Smelting process temperature prediction control method based on improved LVY-GRU network

Abstract

The invention discloses a smelting process temperature prediction control method based on an improved LVY-GRU network, and belongs to the field of prediction control. The method comprises the steps that four types of operation variable data and top-blown furnace temperature data in the tin smelting process are collected and preprocessed; calculating correlation coefficients of the processed data by using a Pearson correlation coefficient method, and eliminating variables with weak correlation by comparing threshold values to construct a data set; using a hedera helix optimization algorithm in a swarm agent algorithm to construct an improved gating value circulation unit algorithm to predict the temperature number of the top-blown furnace so as to realize dynamic modeling; and an improved gating value circulation unit algorithm after the test is calculated through a gradient descent algorithm, an optimal control law is obtained, and temperature control is achieved. According to the improved LVY optimization algorithm, the differential evolution algorithm is used, population differential information is utilized, the potential optimal area is more efficiently pointed, the convergence speed of the algorithm can be increased, the gradient descent method is combined to process the online optimization problem, the optimal control law is obtained, and stable and efficient control over the temperature in the reduction smelting process is achieved.

Description

Technical Field

[0001] The present invention relates to the field of predictive control, and in particular to a method for predictive control of smelting process temperature based on an improved LVY-GRU network. Background Art

[0002] Metals such as tin and copper are important raw materials for industrial development and are widely used in electronics, chemicals, national defense and other fields. Among them, the reduction smelting of metals is a complex process involving high temperature, high pressure, and a highly corrosive environment. In this process, precise temperature control is key to achieving high-quality products and efficient production. In the molten pool, the charge added through the charging port on the furnace top and the fuel, reducing agent, and oxygen added through the spray gun undergo a series of gas-gas, gas-solid, gas-liquid, and liquid-solid reactions. The reaction process within the molten pool is characterized by large inertia, long time lag, slow time variation, and strong nonlinearity. The control methods used in similar industrial control systems mainly include PID control and traditional model predictive control.

[0003] However, while traditional PID control methods have performed well in many industrial control systems, they suffer from poor robustness to parameter variations in highly complex processes such as metal reduction smelting, making precise control difficult. While traditional model predictive control (MPC) can effectively address nonlinearities and time lags, it requires a precise mathematical model based on the actual industrial process or dynamics model as a prediction model. However, the complex nature of the metal reduction smelting process makes it difficult to establish a precise mathematical model, making traditional MPC methods inapplicable. To address the difficulty of MPC modeling, researchers have proposed various MPC methods based on deep learning networks. Among them, GRU networks and their variants have been effectively applied in various process industries due to their unique memory units and excellent time series data prediction performance. Therefore, to address the characteristics of the metal reduction smelting process, such as high inertia, long time lags, slow time variations, and strong nonlinearity, a temperature predictive control method for the reduction smelting process based on an improved LVY-GRU learning network is proposed. Summary of the Invention

[0004] In order to solve the above technical problems, the present invention provides a method for temperature prediction and control of a smelting process based on an improved LVY-GRU network.

[0005] To implement the above technology, the steps are as follows:

[0006] S1. Collect data on four types of operational variables and top-blown furnace temperature during the tin smelting process and preprocess the collected data; calculate the correlation coefficient of the processed data using the Pearson correlation coefficient method; and construct a data set by comparing the obtained correlation coefficients with a threshold value to eliminate variables with weak correlation.

[0007] Here are the steps:

[0008] S1.1. Collect a preset number of batches of data on four types of operational variables during the tin smelting process and top-blown furnace temperature data. The four types of operational variables during the tin smelting process are: coal consumption, oxygen consumption, primary ventilation, and secondary air supply.

[0009] The collection method of the preset batch is: perform data collection operation every 30 seconds;

[0010] S1.2. Use the normalization method to preprocess the collected data to a value with a mean of 0 and a standard deviation of 1. The expression is as follows:

[0011]

[0012] In the formula, X' represents the coal consumption, oxygen consumption, primary ventilation, secondary air supply and temperature data, X' i' It represents the i'th sampling data; represents the average value of each sampling data in the preset batch; σ represents the standard deviation of each sampling data in the preset batch;

[0013] S1.3. For the processed data, the Pearson correlation coefficient method is used to calculate the correlation coefficient with the top-blown furnace temperature. The expression is as follows:

[0014]

[0015] Where, Represents the average value of the i'th sampling data;

[0016] The strong correlation threshold was set to 0.7, the weak correlation threshold was set to 0.3, and the moderate correlation threshold was set to 0.3–0.7. Data with a correlation coefficient below 0.3 were eliminated.

