Method for optimizing the desulphurization of pig iron

Abstract

The invention relates to a method for optimizing pig iron desulphurization (1), wherein the pig iron desulphurization (1) is achieved by dispensing one or more desulphurizing agents (2, 3, 4) into a pig ladle (5) or torpedo car, wherein depending on the type of the dispensed desulphurizing agent (2, 3, 4) several treatment modes are available. The optimization is based on creating a separate model (8) for each of the several treatment modes to optimize the amount of desulphurizing agent (2, 3, 4) added for the respective treatment mode and training (9) the created separate models (8) for each of the several treatment modes after a predetermined number of pig iron desulphurizations (1) using the respective treatment mode. The future pig iron desulphurization (1) is optimized based on the predictions of the separate models (8), wherein the optimization determines the treatment mode and the amount of desulphurizing agent (2, 3, 4) in the treatment mode by comparing the predictions of the separate models (8). The invention also relates to a system and a computer program configured to perform the method of the invention.

Description

[0001] This invention relates to a method for optimizing the desulfurization of pig iron, wherein the desulfurization of pig iron is achieved by distributing one or more desulfurizing agents to a pig iron pot or torpedo car, wherein several processing modes are available depending on the type of desulfurizing agent distributed.

[0002] Desulfurization of pig iron is typically set up between the blast furnace and the converter. Desulfurizing agents such as magnesium (Mg), calcium carbide (CaC2), and lime (CaO) are usually present in different hoppers. For example, the desulfurizing agents are characterized by varying degrees of effectiveness in desulfurization and different material costs. The order of magnesium, calcium carbide, and lime as desulfurizing agents is a descending order of desulfurization effectiveness and material cost.

[0003] During the process, the desulfurizing agent is injected as powder into the pig iron pot or torpedo car via a spray gun. Depending on the processing conditions, one or more blow molding stages may occur, in which only calcium carbide or lime is added (single injection), or a mixture of calcium carbide or lime and magnesium is added (co-injection). The corresponding processing modes, namely lime single injection, calcium carbide single injection, lime-magnesium co-injection, and calcium carbide-magnesium co-injection, take into account the required desulfurization work, such as based on pig iron weight, initial sulfur content, and target sulfur content, as well as existing constraints regarding material availability and cost, maximum possible blowing rate, and predetermined distribution time.

[0004] In practice, analytical and numerical process models are known to help operators select the appropriate treatment mode and the amount of desulfurization agent to be added. These models typically incorporate empirical methods, describing the amount of desulfurization agent to be added as a function of several process variables, such as pig iron quantity, pig iron temperature, initial sulfur content, and target sulfur content. The accuracy of such empirical methods is naturally limited, especially in this case, where equipment-specific effects, such as yield coefficients, must also be considered, as these effects may change during operation. Ideally, during initial commissioning, the process model is adjusted for equipment specificity based on historical treatment data. However, further model adjustments during operation are often difficult. This may be related to one or more of the following reasons: the opacity of the model used or the influence of individual model parameters, the highly nonlinear relationship between the model's input and output variables, and the need for structured evaluation of large amounts of data. Therefore, individual treatments may result in non-target additions of desulfurizers that are inexpensive and / or require post-treatment.

[0005] Therefore, one object of the present invention is to optimize the amount of each desulfurizing agent added during the desulfurization process of pig iron so as to achieve the target sulfur content as reliably and economically as possible in each treatment.

[0006] This invention achieves this objective by means of the method according to claim 1. Advantageous designs of the method are defined in the dependent claims. The system according to claim 21 and the computer program according to claim 22 also achieve this objective.

