CN-121975989-A - Converter blowing-stop temperature prediction method based on deep learning neural network

CN121975989ACN 121975989 ACN121975989 ACN 121975989ACN-121975989-A

Abstract

The invention discloses a converter blowing-stop temperature prediction method based on a deep learning neural network, and belongs to the field of automatic steelmaking control. The method comprises a converter blowing process history furnace time data acquisition module, a converter blowing process history furnace time data preprocessing module, a converter blowing process raw material utilization rate calculation module, a deep learning neural network (DNN) model training module and a current furnace time converter endpoint prediction module. The accurate prediction and dynamic adjustment of the blowing-stop temperature in the converter smelting process are realized, and the accuracy of the blowing-stop temperature prediction of the converter is up to 91%. The technical breakthrough not only greatly improves the automation and intelligence level of converter smelting, but also obviously reduces errors caused by human factors, thereby improving the production efficiency and the product quality. Through accurate control blowing-stopping temperature, the model has effectively reduced the excessive use of coolant, has reduced the wasting of resources, has further reduced manufacturing cost.

Inventors

HE XIAODONG
LUO JUNBIN
YU JIEBIN
LIU RUIPENG
CUI XING
HE ZONGGUI
MA DEWEN

Assignees

宝钢湛江钢铁有限公司
上海宝信软件股份有限公司

Dates

Publication Date: 20260505
Application Date: 20260123

Claims (5)

