A digital heating furnace simulation method based on a neural network model

Abstract

The present application belongs to the technical field of billet heating furnace system simulation, and discloses a digital heating furnace simulation method based on a neural network model. The neural network uses actual production data of the heating furnace as input training samples and uses furnace temperature as prediction output to obtain a neural network model of the relationship between measurable variables in the heating furnace and the furnace temperature. The furnace temperature prediction of the neural network model simulates the actual furnace temperature of the heating furnace, and is combined with a two-level intelligent control system to construct a digital heating furnace simulation platform. The digital heating furnace simulation platform simulates the actual production environment of the heating furnace, provides an experimental environment for the control algorithm of the heating furnace system, reduces the project development cycle, and saves the project implementation cost.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of billet heating furnace system simulation, and relates to a digital heating furnace simulation method based on a neural network model. BACKGROUND

[0002] Since the heating furnace system is a strongly coupled, large time-delay, multi-input and multi-output nonlinear complex system, research on heating furnace temperature control is of great significance to steel production. SUMMARY

[0003] In order to overcome the shortcomings of the prior art, the purpose of the present application is to provide a digital heating furnace simulation method based on a neural network model, which can be used to experiment on the designed furnace temperature or steel temperature control algorithm in actual heating furnace engineering project development, and verify the reliability of the control algorithm. In this way, the feasibility of the designed control algorithm can be analyzed before the actual project is implemented, thereby reducing the project development cycle, reducing the time for on-site debugging, and saving project costs.

[0004] The above purpose of the present application is achieved by the following technical solutions:

[0005] A digital heating furnace simulation method based on a neural network model uses a neural network model to train actual production measurable variable data of the heating furnace as input samples, and takes furnace temperature as a predicted output result; the input samples are the temperature of the current furnace section, the temperature of the previous furnace section, the temperature of the next furnace section, the gas flow of the current furnace section, the air flow of the current furnace section, the total volume of the steel billets in the current furnace section, and the heat content of the steel billets in the current furnace section.

[0006] The heating furnace is a nonlinear system with multiple variables and mutual coupling, and an LSTM neural network is used to model the furnace temperature; since there are seven heating sections in the heating furnace, a neural network model is established for one of the heating sections.

[0007] The mathematical model relationship between the furnace temperature and other variables is as follows:

[0008] In formula (1.1), the parameters are: T z (k) represents the output result of the neural network prediction of the current furnace section at time k, T represents temperature (℃), z represents the furnace section of the heating furnace, z∈{1, 2, 3, 4, 5, 6, 7}, k represents a certain discrete time, (k-1), T z (k-1), T z-1 (k-1), T z+1 (k-1) represents the furnace temperature at time k-1 of the current furnace section, the previous furnace section, and the next furnace section; Gz (k-1) represents the gas flow at time k-1; A z (k-1) represents the air flow at time k-1; A represents the total heat content of the billets in the furnace section at time k-1, represents the volume of each billet, T n represents the temperature of each billet; represents the total volume of the billets in the furnace section at time k, represents the volume of each billet;

[0009] f(·) is a nonlinear function, and n u , n k is the order of the system input and output, R is a set of real numbers, L 2 (R n ) is a Hilbert space.

[0010] An LSTM neural network is used to establish a furnace temperature prediction model.

[0011] The training data sample of the LSTM neural network is the data of the variables in formula (1.1) at time k.

[0012] The input of the neural network is:

[0013]

[0014] The output of the network is: T z (k).

[0015] The activation function is selected as the ReLU function, which is a nonlinear activation function.

[0016] The gradient optimization algorithm uses the Adam algorithm. The formula of the Adam algorithm is as follows: m t = μm t-1 +(1-β1)g t , m t is the first moment estimation of the gradient; n t is the second moment estimation of the gradient; μ, v∈[0, 1), which are the exponential decay rates of the moment estimation; is the correction of m t ; is the correction of n t ; η is the learning rate, Δθ t is the updated parameter, g t is the gradient of the time step, and ε is the error rate.

