A converter endpoint prediction method and device based on multi-task learning

Abstract

The present disclosure provides a converter endpoint prediction method based on multi-task learning, including: obtaining parameter information of a converter; preprocessing the parameter information of the converter to obtain an input vector; inputting the input vector into a pre-trained prediction model to simultaneously obtain the predicted molten steel temperature and carbon content of the converter; obtaining a predicted converter endpoint range based on the predicted molten steel temperature and carbon content of the converter; starting the smelting of hot metal, and generating a termination reminder when it is determined that the molten steel temperature and the carbon content of the converter reach the predicted converter endpoint range. The present disclosure can consider the diversity of process conditions and the complexity of the steel smelting process, and can also avoid the problem that traditional data-driven models are prone to falling into local minima, realizing the accurate simultaneous prediction of the converter endpoint temperature and carbon content.

Description

Technical Field

[0001] The present disclosure relates to the technical field of steelmaking automation control, and particularly to a converter endpoint prediction method based on multi-task learning. Background Art

[0002] Converter steelmaking is a key link in the production and manufacturing of the iron and steel industry, and realizing intelligent manufacturing has a certain leading and exemplary role in the industry. Converter steelmaking is a complex high-temperature physical and chemical change process. It blows oxygen on the molten iron bath to achieve the purpose of decarburization, temperature increase, and change of molten steel composition. Its ultimate goal is to obtain qualified molten steel temperature and composition. With the development of refined production in steel mills, the control requirements for molten steel temperature and composition are becoming increasingly strict. In order to shorten the time from converter to tapping and reduce the negative impact of rapid direct tapping, developing a model with high prediction accuracy can meet the production requirements of efficient, compact, and modern steelmaking.

[0003] The molten steel temperature at the converter endpoint is a key index to ensure the quality of molten steel and continuous casting billets, and it is also a necessary factor to ensure the smooth progress of the refining and continuous casting production rhythm. Accurately controlling the carbon content at the converter endpoint can not only avoid overoxidation of molten steel, reduce the loss of alloy combustion after the furnace, but also reduce carbon emissions in the steelmaking process to a certain extent. Therefore, accurately controlling the molten steel endpoint temperature and endpoint carbon content is the focus of the converter smelting process.

[0004] High-precision prediction of the converter endpoint is an important basis for realizing intelligent converter smelting. The prediction methods of the converter endpoint are divided into mechanism methods and data-driven methods. The mechanism model is of great significance for predicting the converter endpoint temperature and composition because it can better analyze the relationship between influencing factors and the converter endpoint temperature and composition. However, due to the many theoretical assumptions and parameters involved in the mechanism method, and due to the nonlinearity and complexity of the converter process, it is difficult for the mechanism model to achieve high prediction accuracy. Therefore, in order to improve the prediction accuracy of converter endpoint control, many researchers have proposed using data-driven methods to predict the temperature and composition of the converter endpoint. When facing the diversity of process conditions and the complexity of the steel smelting process, the data-driven model lacks the analysis of the correlation between target values, and the traditional data-driven model is prone to falling into local minima. Summary of the Invention

[0005] The present disclosure provides a converter endpoint prediction method based on multi-task learning. Aiming at the problem that the data-driven model lacks the analysis of the correlation between target values, the present disclosure can not only consider the diversity of process conditions and the complexity of the steel smelting process, but also avoid the problem that the traditional data-driven model is prone to falling into local minima.

[0006] To achieve the above-mentioned invention purpose, the technical solution provided by the present disclosure is as follows:

[0007] On the one hand, a converter endpoint prediction method based on multi-task learning is provided, including:

[0008] S1: Obtain the parameter information of the converter, and the parameter information of the converter includes: furnace number, weight of hot metal charged, hot metal temperature, carbon content of hot metal, silicon content of hot metal, manganese content of hot metal, phosphorus content of hot metal, sulfur content of hot metal, scrap addition, weight of high-calcium lime, weight of dolomite, and oxygen supply;

[0009] S2: Preprocess the parameter information of the converter to obtain an input vector, and the preprocessing includes data screening, data cleaning, standardization, and constructing an input vector according to the furnace number;

[0010] S3: Input the input vector into a pre-trained prediction model, and simultaneously obtain the predicted converter molten steel temperature and converter carbon content. The prediction model is a multi-task learning model optimized based on the genetic algorithm. The multi-task learning model has a shared fully connected layer and two tasks, and the two tasks include separate hidden layers;

[0011] S4: Based on the predicted converter molten steel temperature and converter carbon content, obtain the predicted converter endpoint range;

[0012] S5: Start smelting the hot metal. When it is determined that the molten steel temperature of the converter and the carbon content of the converter reach the predicted converter endpoint range, a termination reminder is generated.