[0017] S2. Based on the Ivy Optimization Algorithm (LVY), an improved Gated Value Recurrent Unit (GRU) algorithm is constructed to predict the temperature of the top-blown furnace. The data set is divided into a training set and a test set according to a preset ratio. The training set is used to train the improved Gated Value Recurrent Unit (GRU) algorithm to obtain the optimal hyperparameters of the improved Gated Value Recurrent Unit (GRU) algorithm. The test set is used to test the performance of the improved Gated Value Recurrent Unit (GRU) algorithm.

[0018] Here are the steps:

[0019] S2.1. Specify the hyperparameters to be optimized for the Gated Recurrent Unit (GRU) algorithm, including the number of hidden layer neurons, learning rate, and regularization coefficient.

[0020] S2.2. Initialize the parameters of the Ivy Optimization Algorithm (LVY), including: the number of ivy stems N; the problem dimension D; the search space (X min ~X max ); maximum number of iterations T; growth step α(t); branch probability p branch ; Attachment radius r attach ;

[0021] The growth step length α(t) is expressed as follows:

[0022]

[0023] Where α0 represents the initial growth step; t represents the current iteration step;

[0024] Attachment radius r attach The expression is as follows:

[0025] r attach (t) = r0·e -λt

[0026] Where r0 represents the initial attachment radius; λ represents the attenuation coefficient;

[0027] Randomly generate N initial solutions X = {X1, X2, ..., X N}, where each solution X i The dimensions are:

[0028] S2.3. After dividing the dataset into training and test sets in a ratio of 8:2, the gated recurrent unit (GRU) algorithm constructed using the training set is trained. A loss function is constructed using the predicted and true temperatures to describe the deviation between the predicted and true temperatures. The root mean square error (RMSE) is set as the objective function. When the objective function is less than a threshold, the iteration is stopped to obtain the minimum error.

[0029] The loss function expression is as follows:

[0030]

[0031] Where y r (t) represents the actual temperature value of the t-th iteration step; y p (t) represents the predicted temperature of the GRU algorithm at the t-th iteration step;

[0032] S2.4. Use the Ivy Optimization Algorithm (LVY) to solve the hyperparameters to be optimized for the Gated Value Recurrent Unit (GRU) algorithm through the growth phase, branching phase, and attachment phase. The steps are as follows:

[0033] S2.4.1, Ivy Optimization Algorithm (LVY) Growth Phase, each solution X i The new solution is generated by perturbation, and the expression is as follows:

[0034] X i new =X i +α(t)·ΔX i

[0035] Where ΔX i represents random disturbance;

[0036] S2.4.2, Ivy Optimization Algorithm (LVY) branching stage, with branch probability p branch Execute the branch operation, the expression is as follows:

[0037] X branch =X rand +β·(X rand -X center )

[0038] Where, X rand represents the current solution randomly selected; β represents the expansion factor; X center represents the population center, and the expression is as follows:

[0039]

[0040] S2.4.3, Ivy Optimization Algorithm (LVY) branching stage, select the differential evolution strategy to generate the mutation vector V i , the expression is as follows:

[0041]

[0042] Where F represents the scaling factor; X best represents the global historical optimal solution; X r1 , X r2 represents the first and second individuals randomly selected from the current population, X r1 ≠X r2 ; Select and retain V by crossover probability CR i or The expression is as follows:

[0043]

[0044] Where, j rand represents the randomly selected dimension; j represents the currently selected dimension; at the same time, the fitness function is used to calculate U i ,like Then use U i As a result of the attachment, otherwise retain Among them, Ui Represents the dimension of the component;

[0045] S2.5. Stop iteration when the objective function is less than the objective function threshold to obtain the optimal hyperparameters; use the test set to test the performance of the gated value recurrent unit (GRU) algorithm using the Ivy optimization algorithm (LVY) and the gated value recurrent unit (GRU) algorithm;

[0046] The termination condition of the traditional LVY algorithm is to stop when the maximum number of iterations is reached. However, in order to improve the calculation speed, the termination condition of the algorithm is improved. The present invention changes the termination condition to terminate the iteration when the objective function (RMSE) is less than the threshold ζ=0.05. If it is met, the optimal parameters are assigned to the GRU and prediction is performed.

