; The switching formula of inertia weight value is: Where r is a uniformly distributed random number in the range [0,4), K is the number of iterations, G is the maximum value, and ω s With ω e are the starting and ending values ​​of the inertia weight of each step, respectively; Fitenss3 is the autonomously set value; ω is the inertia weight value after adaptive changes in each stage after the inertia weight value is set.

6. The method for constructing a steel temperature prediction model based on data drive and improved particle swarm algorithm according to claim 3 is characterized by: In step S4, when the chromosome mutation strategy is introduced to increase the diversity of the particle swarm, the method adopted is: compare the size between the individual fitness values ​​of the particles with the average fitness function value of the current particle swarm. If the individual fitness value is higher than the average fitness function value, the particle position is updated according to formula (9): x * ij (k)=x ij (k)-ω×v ij (k)-ω×(g-p i ) (9) In the formula, x* ij is the particle position after the mutation of the i-th particle with dimension j, x ij is the position information of the i-th particle with dimension j before mutation, k is the number of iterations of the current algorithm, v ij is the velocity information of the i-th particle with dimension j before the mutation, p i is the individual's historical optimal position information at the current moment, g is the population's historical optimal position information, ω is the inertia weight coefficient, j is the randomly selected dimension, and i is the particle number.

Images