[0013] Preferably, before the S3, the method further includes:

[0014] S0: Train an initial multi-task learning neural network model with training data to obtain a pre-trained prediction model;

[0015] The training of the initial multi-task learning neural network model with training data in S0 to obtain a pre-trained prediction model includes:

[0016] S01: Collect training data;

[0017] S02: Construct an initial multi-task learning neural network model for predicting the converter molten steel temperature and converter carbon content. The initial model includes: an input layer, a shared network layer, task one, the hidden layer of task one, task two, and the hidden layer of task two;

[0018] S03: Define loss functions for task one and task two. The specific formula of the loss function is as follows:

[0019]

[0020] In the formula: α is the loss weight coefficient of task one, m is the number of samples, Y T is the actual converter molten steel temperature, Y′T is the predicted value of the molten steel temperature in the converter; β is the loss weight coefficient of Task 2, and Y C is the actual carbon content in the converter, and Y′ C : is the predicted value of the converter carbon content, where α + β = 1;

[0021] S04: Use the genetic algorithm to optimize the neural network model of the initial multi-task learning to obtain the optimized model parameters;

[0022] S05: According to the optimized model parameters, construct an optimized neural network model for multi-task learning. The optimized model parameters include the loss weight coefficient of Task 1, the loss weight coefficient of Task 2, the number of nodes in the shared network layer in the neural network structure, and the number of nodes in the hidden layer of each individual task;

[0023] S06: Input the training data into the optimized neural network model for multi-task learning, and train based on the loss model to obtain a pre-trained prediction model.

[0024] Preferably, the collection of training data in S01 includes:

[0025] Collect the actual production data of the converter smelting process from different converters in the steel plant or different heats of the same converter;

[0026] Extract the converter parameters from the actual production data. The converter parameters include furnace number / heat, weight of hot metal charged, hot metal temperature, hot metal carbon content, hot metal silicon content, hot metal manganese content, hot metal phosphorus content, hot metal sulfur content, scrap addition, weight of high-calcium lime, weight of dolomite, and oxygen supply;

[0027] Preprocess the converter parameters, and use one heat as a sample input vector;

[0028] Randomly select a part of the sample input vectors as the training set according to the selected ratio, and use the remaining sample input vectors as the test set.

[0029] Preferably, the use of the genetic algorithm to optimize the neural network model of the initial multi-task learning in S04 to obtain the optimized model parameters includes:

[0030] S041: Determine the number of populations and generate the initial population;

[0031] S042: Set the individual fitness function. The specific formula of the individual fitness function is as follows:

[0032]

[0033] In the formula, F(C,T) is the fitness function, n is the number of heats, and n(C,T) is the number of heats / times of the converter molten steel temperature and converter carbon content within the error range;

[0034] S043: Set the crossover and mutation operators;

[0035] S044: After performing the selection operation on the initial population, determine the individuals, and according to the crossover operation or mutation operation, generate the selected individuals into a population representing a new solution set;

[0036] S045: Set the termination condition, and repeat step S044 until the termination condition is met;

[0037] S046: When the termination condition is reached, decode the optimal individual in the last population to obtain the loss weight coefficient of task one, the loss weight coefficient of task two, the number of nodes in the shared network layer in the neural network structure, and the number of nodes in the hidden layers of each individual task.

[0038] Preferably, the obtaining of the predicted converter end range based on the predicted converter molten steel temperature and converter carbon content in S4 includes:

[0039] Add a ±15°C error range to the predicted converter molten steel temperature to obtain the predicted converter molten steel temperature range;

[0040] Add a ±0.02% error range to the predicted converter carbon content to obtain the predicted converter carbon content range;

[0041] Combine the predicted converter molten steel temperature range and the predicted converter carbon content range to obtain the predicted converter end range.

[0042] On the other hand, a converter end point prediction device based on multi-task learning is provided. The device includes:

[0043] Parameter unit: used to obtain the parameter information of the converter, and the parameter information of the converter includes: furnace number, weight of hot metal charged, hot metal temperature, hot metal carbon content, hot metal silicon content, hot metal manganese content, hot metal phosphorus content, hot metal sulfur content, scrap addition, weight of high-calcium lime, weight of dolomite, and oxygen supply;

[0044] Preprocessing unit: used to preprocess the parameter information of the converter to obtain an input vector, and the preprocessing includes data screening, data cleaning, standardization, and constructing an input vector according to the furnace number;

[0045] Prediction unit: used to input the input vector into a pre-trained prediction model, and simultaneously obtain the predicted converter molten steel temperature and converter carbon content. The prediction model is a multi-task learning model optimized based on the genetic algorithm. The multi-task learning model has a shared fully connected layer and two tasks, and the two tasks include separate hidden layers;

[0046] End point unit: used to obtain the predicted converter end range based on the predicted converter molten steel temperature and converter carbon content;

[0047] Reminder unit: used to start smelting molten iron, and generate a termination prompt when it is determined that the molten steel temperature and the carbon content of the converter reach the predicted converter end point range.