[0007] This invention relates to a method for optimizing pig iron desulfurization, wherein pig iron desulfurization is achieved by distributing one or more desulfurizing agents into a pig iron pot or torpedo car, wherein several processing modes are available depending on the type of desulfurizing agent distributed, including the following steps: Determine the weight of the pig iron to be desulfurized, the initial sulfur content of the pig iron to be desulfurized, and the target sulfur content of the pig iron to be desulfurized; The type and amount of desulfurizing agent dispensed, as well as the actual sulfur content achieved after pig iron desulfurization, are detected and used for pig iron desulfurization. A separate model is created for each of the multiple processing modes to optimize the amount of desulfurizer added for the corresponding processing mode. The separate models are based on machine learning to predict the achievable sulfur content and the amount of desulfurizer to be allocated for future pig iron desulfurization. After desulfurizing a predetermined quantity of pig iron using the corresponding processing modes, a separate model is trained for each of the multiple processing modes. The training of each model is based on determining the weight of the pig iron to be desulfurized, its initial sulfur content, and its target sulfur content, and detecting the type and amount of desulfurizing agent allocated, as well as the actual sulfur content achieved after pig iron desulfurization. Optimize future pig iron desulfurization based on predictions from individual models, where optimization is achieved by comparing predictions from individual models to determine the treatment mode and the amount of desulfurizing agent added in the treatment mode.

[0008] In the first step, for each desulfurization of pig iron, the weight of the pig iron to be desulfurized, the initial sulfur content of the pig iron to be desulfurized, and the target sulfur content of the pig iron to be desulfurized are determined.

[0009] According to the method of the present invention, certain data are detected during the desulfurization of pig iron, particularly the type and amount of desulfurizing agent dispensed, and the actual sulfur content achieved after the desulfurization. Furthermore, the weight of the pig iron is measured during the desulfurization process. Therefore, historical data on the desulfurization of pig iron is collected, wherein the collected historical data relates to parameters relevant to the desulfurization. For example, the collected historical data is stored in a database.

[0010] For each of the several treatment modes, a separate model is built to optimize the amount of desulfurizer added for that mode. The separate models are based on different data because each model only considers data relevant to the specific treatment mode, ignoring data relevant to other treatment modes. Furthermore, it is best not to exchange data between the separate models. These separate models are based on machine learning and can predict the achievable sulfur content for future pig iron desulfurization and the required amount of desulfurizer to be allocated.

[0011] After a predetermined number of pig iron desulfurization processes are performed using the corresponding processing modes, a separate model created according to the present invention for each of the various processing modes is trained. The training of the separate models is based on the weight of the pig iron to be desulfurized, the initial sulfur content of the pig iron to be desulfurized, the target sulfur content of the pig iron to be desulfurized, and relevant historical data detected, particularly the weight of pig iron during the desulfurization process, the type and amount of desulfurizing agent allocated, and the actual sulfur content achieved after the desulfurization process. The training may also consider the weight of pig iron optionally detected during the desulfurization process. When training the separate models, only the performed pig iron desulfurization data and the corresponding processing modes are considered; therefore, different processing modes are strictly considered separately, and the models are trained separately. This eliminates the mutual influence between processing modes or models.

[0012] Using the established model, future pig iron desulfurization can be predicted and optimized. The optimization results determine a treatment mode and the corresponding amount of desulfurizing agent to be added. According to the invention, optimization is achieved by comparing the predictions of different models for different treatment modes.

[0013] According to an advantageous variation of the invention, the desulfurizing agent is selected from magnesium, calcium carbide, and / or lime. Theoretically, any substance with a higher affinity for sulfur than iron itself can be used for desulfurization (e.g., magnesium, cerium, calcium, sodium, etc.). For economic and technical reasons, magnesium, calcium carbide, and / or lime are preferred in practice as the aforementioned desulfurizing agents.

[0014] According to a variation of the invention, the method considers the following processing modes: lime single injection, calcium carbide single injection, lime-magnesium co-injection, calcium carbide-magnesium co-injection, and calcium carbide-magnesium co-injection with upstream and downstream lime single injections.

[0015] In a preferred embodiment of the invention, the individual model is constructed as a regression model. The regression model, based on statistical analysis methods, describes the relationship between a dependent variable (predictor variable) and one or more independent variables (control variables or regression variables). The regression model is used to quantitatively describe the relationship and predict the value of the dependent variable. Specifically, the invention relates to linear regression models. Regression models can be implemented, in particular, through machine learning.