[0047] S3. Optimize the gated recurrent unit (GRU) algorithm after calculation and testing through the gradient descent algorithm to obtain the optimal control law, which is used to solve the online optimization problem and realize temperature closed-loop control.

[0048] Here are the steps:

[0049] S3.1. Initialize the improved Gated Value Recurrent Unit (GRU) algorithm prediction control parameters, including: the weight factor a of the first cost function, the weight factor b of the second cost function, the prediction time domain H p and control time domain H u ;

[0050] S3.2. Construct a reference trajectory r(t) to meet the production requirements of slow temperature rise-fast temperature rise-slow temperature rise;

[0051] A reference trajectory is constructed based on the actual process of tin smelting. In order to meet actual production needs, the temperature needs to increase slowly at the beginning and end and faster in the middle. An S-shaped curve is constructed as the reference trajectory. The reference trajectory expression is as follows:

[0052]

[0053] Where x(t) represents the temperature of the t-th iteration step;

[0054] S3.3. Based on the reference trajectory r(t), the improved gated recurrent unit (GRU) algorithm predicts the output, weight factors, and control law changes, constructs a cost function, and sets constraints on the cost function to minimize the deviation between the predicted output and the reference trajectory and the control law change;

[0055] Based on the reference trajectory r(t), the improved gated recurrent unit (GRU) algorithm predicts the output, weight factor, and control law change, and constructs the cost function, which is expressed as follows:

[0056]

[0057] Where r(t) represents the reference trajectory; represents the predicted output of the improved gated recurrent unit (GRU) algorithm at the t-th iteration step; a represents the weight factor of the first cost function; b represents the weight factor of the second cost function; Δu(t) represents the change in the control law, Δu(t) = u(t+1)-u(t), u(t+1) represents the input control law of t+1 iteration steps, and u(t) represents the input control law of t iteration steps;

[0058] Set constraints on the cost function, the expression is as follows:

[0059]

[0060] Where U(t) represents historical input; Y(t) represents historical output; u min Indicates the minimum limit of the predicted input control amount; u max Indicates the maximum limit of the predicted input control amount; Indicates the minimum limit for the predicted output temperature; Indicates the maximum limit of the predicted output temperature;

[0061] S3.4. Use the gradient descent algorithm to minimize the cost function and obtain the control law u(t), which is expressed as follows:

[0062]

[0063] Where λ2 represents the learning rate of gradient descent, and λ2>0. The cost function is used to differentiate the control law, and the expression is as follows:

[0064]

[0065] After sorting, we get the following expression:

[0066]

[0067] To solve the Jacobian matrix Introducing the substitute quantity P, the expression is as follows:

[0068] P=[y(t-1),y(t-2),…,y(tn),u(t-1),u(t-2),…,u(tn)];

[0069] The solution results are as follows:

[0070]

[0071] Where n = 1, 2, ..., H p ,m=1,2,…,H u , m∈n-1;

[0072] The final simplified result is expressed as:

[0073]

[0074] For the GRU network, in,

[0075] S3.5. Reduce the control law calculation process by changing the control time domain;

[0076] By changing the control time domain H u =1, which reduces the amount of calculation in the optimization process without reducing the control performance;

[0077] When the control time domain H u =1, we can get the expression:

[0078]

[0079] Where E(t) represents the control error, g(t) represents the gradient of the control law u(t),

[0080] S3.6. Input the control law obtained in S3.5 into the oxygen control valve for control.

[0081] Beneficial effects of the present invention:

[0082] In order to obtain variables with strong correlation with the top-blown furnace temperature and thus efficiently control the top-blown furnace temperature, the present invention introduces the Pearson correlation coefficient method to analyze the correlation coefficient between the operating variables and the top-blown furnace temperature, and eliminates variables with weak correlation.