[0048] Preferably, the parameter unit includes:

[0049] Production data module: used to collect the actual production data of the converter smelting process from different converters in the steel plant or different heats of the same converter;

[0050] Parameter module: used to extract converter parameters from the actual production data, and the converter parameters include furnace number / heat number, weight of hot metal charged, hot metal temperature, hot metal carbon content, hot metal silicon content, hot metal manganese content, hot metal phosphorus content, hot metal sulfur content, scrap addition, weight of high-calcium lime, weight of dolomite, and oxygen supply;

[0051] Pretreatment module: used to preprocess the converter parameters, and one heat is used as a sample input vector;

[0052] Training data module: used to randomly select a part of the sample input vectors as the training set according to a selected ratio, and the remaining sample input vectors as the test set.

[0053] Preferably, the end point unit includes:

[0054] Molten steel temperature module: used to add an error range of ±15°C to the predicted molten steel temperature of the converter to obtain the predicted molten steel temperature range of the converter;

[0055] Carbon content module: used to add an error range of ±0.02% to the predicted carbon content of the converter to obtain the predicted carbon content range of the converter;

[0056] End point range module: used to combine the predicted molten steel temperature range and the predicted carbon content range of the converter to obtain the predicted converter end point range.

[0057] On the other hand, an electronic device is provided, and the electronic device includes: a housing, a processor, a memory, a circuit board, and a power supply circuit. Among them, the circuit board is arranged inside the space surrounded by the housing, and the processor and the memory are arranged on the circuit board; the power supply circuit is used to supply power to each circuit or device of the above-mentioned electronic device; the memory is used to store executable program codes; the processor runs the program corresponding to the executable program codes by reading the executable program codes stored in the memory, and is used to execute the above-mentioned evaluation assistance method.

[0058] On the other hand, a computer-readable storage medium is provided, and the computer-readable storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the above-mentioned evaluation assistance method.

[0059] The above technical solution has at least the following beneficial effects compared with the prior art:

[0060] The above solution, a converter endpoint prediction method based on multi-task learning, takes into account the potential mutual influence and connection between prediction targets by using the multi-task learning neural network algorithm. On this basis, the genetic algorithm is used to optimize the weight coefficients of tasks and the number of neuron nodes, realizing the accurate simultaneous prediction of the converter endpoint temperature and carbon content. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] In order to more clearly illustrate the technical solutions in the embodiments of the present disclosure, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present disclosure. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0062] Figure 1 It is a flowchart of a converter endpoint prediction method based on multi-task learning provided by the present disclosure;

[0063] Figure 2 It is a schematic diagram of a neural network model based on multi-task learning provided by the present disclosure;

[0064] Figure 3 It is a schematic diagram of a converter endpoint prediction device based on multi-task learning provided by the present disclosure. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0065] To make the objectives, technical solutions, and advantages of the embodiments of the present disclosure clearer, the following will clearly and completely describe the technical solutions of the embodiments of the present disclosure with reference to the drawings of the embodiments of the present disclosure. Obviously, the described embodiments are some, but not all, of the embodiments of the present disclosure. All other embodiments obtained by those of ordinary skill in the art based on the described embodiments of the present disclosure without creative efforts fall within the scope of protection of the present disclosure.

[0066] Unless otherwise defined, the technical terms or scientific terms used in this disclosure shall have the ordinary meanings as understood by those of ordinary skill in the art to which this disclosure pertains. The terms "first", "second" and similar words used in this disclosure do not denote any order, quantity or importance, but are only used to distinguish different components. Similarly, words such as "a", "an" or "the" do not denote a quantity limitation, but mean that there is at least one. Words such as "comprising" or "including" mean that the elements or objects appearing before this word cover the elements or objects listed after this word and their equivalents, without excluding other elements or objects. Words such as "connected" or "coupled" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect.

[0067] It should be noted that the terms "upper", "lower", "left", "right", "front", "rear", etc. used in this disclosure are only used to represent relative positional relationships. When the absolute position of the object being described changes, the relative positional relationship may also change accordingly.

[0068] Aiming at the problem that the existing data-driven models lack the analysis of the correlation between target values, this disclosure provides a method that can consider the potential mutual influences and connections between prediction targets by using a multi-task learning neural network algorithm, and can achieve the simultaneous accurate prediction of the converter endpoint temperature and carbon content.

[0069] As Figure 1 shown, the embodiments of this disclosure provide a converter endpoint prediction method based on multi-task learning. The method includes the following steps:

[0070] S1: Obtain the parameter information of the converter. The parameter information of the converter includes: furnace number, weight of hot metal charged, hot metal temperature, hot metal carbon content, hot metal silicon content, hot metal manganese content, hot metal phosphorus content, hot metal sulfur content, scrap addition, weight of high-calcium lime, weight of dolomite, and oxygen supply;

[0071] It should be further noted that the approximate ranges of the raw material addition amounts in converter steelmaking are obtained through mechanism model calculations or are artificially set values. Before smelting, parameters such as the amount of hot metal, hot metal composition, addition amounts and compositions of slag-making materials (such as dolomite, lime, etc.), and scrap addition amount have been measured and determined by the steel plant.