[0016] According to a variation of the invention, the method takes into account the maximum possible addition rate and / or possible distribution time of several desulfurizing agents during the optimization process. This ensures that the selected treatment mode is technically feasible in practice.

[0017] According to another variation of the invention, the individual model considers the availability, quantity, and / or cost of various desulfurizing agents during the optimization process. Therefore, if the amount of desulfurizing agent required for a treatment mode is unavailable, the corresponding model will not produce results considered during optimization. By taking into account the cost of the desulfurizing agents, the material costs associated with various treatment modes, which differ in the desulfurizing agents used, can be determined.

[0018] In an advantageous variant of the invention, the method determines the amount of desulfurizing agent required to achieve the target sulfur content for each treatment mode by predicting a separate model of a certain amount of pig iron to be desulfurized with a specific initial sulfur content. Specifically, a treatment mode that can be implemented based on the amount of desulfurizing agent available and at the lowest cost will be selected or proposed below.

[0019] According to an advantageous variation of the invention, when training a separate model for each of the multiple processing modes, only data on pig iron desulfurization that achieves the target sulfur content with a predetermined accuracy are considered. Therefore, by specifying the predetermined accuracy, it can be determined which pig iron desulfurization processes need to be considered during model training. Only pig iron desulfurization with sufficient accuracy at the target sulfur content is considered. This ensures that the model predicts pig iron desulfurization with considerable accuracy at the target sulfur content. This is particularly beneficial for training those pig iron desulfurization processes that cannot be identified as anomalous according to descriptive statistical criteria.

[0020] According to a preferred variant of the invention, when training a separate model for each of the multiple processing modes, only data from a limited number of recently performed processes are considered. This allows for greater consideration of recent pig iron sulfidation processes during model training. This enables the model training to account for slowly changing conditions during pig iron desulfurization, particularly changes in pig iron desulfurization equipment. According to this variant, data regarding older pig iron sulfidation processes that no longer have the current equipment conditions are ignored, and only the most recent data on pig iron sulfidation is considered during model training.

[0021] In a variation of the invention, after training a separate model for each of the multiple processing modes, the trained model is compared with a previous model for each processing mode. Once the model is successfully trained, it is compared with the corresponding previous model. This ensures that the trained model is at least as good as the previous model.

[0022] According to an advantageous variation of the invention, the comparison is based on data from a recently processed number of processes and / or on a generally generated input dataset. Specifically, the comparison between the trained model and a previous model is performed on a number of randomly generated, generally generated input datasets within a predefined interval (Monte Carlo simulation).

[0023] According to a particularly advantageous variation of the invention, the data from the most recently processed dataset is divided into a training dataset and a test dataset. The training dataset is used to train a separate model created for each of the multiple processing modes, and the test dataset is used to compare the trained model with a previous model for each processing mode. This is known as k-fold cross-validation. For example, the invention employs a factor of 5, wherein the dataset is divided into 4 training datasets and 1 test dataset. Advantageously, the training dataset and / or the test dataset is normalized.

[0024] In an advantageous variant of the invention, if the corresponding comparison improves the prediction accuracy, the previous model is replaced with the trained model.

[0025] According to a preferred variant of the invention, the corresponding prior model is only replaced by the corresponding trained model after manual release. This is particularly useful when the trained model does not yield better or even worse comparison results than the prior model.

[0026] According to an alternative variation of the invention, the corresponding prior model is automatically replaced by the corresponding trained model. This is especially true when the trained model provides significantly better comparison results than the prior model.

[0027] In an advantageous variant of the invention, the corresponding previous model is stored after the replacement, particularly along with a timestamp. Therefore, the previous model can be used at any time.

[0028] According to a variation of the invention, the created standalone model is initialized and trained based on historical data or data from comparable pig iron desulfurization.

[0029] According to an advantageous variation of the invention, the method also takes into account the chemical analysis of the pig iron to be desulfurized. This further improves the optimization of pig iron desulfurization.

[0030] In a preferred embodiment of the invention, multiple models are generated for each processing mode, wherein the models are constructed as neural networks. Therefore, several models are available for each processing mode, and for pig iron desulfurization, the model that provides the best results is selected.