[0017] Data preprocessing. First, the data is preprocessed, mainly including outlier detection and filtering. Because there are multiple variable characteristic attributes in the heating furnace data, the maximum value and the minimum value of each variable are not the same, and the normalization is added when the model is constructed, which can increase the learning ability of the model to a certain extent, improve the model precision and speed up the convergence speed.

[0018] The data set is linearly scaled to the range [0, 1] by using linear transformation, Where x min and x max are the minimum and maximum values of the data set respectively. The main advantage of scaling is to avoid the dominance of attributes in a larger numerical range over attributes in a smaller numerical range.

[0019] The number of hidden layers in the neural network is selected as 2, the number of neurons in each hidden layer is 380, the learning rate is 0.01, the number of iterations for neural network training is 500 times, and the gradient descent optimization algorithm is selected as Adam algorithm.

[0020] Performance evaluation. There are several methods that can be used to measure the accuracy of time series prediction models. For this kind of prediction, the prediction accuracy is tested by calculating three commonly used evaluation indicators: root mean square error (RMSE), mean absolute percentage error (MAPE) and mean absolute error (MAE). The expressions of MAPE, RMSE and MAE are: Where and y t are the predicted value and the actual value respectively; k is the number of predictions.

[0021] If RMSE < 0.6, MAPE < 0.05 and MAE < 0.5, the training model is successful.

[0022] The beneficial effects of the present application compared with the prior art are:

[0023] The neural network uses the actual production data of the heating furnace as the input training sample, uses the furnace temperature as the predicted output, obtains the neural network model of the measurable variable in the heating furnace and the relationship between the furnace temperature, uses the furnace temperature prediction of the neural network model to simulate the actual furnace temperature of the heating furnace, combines with the traditional secondary intelligent control system of the steel plant, and constructs a digital heating furnace simulation platform. The digital heating furnace simulation platform is used to simulate the actual production environment of the heating furnace, provide an experimental environment for the control algorithm of the heating furnace system, reduce the project development cycle, and save the project implementation cost. BRIEF DESCRIPTION OF DRAWINGS

[0024] The present application will be further described below in combination with the drawings and examples:

[0025] Figure 1 It is a schematic diagram of the LSTM neural network furnace temperature prediction model.

[0026] Figure 2 Furnace temperature training fitting graph for heating 1st stage in heating furnace.

[0027] Figure 3 Furnace temperature prediction curve graph for heating 1st stage in heating furnace.

[0028] Figure 4 Neural network-based heating furnace simulation system block diagram.

[0029] Figure 5 Actual temperature and predicted temperature curve graph for furnace temperature in example 1. DETAILED DESCRIPTION

[0030] The application will be described in detail below through specific examples, but the protection scope of the application is not limited. Unless otherwise specified, the experimental methods used in the application are conventional methods, and the experimental equipment, materials, reagents, etc. used can be obtained from commercial channels.

[0031] Example 1

[0032] As shown in Figures 1-4 , a neural network-based digital heating furnace simulation method is provided, which uses a neural network model to train the actual production measurable variable data of the heating furnace as an input sample, and takes the furnace temperature as the predicted output result. The input sample is the temperature of the current furnace section, the temperature of the previous furnace section, the temperature of the next furnace section, the gas flow of the current furnace section, the air flow of the current furnace section, the total volume of the current furnace section, and the heat content of the current furnace section.

[0033] The furnace temperature data of the heating furnace is collected in time series, and an LSTM neural network is used to predict the furnace temperature of the heating furnace. The training data is actual field data, which can almost cover the main working conditions of the field. The furnace temperature in the ring heating furnace is taken as the predicted output result, and the previous and next section temperatures, the current section temperature, the gas flow, the air flow, the section billet volume, and the section billet temperature are taken as the input data for model training. The model is trained using the PyTorch architecture, the graphics card is GeForce RTX2060, and the CPU is intel Core i7-9700F. The training set has 20000 groups of data, the sampling period of the data is 10s, the test data amount is 360, and the verification set is 360.