[0072] S2: Preprocess the parameter information of the converter to obtain an input vector. The preprocessing includes data screening, data cleaning, standardization, and constructing an input vector according to the furnace number;

[0073] S3: Input the input vector into a pre-trained prediction model to obtain the predicted temperature of the converter molten steel and the carbon content of the converter simultaneously. The prediction model is a multi-task learning model optimized by a genetic algorithm. The multi-task learning model has a shared fully connected layer and two tasks, and the two tasks include separate hidden layers;

[0074] It should be noted that a neural network is a data-driven intelligent algorithm that processes and predicts information by simulating the transmission process between neurons in biological nerves. Therefore, neural networks can be used for various prediction tasks. Among them, the back propagation neural network (BPNN) is one of the most commonly used neural networks. It trains the network through the back propagation algorithm, enabling the network to learn the mapping relationship between the input and output, and has a high accuracy in solving non-linear problems and certain advantages in solving complex problems. The converter steelmaking process is complex and non-linear, so using BPNN can effectively reduce the influence of adverse factors on the prediction results.

[0075] Multi-task learning is an inductive transfer method. Given the input data and output data of several related tasks, it can make full use of the relevant information between multiple tasks to improve the distinguishability or prediction ability between tasks and learn multiple models simultaneously. Compared with single-task learning, multi-task learning can avoid the problem of model underfitting caused by the neglect of some data features due to the limited number of training samples; in addition, since multiple related tasks are learned simultaneously, potential data features between tasks can be mined, improving the training effect of the model, and the obtained shared model has better generalization ability. Therefore, a multi-task neural network model for predicting the carbon temperature at the end of the converter is established based on the BPNN structure. Compared with the single-task learning neural network, the multi-task learning neural network can effectively identify and learn the common information between subsequent individual tasks due to the setting of the shared layer before each task.

[0076] Before the step S3, the method further includes the following steps:

[0077] S0: Train an initial multi-task learning neural network model with training data to obtain a pre-trained prediction model;

[0078] The step S0 of training the initial multi-task learning neural network model with training data to obtain a pre-trained prediction model includes:

[0079] S01: Collect training data;

[0080] It should be noted that the training data is the actual production data of the converter smelting process collected from different converters in the steel plant or different heats of the same converter; from the actual production data, converter parameters are extracted, and the converter parameters include furnace number / heat number, weight of hot metal charged, hot metal temperature, carbon content of hot metal, silicon content of hot metal, manganese content of hot metal, phosphorus content of hot metal, sulfur content of hot metal, scrap addition, weight of high-calcium lime, weight of dolomite, oxygen supply; the converter parameters are preprocessed, and one heat is used as one input vector; a part of the input vectors are randomly selected as the training set according to the selected ratio, and the remaining input vectors are used as the test set.

[0081] S02: Construct an initial multi-task learning neural network model for predicting the temperature of converter molten steel and the carbon content of the converter, as Figure 2 shown, the initial model includes: an input layer, a shared network layer, Task 1, the hidden layer of Task 1, Task 2, and the hidden layer of Task 2;

[0082] In some embodiments, the initial multi-task learning neural network model selects a shared network layer, each individual task has a hidden layer, and the weight of the loss function of each task is 0.5; then the number of nodes in the input layer is 11, the number of nodes in the shared network layer is 9, the number of nodes in the hidden layer of the individual task is 3, the number of nodes in the output layer is 1, the ReLU activation function is used between the shared layers, the sigmoid activation function is used in the output layer of the network, and the Adam optimizer is used to update the parameters.

[0083] Preferably, the learning rate is set to 0.005, and the number of iterations is 500 times.

[0084] S03: Define loss functions for Task 1 and Task 2, and the specific formulas of the loss functions are as follows:

[0085]

[0086] In the formula: α is the loss weight coefficient of Task 1, m is the number of samples, Y T is the actual temperature of converter molten steel, Y' T is the predicted value of the temperature of converter molten steel; β is the loss weight coefficient of Task 2, Y C is the actual carbon content of the converter, Y' C is the predicted value of the carbon content of the converter, α + β = 1;

[0087] S04: Optimize the initial multi-task learning neural network model using the genetic algorithm to obtain optimized model parameters;

[0088] In some embodiments, the end temperature and carbon content are taken as the prediction objects and different tasks (Task 1, Task 2), and different loss function weights are assigned to the temperature prediction task and the carbon content prediction task. Different weights determine the influence level of the carbon temperature task on the final training effect. The genetic algorithm is used to optimize the loss function weights to make the multi-task learning have better training effects.

[0089] Specifically, the S04 includes:

[0090] S041: Determine the population size and generate an initial population;

[0091] In some embodiments, the determination of the initial population is the determination of the population size. Too large a population size will reduce the algorithm's running efficiency and increase the operation time. Too small a population size will lead to the failure to reach the global optimal solution during the training process, thus affecting the training effect of the algorithm.

[0092] Preferably, the population size can be selected as 50.