[0031] According to a favorable variant, the hyperparameters are selected from: number of layers, number of nodes, number of subsets for k-fold cross-validation, number of training cycles, allowable deviation of sulfur content from target sulfur content to account for the processing during the training phase, learning rate, and momentum.

[0032] The following will provide a more detailed explanation of this objective in conjunction with the embodiments shown in the figures. Wherein: Figure 1 A schematic diagram of pig iron desulfurization is shown; Figure 2 A schematic diagram of a first embodiment of the method for optimizing pig iron desulfurization according to the present invention is shown; Figure 3 A flowchart illustrating the training 9 of a single model 8 within the framework of the method according to the present invention is shown; and Figure 4 Figure 1 A schematic diagram of pig iron desulfurization 1 is shown. Pig iron desulfurization 1 is typically located between the blast furnace and the converter. Pig iron desulfurization 1 is carried out by distributing one or more desulfurizing agents 2, 3, and 4 into the pig iron pot 5. Each desulfurizing agent 2, 3, and 4 is stored in a corresponding hopper and distributed into the pig iron pot 5 via a related pressure conveying container 6 and conveying pipeline 7. Specifically, during the treatment process, the desulfurizing agents 2, 3, and 4 are injected into the pig iron pot 5 as powder through a spray gun 11 connected to the conveying pipeline 7.

[0033] For pig iron desulfurization 1, several treatment modes are available, differing in the types of desulfurizing agents 2, 3, and 4 used. Desulfurizing agents magnesium (Mg) 2, calcium carbide (CaC2) 3, and lime (CaO) 4 are typically found in different hoppers. For example, desulfurizing agents 2, 3, and 4 are characterized by varying desulfurization efficiencies and material costs. The order of desulfurizing agents magnesium 2, calcium carbide 3, and lime 4 is descending in terms of desulfurization efficiency and material cost.

[0034] Depending on the processing, one or more blow molding stages may occur, in which only calcium carbide 3 or lime 4 is added (single injection), or a mixture of calcium carbide 3 or lime 4 and magnesium 2 is added (co-injection). The corresponding processing modes, namely lime single injection, calcium carbide single injection, lime-magnesium co-injection, and calcium carbide-magnesium co-injection, take into account the required desulfurization work, such as based on pig iron weight, initial sulfur content, and target sulfur content, as well as existing constraints regarding material availability and cost, maximum possible blowing rate, and predetermined distribution time.

[0035] Figure 2 A schematic diagram of a first embodiment of a method for optimizing pig iron desulfurization 1 in a steel plant with a blast furnace route according to the present invention is shown. Pig iron desulfurization 1 is carried out by distributing one or more desulfurizing agents 2, 3, 4 into a pig iron pot 5 or a torpedo car, wherein several processing modes are available, differing in the type of desulfurizing agents 2, 3, 4 distributed, such as, for example, a combination of... Figure 1 Detailed description. For general design of pig iron desulfurization in this regard, refer to... Figure 1 .

[0036] According to the present invention, the method for optimizing pig iron desulfurization 1 is based on creating a separate model 8 for each of a variety of treatment modes to optimize the addition amount of desulfurizing agents 2, 3, and 4 in the corresponding treatment mode. The separate model 8 is based on machine learning to predict the sulfur content required for future pig iron desulfurization 1 and the required amount of desulfurizing agents 2, 3, and 4 to be allocated. Figure 2 A schematic diagram of a first embodiment of the method according to the invention is shown, in which a separate model 8 has been created and the method is being used to optimize pig iron desulfurization 1. Creating the separate model 8 can be considered an initialization step, and the created model 8 is then trained as described below to improve its predictive accuracy. For example, initialization is based on historical data or data from similar pig iron desulfurization 1 methods.

[0037] To optimize pig iron desulfurization, the weight of the pig iron to be desulfurized and its initial sulfur content (S) were determined. 起始 ) and the target sulfur content (S) of the pig iron to be desulfurized 目标 The above information can also be provided by higher-level production plans or process controls, which is equivalent to determination in the sense of this invention.