[0034] The heating furnace is a multi-variable and mutually coupled nonlinear system, and an LSTM neural network is used to model the furnace temperature of the heating furnace. Since there are 7 heating sections in the heating furnace, a neural network model is established for one of the heating sections.

[0035] The mathematical model relationship between the furnace temperature and other variables is:

[0036] The parameters in formula (1.1) are: T z (k) represents the output result predicted by the neural network of the furnace section at time k, T represents temperature (℃), z represents the heating furnace section, z∈{1, 2, 3, 4, 5, 6, 7}, k represents a discrete time, and The following is the same; T z (k-1), T z-1 (k-1), T z+1 (k-1) represents the furnace temperature at time k-1 of the current furnace section, the previous furnace section and the next furnace section; G z (k-1) represents the gas flow at time k-1; A z (k-1) represents the air flow at time k; (k-1) represents the total heat content of the billets in the current furnace section at time k-1, V represents the volume of each billet, T n represents the temperature of each billet; (k) represents the total volume of the billets in the current furnace section at time k, V represents the volume of each billet;

[0037] f(·) is a nonlinear function, and n u , n k is the order of system input and output, R is a set of real numbers, L 2 (R n ) is a Hilbert space.

[0038] The furnace temperature prediction model is established by using an LSTM neural network.

[0039] The training data sample of the LSTM neural network is the data of the variables in formula (1.1) at time k.

[0040] The input of the neural network is:

[0041]

[0042] The output of the network is: T z (k).

[0043] The activation function is selected as a ReLU function, which is a nonlinear activation function.

[0044] The gradient optimization algorithm adopts an Adam algorithm. The formula of the Adam algorithm is as follows: m t =μm t-1 +(1-μ)g t , m tis the first moment estimation of the gradient; n t is the second moment estimation of the gradient; μ, v ∈ [0, 1) are the exponential decay rates of the moment estimations; is the correction of m t is the correction of n t η is the learning rate, Δθ t is the updated parameter, g t is the gradient of the time step, ε is the error rate.

[0045] Data preprocessing. First, the data is preprocessed, mainly including outlier detection and filtering. Since there are multiple variable characteristic attributes in the furnace data, the maximum and minimum values of each variable are different. When building the model, adding normalization can increase the learning ability of the model to a certain extent, improve the model precision, and speed up the convergence speed.

[0046] The linear transformation is used to linearly scale the data set to the range [0, 1], where x min and x max are the minimum and maximum values of the data set, respectively. The main advantage of scaling is to avoid attributes in a larger numerical range dominating attributes in a smaller numerical range.

[0047] The number of hidden layers of the neural network is selected as 2, the number of neurons in each hidden layer is 380, the learning rate is 0.01, the number of training iterations of the neural network is 500 times, and the gradient descent optimization algorithm is selected as Adam algorithm.

[0048] Performance evaluation. There are several methods that can be used to measure the accuracy of time series prediction models. For this type of prediction, the prediction accuracy is tested by calculating three commonly used evaluation indicators: root mean square error (RMSE), mean absolute percentage error (MAPE), and mean absolute error (MAE). The expressions of MAPE, RMSE, and MAE are: where and y t are the predicted and actual values, respectively; k is the number of predictions.

[0049] If RMSE < 0.6, MAPE < 0.05, and MAE < 0.5, the training model is successful.

[0050] Experimental results: Figure 2 is the fitting curve of the training process of the furnace temperature in heating section 1. It can be seen that the predicted furnace temperature and the actual value fit very well. Figure 3 ​In order to heat the first section of the LSTM-NN (NN is the abbreviation of Neural Network) furnace temperature prediction graph, the number of prediction samples is 360. The blue curve is the actual value, and the red curve is the furnace temperature prediction value. It can be seen that each prediction value curve has no special large dispersion degree from the actual value curve, and the prediction value is almost close to the true value.