[0093] S042: Set the individual fitness function, and the specific formula of the individual fitness function is as follows:

[0094]

[0095] In the formula, F(C,T) is the fitness function, n is the number of furnace operations, and n(C,T) is the number of furnaces / operations in which the converter molten steel temperature and the converter carbon content are within the error range;

[0096] It should be noted that the definition method of the fitness function should be designed differently for different problems. The fitness function is used to judge the quality of individuals in the population. The higher the fitness function value, the higher the probability of being selected, and at the same time, the better the quality of its individual as the solution to the problem.

[0097] Preferably, the fitness function of the present disclosure is the reciprocal of the double hit rate of carbon and temperature.

[0098] S043: Set the crossover and mutation operators;

[0099] It should be noted that the size of the crossover probability will affect the speed of finding the optimal individual; the mutation operator can avoid the algorithm falling into a local optimal solution. Too large a value will cause the algorithm to become a random search algorithm, and too small a value is not conducive to the generation of new individuals.

[0100] Preferably, the crossover probability is generally taken as 0.4 - 0.9; usually, the value of the mutation operator is between 0.001 and 0.1.

[0101] Preferably, to obtain better results, the crossover probability is set to 0.8 and the mutation probability is set to 0.01.

[0102] S044: After performing a selection operation on the initial population, individuals are determined, and through crossover operation or mutation operation, the selected individuals are generated into a population representing a new solution set;

[0103] S045: Set termination conditions and repeat step S0234 until the termination conditions are met;

[0104] It should be noted that in the genetic algorithm, a maximum number of iterations needs to be set in advance. When the evolution reaches the maximum number of iterations, the optimal individual in the final population is decoded, and its solution can be used as the optimal solution to the problem to be solved.

[0105] Preferably, the maximum number of iterations is set to 100.

[0106] S046: After reaching the termination conditions, decode the optimal individual in the final population to obtain the loss weight coefficient of Task 1, the loss weight coefficient of Task 2, the number of nodes in the shared network layer in the neural network structure, and the number of nodes in the hidden layers of each individual task.

[0107] It should be noted that the genetic algorithm is a stochastic global search optimization method that mimics the genetic variation process in biology. It searches for optimization based on the principles of natural selection and genetics, simulating phenomena such as replication, crossover, and mutation that occur in natural selection and genetics. The basis of the genetic algorithm is the principle of biological evolution, "survival of the fittest, elimination of the unfit". Starting from the initial population, it forms an encoded tandem population, and then through operations such as random selection, crossover, and mutation, it is screened according to the designed fitness function, retaining individuals with high fitness, forming a new population that continuously reproduces and evolves until the limiting conditions are met, thereby solving complex problems that are difficult to solve by classical mathematics. The genetic algorithm has a simple principle and can be processed in parallel.

[0108] It should be further noted that in this disclosure, the genetic algorithm mainly optimizes the loss weight coefficients of each task and the number of neuron nodes in each layer of the neural network of the multi-task learning neural network model, uses the fitness function to reflect the superiority and inferiority of individuals, and after being processed by genetic operators and reaching the final number of iterations, selects the individual with the best fitness. Its result can be approximately used as the approximate solution to the current problem, that is, the loss weight coefficients of each task and the number of neuron nodes in each layer of the neural network, and then the result is applied to the multi-task learning neural network model for training to improve the hit rate of the end point temperature and the end point carbon content.

[0109] S05: According to the optimized model parameters, construct an optimized multi-task learning neural network model, where the optimized model parameters include the loss weight coefficient of Task 1, the loss weight coefficient of Task 2, the number of nodes in the shared network layer in the neural network structure, and the number of nodes in the hidden layers of each individual task;

[0110] In some embodiments, a genetic algorithm is used to optimize the task loss weight coefficients and the number of neuron nodes in each layer of the network. The solution results show that for the multi-task learning neural network model, the loss weight coefficient of Task 1 is 0.710, the loss weight coefficient of Task 2 is 0.290, the number of nodes in the input layer of the neural network structure is 11, the number of nodes in the shared network layer is 13, the number of nodes in the individual hidden layer for each task is 5, and the number of nodes in the output layer is 1. Based on this, the multi-task learning neural network model is reconstructed.

[0111] S06: Input the training data into the optimized multi-task learning neural network model and train it based on the loss model to obtain a pre-trained prediction model.

[0112] Furthermore, as shown in Table 1, the actual production data of a 300t converter in a certain steel plant is used to train the initial multi-task learning neural network model. The errors between the predicted converter molten steel temperature and converter carbon content and the actual values are concentrated within ±15°C and ±0.02%, and the double hit rate reaches 73.75%; however, when the error range is selected as temperature ±5°C and carbon content ±0.01%, the prediction accuracy is not high, and the carbon-temperature double hit rate is only 28.75%.

[0113]

[0114] Table 1: Error distribution of the prediction results of the initial multi-task learning neural network model.