[0038] Based on this information, according to the present invention, pig iron desulfurization 1 is optimized by providing predictions of achievable sulfur content and the required amounts of desulfurizing agents 2, 3, and 4 to be allocated for individual models 8 of different treatment modes. For example, the treatment mode with the lowest predicted sulfur content and / or the lowest consumption cost of desulfurizing agents 2, 3, and 4 is selected.

[0039] Optimizing pig iron desulfurization 1 based on the created model 8 can effectively consider the maximum possible addition rate and / or possible distribution time of several desulfurizing agents 2, 3, and 4, as well as the available quantity and / or cost of several desulfurizing agents 2, 3, and 4. According to the invention, this method determines the amount of desulfurizing agents 2, 3, and 4 required for each treatment mode to achieve the target sulfur content by predicting a specific amount of pig iron to be desulfurized using a separate model 8 with a specific initial sulfur content, particularly selecting the treatment mode that is convertible and has the lowest cost based on the available quantity of desulfurizing agents 2, 3, and 4.

[0040] After or during pig iron desulfurization 1, the types of desulfurizing agents 2, 3, and 4, the amounts of desulfurizing agents 2, 3, and 4 allocated, and the actual sulfur content (S) achieved are detected. 结束 Furthermore, the weight of the pig iron can be measured during pig iron desulfurization 1, as this may differ from the determined or predetermined weight of the pig iron to be desulfurized. Data collected regarding pig iron desulfurization 1, particularly the weight of the pig iron to be desulfurized and the initial sulfur content (S) of the pig iron to be desulfurized, are crucial. 起始 ) and the target sulfur content (S) of the pig iron to be desulfurized 目标The types of desulfurizing agents 2, 3, and 4 allocated, the amounts of desulfurizing agents 2, 3, and 4 allocated, and the actual sulfur content achieved (S) 结束 ), stored in database 10.

[0041] After performing a predetermined amount of pig iron desulfurization 1 using the corresponding processing mode, a related model 8 is trained for that processing mode. Training is conducted based on information stored in database 10, namely, determining the weight of the pig iron to be desulfurized, the initial sulfur content of the pig iron to be desulfurized, and the target sulfur content of the pig iron to be desulfurized, and detecting the types and amounts of the allocated desulfurizing agents 2, 3, and 4, as well as the actual sulfur content achieved after pig iron desulfurization 1. The weight of pig iron that can be optionally detected during pig iron desulfurization 1 can also be considered in the training 9 of model 8. Each model 8 is trained independently because each processing mode will achieve the predetermined amount of pig iron desulfurization 1 at different time points. The performed pig iron desulfurization 1 and the created model 8 are respectively associated with specific processing modes, ensuring that models 8 for different processing modes do not interfere with each other.

[0042] When creating a separate model 8 for each of the multiple processing modes, only data from pig iron desulfurization 1 that achieved the target sulfur content with a predetermined accuracy are considered. Therefore, pig iron desulfurization 1 that failed to achieve the target sulfur content with a predetermined accuracy is ignored. Furthermore, the number of pig iron desulfurization 1s considered in training 9 can be limited, particularly the number of most recently performed pig iron desulfurization 1s.

[0043] After training model 8 (9), the trained model 8 for the processing mode is compared with the previous model 8 for the processing mode. For example, the comparison can be based on data from a certain number of recent processing iterations and / or on a generally generated input dataset. If the comparison results in improved prediction accuracy, the previous model 8 is replaced with the trained model 8. The replacement can be performed manually by the user or automatically. The replaced previous model 8 is conveniently stored, for example, in database 10.

[0044] According to an advantageous variation of the invention, multiple models 8 are generated for each processing mode, wherein the models 8 are constructed as neural networks. For example, hyperparameters are selected from: number of layers, number of nodes, number of subsets for k-fold cross-validation, number of training epochs, allowable deviation between the sulfur content and the target sulfur content to take into account the processing during the training phase, learning rate, and momentum.