[0083] In order to make up for the fact that the traditional model predictive control method cannot establish an accurate mathematical model when facing the metal reduction smelting process with strong nonlinearity and time-varying, the present invention introduces an improved LVY-GRU learning network to fit the metal reduction smelting process model.

[0084] In order to improve the prediction accuracy of LVY-GRU, the present invention proposes an improved LVY optimization algorithm, which uses the differential evolution algorithm to utilize population differential information to more efficiently point to the potential optimal area, thereby improving the convergence speed of the algorithm.

[0085] The present invention uses the trained improved LVY-GRU learning network model as the prediction model of model predictive control, and then combines it with the gradient descent method to process the online optimization problem to obtain the optimal control law, thereby achieving stable and efficient temperature control in the reduction smelting process. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] Figure 1 A flow chart of the steps of the present invention;

[0087] Figure 2 The basic framework diagram of the present invention;

[0088] Figure 3 Temperature prediction results of different models;

[0089] Figure 4 Temperature prediction error plots for different models;

[0090] Figure 5 Control result diagram of different control methods. DETAILED DESCRIPTION

[0091] The present invention is further described in detail below with reference to specific embodiments.

[0092] like Figure 1 and Figure 2 As shown, a method for temperature prediction and control of a smelting process based on an improved LVY-GRU network includes the following steps:

[0093] S1. Collect data on four types of operational variables and top-blown furnace temperature during the tin smelting process and preprocess the collected data; calculate the correlation coefficient of the processed data using the Pearson correlation coefficient method; and construct a data set by comparing the obtained correlation coefficients with a threshold value to eliminate variables with weak correlation.

[0094] Here are the steps:

[0095] S1.1. Collect a preset number of batches of data on four types of operational variables during the tin smelting process and top-blown furnace temperature data. The four types of operational variables during the tin smelting process are: coal consumption, oxygen consumption, primary ventilation, and secondary air supply.

[0096] In this example, the production data of a tin smelting plant is collected, and the four types of operation variable data and top-blowing furnace temperature data of 6 batches (m) of tin smelting process are collected, which are expressed as: Z of 6 batches (m) m×5 =[X m×4 Y m×1 ], where X m×4 are the four types of operating variable data of the tin smelting process, Y m×1 is the temperature data;

[0097] The collection method of the preset batch is: perform data collection operation every 30 seconds;

[0098] S1.2. Use the normalization method to preprocess the collected data to a value with a mean of 0 and a standard deviation of 1. The expression is as follows:

[0099]

[0100] In the formula, X' represents the coal consumption, oxygen consumption, primary ventilation, secondary air supply and temperature data, X' i' It represents the i'th sampling data; represents the average value of each sampling data of 6 batches; σ represents the standard deviation of each sampling data of 6 batches;

[0101] S1.3. For the processed data, the Pearson correlation coefficient method is used to calculate the correlation coefficient with the top-blown furnace temperature. The expression is as follows:

[0102]

[0103] Where, Represents the average value of the i'th sampling data;

[0104] The strong correlation threshold was set to 0.7, the weak correlation threshold was set to 0.3, the moderate correlation threshold was set to 0.3–0.7, and data with a correlation coefficient below 0.3 were eliminated;

[0105] In this embodiment, the data with the highest correlation with the top-blown furnace temperature is: oxygen.

[0106] S2. Based on the Ivy Optimization Algorithm (LVY), an improved Gated Value Recurrent Unit (GRU) algorithm is constructed to predict the temperature of the top-blown furnace. The data set is divided into a training set and a test set according to a preset ratio. The training set is used to train the improved Gated Value Recurrent Unit (GRU) algorithm to obtain the optimal hyperparameters of the improved Gated Value Recurrent Unit (GRU) algorithm. The test set is used to test the performance of the improved Gated Value Recurrent Unit (GRU) algorithm.

[0107] Here are the steps:

[0108] S2.1. Specify the hyperparameters to be optimized for the Gated Recurrent Unit (GRU) algorithm, including the number of hidden layer neurons, learning rate, and regularization coefficient.