[0115] As shown in Table 2, when training the optimized multi-task learning neural network model, the errors between the predicted converter molten steel temperature and converter carbon content and the actual values are concentrated within ±15°C and ±0.02%, and the hit rate is increased by 13.75%; when the error range is selected as temperature ±5°C and carbon content ±0.01%, the prediction accuracy is not high, and the hit rate is increased by 12.5%.

[0116]

[0117]

[0118] Table 2: Error distribution of the prediction results of the optimized multi-task learning neural network model.

[0119] S4: Obtain the predicted converter end point range based on the predicted converter molten steel temperature and converter carbon content;

[0120] It should be noted that: S4 includes:

[0121] Add an error range of ±15°C to the predicted converter molten steel temperature to obtain the predicted converter molten steel temperature range;

[0122] Add an error range of ±0.02% to the predicted converter carbon content to obtain the predicted converter carbon content range;

[0123] Combine the predicted converter molten steel temperature range and the predicted converter carbon content range to obtain the predicted converter end point range.

[0124] In some embodiments, the predicted converter molten steel temperature range and the predicted converter carbon content range are related to the double hit rate. In the present disclosure, the double hit rates of the converter molten steel temperature and the converter carbon content with errors within ±5°C and ±0.01%, ±10°C and ±0.015%, and ±15°C and ±0.02% of the multi-task learning model optimized by the genetic algorithm are 41.25%, 65%, and 87.5% respectively. Because ±15°C and ±0.02% have a relatively high double hit rate, they are selected as the final end point range.

[0125] S5: Start smelting the hot metal. When it is determined that the molten steel temperature and the converter carbon content of the converter reach the predicted converter end point range, generate a termination reminder.

[0126] The above is the introduction of the method embodiments. The following further illustrates the solution of the present disclosure through device embodiments.

[0127] As Figure 3 shown, the present disclosure provides a converter end point prediction device based on multi-task learning. The device includes:

[0128] A parameter unit 310: configured to obtain parameter information of the converter before smelting, where the parameter information of the converter includes: furnace number, weight of hot metal charged, hot metal temperature, hot metal carbon content, hot metal silicon content, hot metal manganese content, hot metal phosphorus content, hot metal sulfur content, scrap addition, weight of high-calcium lime, weight of dolomite, oxygen supply;

[0129] A preprocessing unit 320: configured to preprocess the parameter information of the converter, and the preprocessing includes data screening, data cleaning, standardization, and constructing an input vector according to the furnace number;

[0130] A prediction unit 330: configured to input the input vector into a pre-trained prediction model, and simultaneously obtain the predicted converter molten steel temperature and converter carbon content. The prediction model is a multi-task learning model optimized by a genetic algorithm. The multi-task learning model has a shared fully connected layer and two tasks, and the two tasks include separate hidden layers;

[0131] An end point unit 340: configured to obtain a predicted converter end point range based on the predicted converter molten steel temperature and converter carbon content;

[0132] Reminder unit 350: Used to generate a termination reminder when it is determined that the temperature of the converter molten steel and the carbon content of the converter reach the predicted converter end point range.

[0133] Preferably, the parameter unit includes:

[0134] Production data module: Used to collect the actual production data of the converter smelting process from different converters in the steel plant or different heats of the same converter;

[0135] Parameter module, used to extract converter parameters, the converter parameters include: furnace number / heat number, weight of hot metal charged, hot metal temperature, hot metal carbon content, hot metal silicon content, hot metal manganese content, hot metal phosphorus content, hot metal sulfur content, scrap addition, weight of high-calcium lime, weight of dolomite, oxygen supply;

[0136] Pretreatment module: Used to preprocess the converter parameters, with one heat as one input vector;

[0137] Training data module: Used to randomly select a part of the input vectors as the training set according to the selected ratio, and the remaining input vectors as the test set.

[0138] Preferably, the end point unit includes:

[0139] Molten steel temperature module: Used to add a ±15°C error range to the predicted converter molten steel temperature to obtain the predicted converter molten steel temperature range;

[0140] Carbon content module: Used to add a ±0.02% error range to the predicted converter carbon content to obtain the predicted converter carbon content range;

[0141] End point range module: Used to combine the predicted converter molten steel temperature range and the predicted converter carbon content range to obtain the predicted converter end point range.

[0142] In an exemplary embodiment, an electronic device is also provided. The electronic device includes: a housing, a processor, a memory, a circuit board, and a power supply circuit. Among them, the circuit board is arranged inside the space surrounded by the housing, and the processor and the memory are arranged on the circuit board; the power supply circuit is used to supply power to each circuit or device of the above-mentioned electronic device; the memory is used to store executable program codes; the processor runs the program corresponding to the executable program codes by reading the executable program codes stored in the memory.

[0143] In an exemplary embodiment, a computer-readable storage medium is also provided. The computer-readable storage medium stores one or more programs, and the one or more programs can be executed by one or more processors.

[0144] The following points need to be explained:

[0145] (1) The accompanying drawings of the embodiments of the present disclosure only relate to the structures involved in the embodiments of the present disclosure, and other structures may refer to the general design.