1. The converter blowing-stop temperature prediction method based on the deep learning neural network is characterized by comprising the following steps of: S1, collecting historical furnace times of a converter converting process, namely automatically collecting the historical furnace times of the converter converting process from a converter data server by using a data collecting program; S2, preprocessing historical furnace times of the converter converting process, namely further screening and processing data acquired from the historical furnace times of the converter converting process, firstly removing the furnace times with abnormal data items by taking 95% as confidence intervals according to normal distribution conditions of all the data items, and then carrying out normalization processing on the screened data to enable the data to be in the same order of magnitude; s3, calculating the raw material utilization rate in the converting process of the converter, namely predicting or calculating the raw material utilization rate in the later stage of converting by utilizing a metallurgical mechanism and a linear regression model based on the historical converting data of the converter; S4, training a deep learning neural network model, namely dividing a structured data set obtained from a tapping process historical furnace time data screening and processing module into a training data set and a test data set, firstly constructing a deep learning neural network architecture to capture sequence dependence and time dynamics in data, and then training and learning the deep learning neural network model by using the training data set to learn patterns and trends in the data; S5, predicting the blowing-stop temperature of the converter in the current furnace, namely firstly obtaining the data of the current furnace, then carrying out structural processing on the obtained data of the current furnace, and finally predicting the blowing-stop molten steel temperature of the current furnace by using a converter blowing-stop temperature prediction model obtained by a neural network model training module.
2. The method for predicting the converter blowing stop temperature based on the deep learning neural network according to claim 1, wherein in the step S1, in the process of collecting historical converter times in the converter blowing process, each converter time comprises data items including a furnace seat number, a molten iron adding weight, adding weights of various scrap steels, relevant data before and after sublance measurement and target molten steel indexes.
3. The method for predicting the converter stop temperature based on the deep learning neural network according to claim 1, wherein in step S3, the calculation of the raw material utilization rate in the converter converting process specifically includes: ① . Adopting converter process historical data required by calculating the raw material utilization rate as a data set of the module, wherein the data set comprises related data before and after sublance measurement, molten iron and scrap steel adding weight, and dividing the data into a training set and a testing set; ② . The raw material utilization rate in the later stage of smelting is calculated according to a metallurgical mechanism, and the calculation formula of the raw material utilization rate is as follows: In the above-mentioned method, the step of, Is the utilization rate of raw materials; the consumption of raw materials and kg of the raw materials are consumed in the later stage of smelting; is the consumption of raw materials in the later period of smelting, kg; Is the ore addition amount in the later smelting stage, kg; And Respectively measuring component indexes of a sublance in converting and a sublance in blowing stop,%; The weight of molten steel is kg; 、 And The relative atomic masses of the index elements, respectively; And The method is characterized in that the temperature index measured by a blowing-stopping sublance and the temperature index measured by the sublance in blowing are respectively at the temperature DEG C; is the ore temperature, C/kg; And Is the thermal effect value of the molten iron component, KJ; Iron loss rate in converter smelting process,%; And The weight of molten iron and the weight of scrap steel are kg; ③ . The key parameters affecting the utilization rate of the raw materials are taken as independent variables, including the weight of molten steel after tapping and the related data before and after measuring by a sublance, and the utilization rate of the raw materials is taken as the independent variables to establish a linear regression model, wherein the model is formed as follows: In the above-mentioned method, the step of, , ,......, The regression coefficient, c is the error term; ④ . The performance of the model is estimated by using a test set, the accuracy of coefficient estimation is judged according to standard error, the significance of coefficient estimation judges that independent variables have significant influence on response variables, the fitting degree of the coefficient judgment model is determined, the condition number judges the relativity among the independent variables, the problem of multiple collinearity is avoided, and the parameters of the model are adjusted according to the estimation result; ⑤ . The model is deployed in practical application, the molten steel and the temperature measured by a dependent variable blowing-stopping sublance in a regression equation are changed into target molten steel and temperature, and the raw material utilization rate in the later stage of current heat smelting is calculated.
4. The converter blowing-stop temperature prediction method based on the deep learning neural network according to claim 1, wherein in the step S4, the deep learning neural network model uses the deep learning connection in the deep learning neural network model to memorize and update the history information, a test data set is periodically used for verifying the training effect of the deep learning neural network model along with the training, if the training effect reaches the expected standard, the training is stopped, and the finally obtained deep learning neural network model is saved as a converter end point control model.
5. The converter blowing-stop temperature prediction method based on the deep learning neural network according to claim 1, wherein in step S4, the specific process of training the deep learning neural network model is as follows: ① . Setting basic parameters of a neural network algorithm of a deep learning neural network model, setting the time step to be more than 19, setting the number of output layer nodes to be 4, setting the output layer node function to be a linear function purelin, and setting the number of the output layer nodes to be consistent with the number of output variables; ② . Calculating hidden states at this time using initialized neural network parameters Each time step t of the deep learning neural network model has a hidden state This hidden state is based on the current input And hidden state of previous time step Calculated, hidden state Is calculated as follows: in the above formula, W is the weight matrix from the previous hidden state to the current hidden state, U is the weight matrix currently input to the current hidden state, b is the bias term of the hidden layer, An activation function tanh which is an implicit layer; ③ . In the deep learning neural network model, based on the current hidden state Calculation output : In the above formula, V is a weight matrix implying a state to the output layer, Is a bias term for the output layer, Is the activation function purelin of the output layer; ④ . When a prediction model is trained, a loss function is constructed according to the difference between the result of model prediction in the ③ th step and a true value so as to quantify the accuracy of prediction, and the model takes the average square error as the loss function and comprises the following calculation method: in the above formula, y is the actual output, Is the calculation output of the deep learning neural network model; ⑤ . In order to optimize parameters of the deep learning neural network model, calculating gradients of a loss function L on each parameter, training the deep learning neural network model by BPTT algorithm, wherein the algorithm is an expansion of a back propagation algorithm on a time sequence, three parameters needing optimizing are observed according to the structure of the deep learning neural network model, namely U, V, W, historical data before tracing an optimizing process of the two parameters of U, W, and the parameter V only pays attention to the current; The partial derivative of parameter V is as follows: The general formula of the partial derivatives of L on U and W at time t is as follows: The integral partial guide formula is to accumulate the partial guide formulas according to time; the activation function is sleeved into the partial derivative formula, and the intermediate cumulative part is taken out, wherein the formula is as follows: ⑥ . Deep learning neural network model training adopts a gradient descent method to optimize weights The update rule formula of the weight w is as follows: in the above formula, w new is the updated weight matrix, and γ is the learning rate; to calculate the gradient, the error term is located as follows: Calculating error items of the output layer: the error terms for the recursive computation of the output node and hidden node are: Updating the weight matrix: Output layer: input layer: hidden layer: ⑦ . The above forward propagation, loss calculation and back propagation processes are repeated for a number of iterations across the training dataset until the model converges, i.e. the loss function reaches a small value and the performance of the model is no longer significantly improved.