[0109] S2.2. Initialize the parameters of the Ivy Optimization Algorithm (LVY), including: the number of ivy stems N; the problem dimension D; the search space (X min ~X max ); maximum number of iterations T; growth step α(t); branch probability p branch ; Attachment radius r attach ;

[0110] In this embodiment, the number of hyperparameters to be optimized is 3, so the problem dimension D is set to 3; too many ivy stems will increase the computational cost, so setting the number of ivy stems N to 6 can cover enough initial solutions to avoid redundant computations caused by an overly large population; the search space is set to X min=0, X max =100; maximum number of iterations T = 1000; branch probability p branch =0.15;

[0111] The growth step length α(t) is expressed as follows:

[0112]

[0113] Where α0 represents the initial growth step; t represents the current iteration step;

[0114] Attachment radius r attach The expression is as follows:

[0115] r attach (t) = r0·e -λt

[0116] Where r0 represents the initial attachment radius; λ represents the attenuation coefficient;

[0117] Randomly generate N initial solutions X = {X1, X2, ..., X N}, where each solution X i The dimensions are:

[0118] S2.3. After dividing the dataset into training and test sets in a ratio of 8:2, the gated recurrent unit (GRU) algorithm constructed using the training set is trained. A loss function is constructed using the predicted and true temperatures to describe the deviation between the predicted and true temperatures. The root mean square error (RMSE) is set as the objective function. When the objective function is less than a threshold, the iteration is stopped to obtain the minimum error.

[0119] The loss function expression is as follows:

[0120]

[0121] Where y r (t) represents the actual temperature value of the t-th iteration step; y p (t) represents the predicted temperature of the GRU algorithm at the t-th iteration step;

[0122] S2.4. Use the Ivy Optimization Algorithm (LVY) to solve the hyperparameters to be optimized for the Gated Value Recurrent Unit (GRU) algorithm through the growth phase, branching phase, and attachment phase. The steps are as follows:

[0123] S2.4.1, Ivy Optimization Algorithm (LVY) Growth Phase, each solution X i The new solution is generated by perturbation, and the expression is as follows:

[0124] X i new =X i +α(t)·ΔX i

[0125] Where ΔX i represents random disturbance, and Gaussian disturbance is selected in this embodiment;

[0126] S2.4.2, Ivy Optimization Algorithm (LVY) branching stage, with branch probability p branch Execute the branch operation, the expression is as follows:

[0127] X branch =X rand +β·(X rand -X center )

[0128] Where, X rand represents the current solution randomly selected; β represents the expansion factor, in this embodiment, β = 1.5; X center represents the population center, and the expression is as follows:

[0129]

[0130] S2.4.3, Ivy Optimization Algorithm (LVY) branching stage, select the differential evolution strategy to generate the mutation vector V i , the expression is as follows:

[0131]

[0132] Wherein, F represents the scaling factor. In this embodiment, F=0.7; X best represents the global historical optimal solution; X r1 , X r2 represents the first and second individuals randomly selected from the current population, X r1 ≠X r2 ; Select and retain V by crossover probability CR i or The expression is as follows:

[0133]

[0134] Where, j rand represents the randomly selected dimension; j represents the currently selected dimension; at the same time, the fitness function is used to calculate U i ,like Then use U i As a result of the attachment, otherwise retain Among them, U i Represents the dimension of the component;

[0135] S2.5. Stop iteration when the objective function is less than the objective function threshold to obtain the optimal hyperparameters; use the test set to test the performance of the gated value recurrent unit (GRU) algorithm using the Ivy optimization algorithm (LVY) and the gated value recurrent unit (GRU) algorithm;

[0136] The traditional LVY algorithm terminates when the maximum number of iterations is reached. However, in order to improve the operation speed, the termination condition of the algorithm is improved. The present invention changes the termination condition to terminate the iteration when the objective function (RMSE) is less than the threshold ζ = 0.05. If the condition is met, the optimal parameters are assigned to the GRU and prediction is performed. In this embodiment, the optimal hyperparameter combination is shown in Table 1.