[0146] (2) For clarity, in the drawings used to describe the embodiments of the present disclosure, the thickness of layers or regions is enlarged or reduced, that is, these drawings are not drawn to actual scale. It can be understood that when an element such as a layer, film, region, or substrate is referred to as being "on" or "under" another element, the element can be "directly" on or under the other element or there can be intermediate elements.

[0147] (3) Without conflict, the embodiments of the present disclosure and the features in the embodiments can be combined with each other to obtain new embodiments.

[0148] The above is only the specific implementation manner of the present disclosure, but the protection scope of the present disclosure is not limited thereto. The protection scope of the present disclosure shall be subject to the protection scope of the claims.

Claims

1. A converter endpoint prediction method based on multi-task learning, characterized in that, Including: S1: Obtain the parameter information of the converter. The parameter information of the converter includes: furnace number, weight of hot metal charged, hot metal temperature, carbon content of hot metal, silicon content of hot metal, manganese content of hot metal, phosphorus content of hot metal, sulfur content of hot metal, scrap addition, weight of high-calcium lime, weight of dolomite, and oxygen supply; S2: Preprocess the parameter information of the converter to obtain an input vector. The preprocessing includes data screening, data cleaning, standardization, and constructing an input vector according to the furnace number; S0: Train the initial multi-task learning neural network model with training data to obtain a pre-trained prediction model; The step of S0 training the initial multi-task learning neural network model with training data to obtain a pre-trained prediction model includes: S01: Collect training data; S02: Construct an initial multi-task learning neural network model for predicting the molten steel temperature and carbon content of the converter. The initial model includes: an input layer, a shared network layer, a hidden layer for task one, and a hidden layer for task two; S03: Define loss functions for task one and task two. The specific formulas of the loss functions are as follows: Where: α is the loss weight coefficient of Task 1, m is the number of samples, Y T is the actual temperature of molten steel in the converter, Y T ′ is the predicted temperature of molten steel in the converter; β is the loss weight coefficient of Task 2, Y C is the actual carbon content in the converter, Y C ′ : the predicted carbon content in the converter, α + β = 1; S04: Optimize the initial multi-task learning neural network model using the genetic algorithm to obtain optimized model parameters; S05: According to the optimized model parameters, construct an optimized multi-task learning neural network model. The optimized model parameters include the loss weight coefficient of task one, the loss weight coefficient of task two, the number of nodes in the shared network layer in the neural network structure, and the number of nodes in the hidden layers of each individual task; S06: Input the training data into the optimized multi-task learning neural network model and train it based on the loss model to obtain a pre-trained prediction model; S3: Input the input vector into the pre-trained prediction model to simultaneously obtain the predicted molten steel temperature and carbon content of the converter. The prediction model is a multi-task learning model optimized based on the genetic algorithm. The multi-task learning model has a shared fully connected layer and two tasks. The two tasks contain separate hidden layers; S4: Based on the predicted molten steel temperature and carbon content of the converter, obtain the predicted end-point range of the converter; S5: Start smelting the hot metal. When it is determined that the molten steel temperature and the carbon content of the converter reach the predicted end-point range of the converter, generate a termination reminder.

2. The converter endpoint prediction method based on multi-task learning according to claim 1, characterized in that, The step of S01 collecting training data includes: Collect the actual production data of the converter smelting process from different converters in the steel plant or different heats of the same converter; Extract the converter parameters from the actual production data. The converter parameters include furnace number / heat number, weight of hot metal charged, hot metal temperature, carbon content of hot metal, silicon content of hot metal, manganese content of hot metal, phosphorus content of hot metal, sulfur content of hot metal, scrap addition, weight of high-calcium lime, weight of dolomite, and oxygen supply; Preprocess the converter parameters, and use one heat as a sample input vector; Randomly select a part of the sample input vectors as the training set according to the selected ratio, and use the remaining sample input vectors as the test set.

3. The converter endpoint prediction method based on multi-task learning according to claim 1, characterized in that, The step of S04 optimizing the initial multi-task learning neural network model using the genetic algorithm to obtain optimized model parameters includes: S041: Determine the population size and generate an initial population; S042: Set the individual fitness function, and the specific formula of the individual fitness function is as follows: In the formula, F(C,T) is the fitness function, n is the number of furnace operations, and n(C,T) is the number of furnaces / operations where the molten steel temperature and carbon content of the converter are within the error range; S043: Set the crossover and mutation operators; S044: Determine individuals after performing a selection operation on the initial population, and generate a population representing a new solution set by performing crossover operations or mutation operations on the selected individuals; S045: Set the termination condition, and repeat step S044 until the termination condition is met; S046: After reaching the termination condition, decode the optimal individual in the last population to obtain the loss weight coefficient of Task 1, the loss weight coefficient of Task 2, the number of nodes in the shared network layer in the neural network structure, and the number of nodes in the hidden layers of each individual task.