[0051] According to the furnace temperature prediction accuracy verification data of the first section of the LSTM-NN heating, the minimum value of the root mean square error RMSE is 0.5551, the minimum value of the mean absolute error MAE is 0.4429, and the minimum value of the mean absolute error MAPR is 0.035, which indicates that the error between the prediction value and the actual value is very small, the accuracy is very high, and the expected furnace temperature prediction result is achieved.

[0052] In the furnace temperature prediction of the first heating section, the true furnace temperature curve is oscillating, and the LSTM-NN can also well predict and follow without overfitting. In the furnace temperature prediction of the second to fourth heating sections, the steep peak in the middle is the temperature rising state, and the LSTM-NN can also accurately predict according to the trend of the curve. In summary, the LSTM-NN can achieve good prediction results for the furnace temperature.

[0053] The working principle and use process of the present application are as follows: Figure 4 The heating furnace simulation system block diagram based on the neural network is first installed with the heating furnace two-level intelligent control system, the production plan is input, and the two-level intelligent control system is run. The two-level intelligent control system will monitor the section furnace temperature, section billet temperature, section billet quantity, section billet total volume, section billet total heat content, coal gas flow, air flow, running speed and the like of the heating furnace, and save to the database in real time with a sampling time of 1s. The two-level intelligent control system has a steel temperature field model calculation module to calculate the billet temperature in real time.

[0054] Load the trained neural network furnace temperature prediction model based on the Pytorch development platform, access the data of the current furnace section temperature, the previous furnace section temperature, the next furnace section temperature, the current furnace section coal gas flow, the current furnace section air flow, the current furnace section billet total volume, the current furnace section billet heat content and the like in the database as the input of the neural network model, and predict the furnace temperature and save to the database. The two-level intelligent control system reads the furnace temperature value predicted by the neural network model in the database in real time, and displays it on the screen.

[0055] When the length or diameter of the billets in the single section of the secondary intelligent control system is large, the volume and total heat content of the billets become large, and the sampling data at this time is saved to the database by the secondary intelligent control system. The neural network model reads the data in the database at this time, and the predicted furnace temperature becomes small, which is consistent with the expected increase of the furnace temperature in the heating furnace.

[0056] When the running speed of the secondary intelligent control system is increased, the heating time of the heated billets in the heating furnace simulation system becomes small, the acceleration of the steel temperature rise becomes low, and the total heat content of the section billets rises slowly. At this time, the sampled data is saved to the database, and after the neural network model reads the data in the database, the predicted furnace temperature is increased compared with the previous time, which is consistent with the expected increase of the furnace temperature in the actual heating furnace. The neural network model takes the data sampled by the secondary intelligent control system as input, and the predicted furnace temperature result is consistent with the actual working condition of the heating furnace.

[0057] When the secondary intelligent control system increases the gas flow and air flow, the section furnace temperature rises: when the secondary system reduces the gas flow and air flow, the section furnace temperature decreases, and the predicted furnace temperature result is consistent with the actual working condition of the heating furnace.

[0058] Example 1

[0059] The method provided in Example 1 is applied to the furnace temperature prediction of the heating furnace of a steel plant in Jiangsu. As shown in Figure 5 , it is a curve diagram of the actual furnace temperature and the predicted furnace temperature of the heating furnace. The predicted furnace temperature curve in the figure fits very well with the actual furnace temperature curve, and the temperature difference between the predicted furnace temperature and the actual furnace temperature is ≤±1℃, achieving a very good prediction effect.

[0060] The above-described embodiments are only preferred embodiments of the present application, and not all the embodiments that can be implemented by the present application. Any obvious modifications made by those skilled in the art without departing from the principles and spirit of the present application should be considered to be included in the protection scope of the claims of the present application.