[0137] Table 1: GRU learning network hyperparameter values

[0138]

[0139] The comparison of prediction performance of different models is shown in Table 2; the temperature prediction results of different models are shown in Table 2. Figure 3 The temperature prediction errors of different models are shown in Figure 4 shown; from Figure 3 It can be seen that compared with the GRU model, the actual predicted temperature of the present invention is closer to the actual value; Figure 4 It can be seen that the error range of the present invention is smaller than that of the GRU model, which shows that the prediction results of the present invention are more accurate.

[0140] Table 2 Comparison of prediction performance of different models

[0141]

[0142]

[0143] As can be seen from Table 2, the gated value recurrent unit (GRU) algorithm after the Ivy optimization algorithm (LVY) has a high determination coefficient (R 2 ), mean square error (MSE) and mean absolute error (MAE) are all better than the gated recurrent unit (GRU) algorithm.

[0144] S3. Optimize the gated recurrent unit (GRU) algorithm after calculation and testing through the gradient descent algorithm to obtain the optimal control law, which is used to solve the online optimization problem and realize temperature closed-loop control.

[0145] Here are the steps:

[0146] S3.1. Initialize the improved Gated Value Recurrent Unit (GRU) algorithm prediction control parameters, including: the weight factor a of the first cost function, the weight factor b of the second cost function, the prediction time domain H p and control time domain H u;

[0147] In this embodiment, the weight factor a of the first cost function is 1; the weight factor b of the second cost function is 0.1; the prediction time domain H p =4; control time domain H u =3;

[0148] S3.2. Construct a reference trajectory r(t) to meet the production requirements of slow temperature rise-fast temperature rise-slow temperature rise;

[0149] A reference trajectory is constructed based on the actual process of tin smelting. In order to meet actual production needs, the temperature needs to increase slowly at the beginning and end and faster in the middle. An S-shaped curve is constructed as the reference trajectory. The reference trajectory expression is as follows:

[0150]

[0151] Where x(t) represents the temperature of the t-th iteration step;

[0152] S3.3. Based on the reference trajectory r(t), the improved gated recurrent unit (GRU) algorithm predicts the output, weight factors, and control law changes, constructs a cost function, and sets constraints on the cost function to minimize the deviation between the predicted output and the reference trajectory and the control law change;

[0153] Based on the reference trajectory r(t), the improved gated recurrent unit (GRU) algorithm predicts the output, weight factor, and control law change, and constructs the cost function, which is expressed as follows:

[0154]

[0155] Where r(t) represents the reference trajectory; represents the predicted output of the improved gated recurrent unit (GRU) algorithm at the t-th iteration step; a represents the weight factor of the first cost function; b represents the weight factor of the second cost function; Δu(t) represents the change in the control law, Δu(t) = u(t+1)-u(t), u(t+1) represents the input control law of t+1 iteration steps, and u(t) represents the input control law of t iteration steps;

[0156] Set constraints on the cost function, the expression is as follows:

[0157]

[0158] Where U(t) represents historical input; Y(t) represents historical output; u min Indicates the minimum limit of the predicted input control amount; u max Indicates the maximum limit of the predicted input control amount; Indicates the minimum limit for the predicted output temperature; Indicates the maximum limit of the predicted output temperature;

[0159] In this embodiment, u min =0%,u max =100%,

[0160] S3.4. Use the gradient descent algorithm to minimize the cost function and obtain the control law u(t), which is expressed as follows:

[0161]

[0162] Where λ2 represents the learning rate of gradient descent, and λ2>0. The cost function is used to differentiate the control law, and the expression is as follows:

[0163]

[0164] After sorting, we get the following expression:

[0165]

[0166] To solve the Jacobian matrix Introducing the substitute quantity P, the expression is as follows:

[0167] P=[y(t-1),y(t-2),…,y(tn),u(t-1),u(t-2),…,u(tn)];

[0168] The solution results are as follows:

[0169]

[0170] Where n = 1, 2, ..., H p ,m=1,2,…,H u , m∈n-1;

[0171] The final simplified result is expressed as:

[0172]

[0173] For the GRU network, in,

[0174] S3.5. Reduce the control law calculation process by changing the control time domain;

[0175] By changing the control time domain H u =1, which reduces the amount of calculation in the optimization process without reducing the control performance;

[0176] When the control time domain H u =1, we can get the expression:

[0177]

[0178] Where E(t) represents the control error, g(t) represents the gradient of the control law u(t),

[0179] S3.6. Input the control law obtained in S3.5 into the oxygen control valve for control.