4. The converter endpoint prediction method based on multi-task learning according to claim 1, characterized in that, The S4 for obtaining the predicted converter end point range based on the predicted molten steel temperature and carbon content of the converter includes: Adding an error range of ±15°C to the predicted molten steel temperature of the converter to obtain the predicted molten steel temperature range of the converter; Adding an error range of ±0.02% to the predicted carbon content of the converter to obtain the predicted carbon content range of the converter; Combining the predicted molten steel temperature range and the predicted carbon content range of the converter to obtain the predicted converter end point range.

5. A converter endpoint prediction device based on multi-task learning, characterized in that, The device is applicable to the method of any one of claims 1-4, and the device includes: Parameter unit: used to obtain the parameter information of the converter, and the parameter information of the converter includes: furnace number, weight of hot metal charged, hot metal temperature, hot metal carbon content, hot metal silicon content, hot metal manganese content, hot metal phosphorus content, hot metal sulfur content, scrap addition, weight of high-calcium lime, weight of dolomite, and oxygen supply; Preprocessing unit: used to preprocess the parameter information of the converter to obtain an input vector, and the preprocessing includes data screening, data cleaning, standardization, and constructing an input vector according to the furnace number; Prediction unit: used to train the initial multi-task learning neural network model with training data to obtain a pre-trained prediction model, and is also used to input the input vector into the pre-trained prediction model to simultaneously obtain the predicted molten steel temperature and carbon content of the converter. The prediction model is a multi-task learning model optimized by a genetic algorithm. The multi-task learning model has a shared fully connected layer and two tasks, and the two tasks include separate hidden layers; and, The training of the initial multi-task learning neural network model with training data to obtain a pre-trained prediction model includes: S01: Collect training data; S02: Construct an initial multi-task learning neural network model for predicting the molten steel temperature and carbon content of the converter. The initial model includes: an input layer, a shared network layer, a hidden layer for Task 1, and a hidden layer for Task 2; S03: Define loss functions for Task 1 and Task 2, and the specific formulas of the loss functions are as follows: Where: α is the loss weight coefficient of Task 1, m is the number of samples, Y T is the actual temperature of molten steel in the converter, Y T ′ is the predicted temperature of molten steel in the converter; β is the loss weight coefficient of Task 2, Y C is the actual carbon content in the converter, Y C ′ : the predicted carbon content in the converter, α + β = 1; S04: Optimize the initial multi-task learning neural network model using a genetic algorithm to obtain optimized model parameters; S05: Construct an optimized neural network model for multi-task learning according to the optimized model parameters, where the optimized model parameters include the loss weight coefficient of Task 1, the loss weight coefficient of Task 2, the number of nodes in the shared network layer in the neural network structure, and the number of nodes in the hidden layer of each individual task. S06: Input the training data into the optimized neural network model for multi-task learning, and train it based on the loss model to obtain a pre-trained prediction model. Endpoint unit: Used to obtain the predicted converter endpoint range based on the predicted converter molten steel temperature and converter carbon content. Reminder unit: Used to start smelting of hot metal, and generate a termination reminder when it is determined that the molten steel temperature and the converter carbon content of the converter reach the predicted converter endpoint range.

6. The converter endpoint prediction device based on multi-task learning according to claim 5, characterized in that, The parameter unit includes: Production data module: Used to collect the actual production data of the converter smelting process from different converters in the steel plant or different heats of the same converter. Parameter module: Used to extract converter parameters from the actual production data, where the converter parameters include furnace number / heat number, weight of hot metal charged, hot metal temperature, hot metal carbon content, hot metal silicon content, hot metal manganese content, hot metal phosphorus content, hot metal sulfur content, scrap addition, weight of high-calcium lime, weight of dolomite, and oxygen supply. Preprocessing module: Used to preprocess the converter parameters, and input one heat as a sample input vector. Training data module: Used to randomly select a part of the sample input vectors as the training set according to the selected ratio, and the remaining sample input vectors as the test set.

7. The converter endpoint prediction device based on multi-task learning according to claim 5, characterized in that, The endpoint unit includes: Molten steel temperature module: Used to add an error range of ±15°C to the predicted converter molten steel temperature to obtain the predicted converter molten steel temperature range. Carbon content module: Used to add an error range of ±0.02% to the predicted converter carbon content to obtain the predicted converter carbon content range. Endpoint range module: Used to combine the predicted converter molten steel temperature range and the predicted converter carbon content range to obtain the predicted converter endpoint range.

8. An electronic device, characterized in that, The electronic device includes: a housing, a processor, a memory, a circuit board, and a power supply circuit. Among them, the circuit board is arranged inside the space enclosed by the housing, and the processor and the memory are arranged on the circuit board; the power supply circuit is used to supply power to each circuit or device of the above-mentioned electronic device; the memory is used to store executable program codes; the processor runs the program corresponding to the executable program codes by reading the executable program codes stored in the memory, and is used to execute the converter endpoint prediction method according to any one of claims 1 to 4.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the converter endpoint prediction method according to any one of claims 1 to 4.

Images