Description

Converter blowing-stop temperature prediction method based on deep learning neural network Technical Field The invention relates to automatic control of converter steelmaking, in particular to a converter blowing-stop temperature prediction method based on a deep learning neural network. Background Currently, a converter blowing stop temperature prediction model mainly depends on two major technologies of a data driving model and a mechanism model. Data-driven models, particularly neural networks and Support Vector Machines (SVMs), have become central forces in this area. The models can accurately capture the dynamic changes of key parameters such as temperature, components and the like in the converter smelting process through the hidden modes in the deep mining historical data, so that the accurate prediction of the blowing-stop temperature is realized. The neural network can still keep higher prediction precision under complex and changeable furnace conditions by the strong nonlinear mapping capability and self-learning capability. And the SVM can be used for effectively classifying and regressing the data by searching the optimal hyperplane, so that another powerful tool is provided for predicting the blowing-stop temperature of the converter. The mechanism model is based on thermodynamic and kinetic principles, and establishes mathematical relations between converter blowing-stop temperature and various influencing factors by simulating chemical reactions and heat transfer processes in the converter. Although the calculation process of the mechanism model is relatively complex and needs more assumption conditions, the theory is strong, the physical and chemical mechanisms behind the prediction result can be revealed, and an important basis is provided for process optimization. The data driving model is known by the strong data processing capability and high adaptability, and can rapidly adapt to the prediction requirements under different furnace conditions. However, such models have a high degree of data dependence, and require sufficient and high quality historical data as support. In addition, as the data volume increases, the computational complexity of the model also increases, and the requirement on computational resources is also higher. In contrast, the mechanism model has more calculation complexity and more assumption conditions, but the prediction result has stronger interpretation. The mechanism model can clearly understand which factors are key factors affecting the blowing-stop temperature and how to adjust the factors to achieve the expected smelting effect. However, due to the complexity and variability of the conditions within the furnace, the accuracy of prediction of the mechanism model in practical applications may be somewhat affected. Disclosure of Invention The invention aims to provide a converter blowing-stop temperature prediction method based on a deep learning neural network, which aims to realize accurate prediction of blowing-stop temperature in the converter smelting process by combining data feedback of a sublance technology through combining data driving advantages of a neural network model and a physical and chemical theoretical basis of a mechanism model. The method can improve the prediction precision of the blowing-stop temperature, reduce production errors and fluctuation, optimize production efficiency, precisely control key process parameters such as oxygen blowing amount, feeding amount and the like according to the prediction temperature, be beneficial to realizing the efficient utilization of energy, reduce the energy consumption and material consumption in the production process, further obviously reduce the production cost, and ensure the stability and consistency of the product quality. The technical scheme is that the converter blowing stop temperature prediction method based on the deep learning neural network comprises a first converter blowing process history furnace time data acquisition module, a second converter blowing process history furnace time data preprocessing module, a third converter blowing process raw material utilization rate calculation module, a fourth deep learning neural network (DNN) model training module and a fifth converter end point prediction module. Historical furnace time data acquisition module for converter blowing process The historical converter times data acquisition module for the converter converting process automatically acquires historical converter times data of the converter converting process from a converter data server by using a data acquisition program, wherein each converter time data comprises data items including a base number, molten iron adding weight, adding weight of various waste steel, related data before and after measuring by a sublance and target molten steel indexes. Converter converting process history furnace time data preprocessing module The converter blowing process history furnace time data p