[0180] The present invention and GRU were simultaneously tested for performance, and the performance indicators selected included the integral absolute error index IAE and the integral square error index ISE; the prediction results are shown in Table 3;

[0181] Table 3 Comparison of control performance of different control methods

[0182]

[0183] The present invention and GRU simultaneously carried out temperature control, and the results were as follows Figure 5 As shown, from Figure 5 It can be seen that the curvature and value of the present invention are closer to the reference curve, which shows that the temperature control of the present invention is more precise and the effect is better.

[0184] It should be noted that the above are only preferred embodiments of the present application and do not limit the scope of patent protection of the present application. Any equivalent structure or equivalent process transformation made using the contents of the description and drawings of this application, or directly or indirectly applied in other related technical fields, are also included in the scope of patent protection of the present application.

Claims

1. A method for temperature prediction and control of a smelting process based on an improved LVY-GRU network, characterized in that: The following steps are involved: S1. Collect data on four types of operational variables and top-blown furnace temperature during the tin smelting process and preprocess the collected data; calculate the correlation coefficient of the processed data using the Pearson correlation coefficient method; and construct a data set by comparing the obtained correlation coefficients with a threshold value to eliminate variables with weak correlation. S2. Based on the Ivy optimization algorithm, an improved gated value cycle unit algorithm is constructed to predict the temperature of the top-blown furnace and dynamically model the system; The data set is divided into a training set and a test set according to a preset ratio. The training set is used to train the improved gated valued recurrent unit algorithm to obtain the optimal hyperparameters of the improved gated valued recurrent unit algorithm; the test set is used to test the performance of the improved gated valued recurrent unit algorithm. S3. The improved gated value recurrent unit algorithm after optimization calculation and testing is obtained through the gradient descent algorithm to obtain the optimal control law, which is used to solve the online optimization problem and realize temperature control.

2. The method for temperature prediction and control of a smelting process based on an improved LVY-GRU network according to claim 1, characterized in that: The steps of collecting four types of operating variable data and top-blown furnace temperature data of the tin smelting process and preprocessing the collected data are as follows: calculating the correlation coefficient of the processed data using the Pearson correlation coefficient method; and eliminating variables with weak correlation by comparing the obtained correlation coefficient with a threshold value. The steps of constructing the data set are as follows: S1.

1. Collect a preset number of batches of data on four types of operational variables during the tin smelting process and top-blown furnace temperature data. The four types of operational variables during the tin smelting process are: coal consumption, oxygen consumption, primary ventilation, and secondary air supply. S1.

2. Use the normalization method to preprocess the collected data to a value with a mean of 0 and a standard deviation of 1; S1.

3. Use the Pearson correlation coefficient method to calculate the correlation coefficient between the processed data and the top-blown furnace temperature. Set the strong correlation threshold to 0.7, the weak correlation threshold to 0.3, and the medium correlation threshold to 0.3-0.

7. Eliminate data with a correlation coefficient below 0.

3.

3. The method for temperature prediction and control of a smelting process based on an improved LVY-GRU network according to claim 1, characterized in that: The improved gated value cyclic unit algorithm is constructed based on the ivy optimization algorithm to predict the temperature of the top-blown furnace and realize dynamic modeling; the data set is divided into a training set and a test set according to a preset ratio, the training set is used to train the improved gated value cyclic unit algorithm to obtain the optimal hyperparameters of the improved gated value cyclic unit algorithm; the test set is used to test the performance of the improved gated value cyclic unit algorithm. The steps are as follows: S2.

1. Specify the hyperparameters to be optimized for the gated recurrent unit algorithm, including the number of hidden layer neurons, learning rate, and regularization coefficient. S2.

2. Initialize the parameters of the ivy optimization algorithm, including: number of ivy stems; problem dimension; search space; maximum number of iterations; growth step size; branching probability; and attachment radius. S2.