Claims

1. A digital heating furnace simulation method based on a neural network model, characterized in that, The neural network model is trained using measurable variable data from actual production in the heating furnace as input samples, with furnace temperature as the predicted output. The input samples are: temperature of the current furnace section, temperature of the previous furnace section, temperature of the next furnace section, gas flow rate of the current furnace section, air flow rate of the current furnace section, total volume of steel billet in the current furnace section, and calorific value of steel billet in the current furnace section. The neural network model is trained using measurable variable data from actual production in the heating furnace as input samples, with furnace temperature as the predicted output. The input samples are: temperature of the current furnace section, temperature of the previous furnace section, temperature of the next furnace section, gas flow rate of the current furnace section, air flow rate of the current furnace section, total volume of steel billet in the current furnace section, and calorific value of steel billet in the current furnace section. The heating furnace is a nonlinear system with multiple variables influencing and coupled with each other. An LSTM neural network is used to model the furnace temperature. Since there are a total of 7 heating zones in the heating furnace, a neural network model is now performed on one of the heating zones. The mathematical model relationship between furnace temperature and other variables is as follows: The parameter in equation (1.1) is: T z (k) represents the furnace temperature predicted by the neural network for this furnace section at time k, where T represents temperature (°C), z represents the furnace section, z∈{1,2,3,4,5,6,7}, and k represents a discrete time. The same applies to the following; T z (k-1), T z-1 (k-1), T z+1 (k-1) represents the furnace temperature at time k-1 for the current furnace section, the previous furnace section, and the next furnace section; G z (k-1) represents the gas flow rate at time k-1; A z (k-1) represents the airflow rate at time k; This represents the total heat content of the steel billets in this furnace section at time k-1. T represents the volume of each steel billet. n To indicate the temperature of each steel billet; This represents the total volume of steel billets in this furnace section at time k. This indicates the volume of each steel billet; f(·) is a nonlinear function, and n u ,n k Let L be the order of the system input and output, R be the set of real numbers, and L be the order of the system input and output. 2 (R n ) is a Hilbert space; A furnace temperature prediction model was established using an LSTM neural network. The training data samples for the LSTM neural network are the data of the variables in equation (1.1) at time k; The input to the neural network is: The network output is: T z (k); The activation function should be the ReLU function. The gradient optimization algorithm uses the Adam algorithm; Data preprocessing; Performance evaluation.

2. The digital heating furnace simulation method based on a neural network model as described in claim 1, characterized in that, The formula for the Adam algorithm is shown below: m t =μm t-1 +(1-β1)g t , m t This is a first-moment estimate of the gradient; n t Let μ be the second moment estimate of the gradient; μ,v∈[0,1), and is the exponential decay rate of the moment estimate; For m t Correction; For n t The correction; η is the learning rate, Δθ t To update parameters, g t Let ε be the gradient over the time step, and ε be the error rate.

3. The digital heating furnace simulation method based on a neural network model as described in claim 2, characterized in that, Data preprocessing: First, the data is preprocessed, including outlier detection and filtering; then, a linear transformation is used to linearly scale the dataset to the range [0,1]. In the formula x min and x max These are the minimum and maximum values ​​of the dataset, respectively. The neural network is configured with 2 hidden layers, 380 neurons in each hidden layer, a learning rate of 0.01, and 500 training iterations. The Adam algorithm is selected for gradient descent optimization.

4. The digital heating furnace simulation method based on a neural network model as described in claim 3, characterized in that, Performance evaluation is conducted by calculating three commonly used evaluation metrics to test prediction accuracy: RMSE, MAPE, and MAE; the expressions for MAPE, RMSE, and MAE are: in and y t These are the predicted value and the actual value, respectively; k is the number of predictions; if RMSE < 0.6, MAPE < 0.05, and MAE < 0.5, then the model training is successful.

Images