3. After dividing the dataset into a training set and a test set according to a preset ratio, the gated recurrent unit algorithm constructed using the training set is trained; Use the predicted temperature and the actual temperature to construct a loss value function to describe the deviation between the predicted temperature and the actual temperature; At the same time, the root mean square error is set as the objective function, and the iteration is stopped when the objective function is less than the threshold to obtain the minimum error; S2.

4. Use the Ivy optimization algorithm to solve the hyperparameters to be optimized of the gated recurrent unit algorithm through the growth phase, branching phase, and attachment phase; S2.

5. Stop iteration when the objective function is less than the objective function threshold to obtain the optimal hyperparameters; use the test set to test the performance of the gated value recurrent unit algorithm using the Ivy optimization algorithm and the gated value recurrent unit algorithm.

4. The method for temperature prediction and control of a smelting process based on an improved LVY-GRU network according to claim 1, characterized in that: The improved gated value recurrent unit algorithm after the gradient descent algorithm optimization calculation test is used to obtain the optimal control law for processing the online optimization problem. The steps for achieving temperature control are as follows: S3.

1. Initialize the improved gated value recurrent unit algorithm prediction control parameters, including: the weight factor of the first cost function, the weight factor of the second cost function, the prediction time domain, and the control time domain; S3.

2. Construct a reference trajectory to meet the production requirements of slow temperature rise, fast temperature rise, and slow temperature rise. The reference trajectory expression is as follows: Where x(t) represents the temperature of the t-th iteration step; S3.

3. Based on the reference trajectory, the improved gated value recurrent unit algorithm predicts the output, weight factors, and control law changes, constructs a cost function, and sets constraints on the cost function to minimize the deviation between the predicted output and the reference trajectory and the control law change; The expression of the cost function is as follows: Where r(t) represents the reference trajectory; represents the predicted output of the improved gated value recurrent unit algorithm under the t-th iteration step; a represents the weight factor of the first cost function; b represents the weight factor of the second cost function; Δu(t) represents the change in the control law, Δu(t) = u(t+1)-u(t), u(t+1) represents the input control law of t+1 iteration steps, and u(t) represents the input control law of t iteration steps; The expression to set the constraint is as follows: Where U(t) represents historical input; Y(t) represents historical output; u min Indicates the minimum limit of the predicted input control amount; u max Indicates the maximum limit of the predicted input control amount; Indicates the minimum limit for the predicted output temperature; Indicates the maximum limit of the predicted output temperature; S3.

4. Use the gradient descent algorithm to minimize the cost function and obtain the control law u(t); S3.

5. Reduce the control law calculation process by changing the control time domain; The way to change the control time domain is: set the control time domain to a fixed value; S3.

6. Input the control law obtained in S3.5 into the oxygen control valve for control.

5. The method for temperature prediction and control of a smelting process based on an improved LVY-GRU network according to claim 3, characterized in that: The steps of using the Ivy optimization algorithm to solve the hyperparameters to be optimized of the gated valued recurrent unit algorithm through the growth phase, the branching phase, and the attachment phase are as follows: S2.4.1, Ivy optimization algorithm growth stage, each solution X i The new solution is generated by perturbation, and the expression is as follows: X i new =X i +α(t)·ΔX i Where ΔX i represents random perturbation; α(t) represents the growth step length; S2.4.

2. In the branching phase of the Ivy Optimization Algorithm, branching operations are performed based on branching probabilities, which are expressed as follows: X branch =X rand +β·(X rand -X center ) Where, X rand represents the current solution randomly selected; β represents the expansion factor; X center represents the population center, and the expression is as follows: S2.4.

3. In the branching phase of the Ivy optimization algorithm, the differential evolution strategy is selected to generate the mutation vector V i , the expression is as follows: Where F represents the scaling factor; X best represents the global historical optimal solution; X r1 , X r2 represents the first and second individuals randomly selected from the current population, X r1 ≠X r2 ; Select and retain V by crossover probability CR i or The expression is as follows: Where, j rand represents the randomly selected dimension; j represents the currently selected dimension; at the same time, the fitness function is used to calculate U i ,like Then use U i As a result of the attachment, otherwise retain Among them, U i Indicates the dimension of the component.

Images