[0045] Figure 3 A flowchart illustrating the training 9 of a single model 8 within the framework of the method according to the present invention is shown. The abbreviations used in the flowchart are explained below: Historical processing of the model training dataset: For each processing mode, a dataset D is first composed of all historical processes i (N 处理，新，i = N 处理，全部，i ). The dataset is then reduced to processes that achieve the target sulfur content S with good precision 目标 , i.e., processes satisfying |S 目标 -S 结果 | / S 目标 <ABW. For the allowable deviation ABW, iteration is performed using the values {0.01, 0.02, 0.05, 0.1, 0.2, 0.5, 1.0, 1.5, 2.0}, so as to obtain different sizes of the resulting dataset D i,red .

[0046] Subsets and complementary subsets of the dataset for testing or training: To reduce the probability of overfitting of the regression model 8, k-fold cross-validation (here k=5) is adopted. Here, the k-th subset K in D is sequentially used as the test dataset, and the set D i,red \K is used as the training dataset. The test and training datasets are subjected to normalization processing. i,red

[0047] Then, further iteration is performed using predefined values for certain hyperparameters of the algorithm.

[0048] For each combination of ABW, the division of test and training datasets, and hyperparameters, the initialization and training 9 of the model 8 are constructed as a neural network.

[0049] The input parameters include chemical analysis, temperature and the amount of pig iron to be desulfurized. The output parameters are the amounts of desulfurizing agents 2, 3, and 4.

[0050] The number and size of hidden layers can be fixedly specified, or can be set as hyperparameters for iteration. During each iteration, the resulting neural network, i.e., the determined weights and the associated test error determined using the corresponding test dataset, are recorded. After all iterations are completed, the neural network with the minimum test error is regarded as the new regression model 8 M for processing mode i 新，i . When comparing the new regression model 8 M 新，i with the previously used M 旧，i , the test error based on the unreduced dataset Di is used again.

[0051] According to the flowchart shown in Figure 3 , the sequence of steps is applied to a dataset of 4898 processes, where ABW=1.1 is processing mode 3 based on real operator data. In each case, the first 1000, 2000, or 3000 processes are used for model training 9, so as to simulate continuous model adjustment during operation. Figure 4 ​The specifications for the corresponding determined regression model 8 and the actual amounts of lime 4 and magnesium 2 (in kilograms) added are shown. The columns from left to right contain the model results, sorted in ascending order by the number of treatments on which training 9 was based. The comparisons shown include all treatments not considered in the training or testing of the corresponding model 8. Thus, successive model improvements are achieved from left to right. The corresponding error calculations are as follows. List of reference numerals 1. Desulfurization of pig iron 2. Desulfurizing agent (magnesium-Mg) 3. Desulfurizing agent (calcium carbide - CaC2) 4. Desulfurizing agent (lime-CaO) 5. Cast iron pot 6. Pressure Transfer Container 7. Delivery pipeline 8 Models 9 Training 10 Databases 11 Spray gun

Claims

1. A method for optimizing desulfurization of pig iron (1), wherein, The desulfurization (1) of the pig iron is achieved by distributing one or more desulfurizing agents (2, 3, 4) into the pig iron pot (5) or torpedo car, wherein, depending on the type of desulfurizing agent (2, 3, 4) distributed, several processing modes are available, including the following steps: Determine the weight of the pig iron to be desulfurized, the initial sulfur content of the pig iron to be desulfurized, and the target sulfur content of the pig iron to be desulfurized; The type of desulfurizing agent (2, 3, 4) and the amount of desulfurizing agent (2, 3, 4) distributed are detected, as well as the sulfur content actually reached during and after the desulfurization of pig iron (1), for the purpose of carrying out the desulfurization of pig iron (1). A separate model (8) is created for each of the multiple processing modes to optimize the amount of desulfurizing agent (2, 3, 4) added for the corresponding processing mode, wherein the separate model (8) is based on machine learning to predict the achievable sulfur content and the amount of desulfurizing agent (2, 3, 4) to be allocated for future pig iron desulfurization (1), respectively. After performing a predetermined number of pig iron desulfurization (1) using the corresponding processing modes, for each of the multiple processing modes, a separate model (8) created by (9) is trained, wherein the training (9) of the separate model (8) is based on determining the weight of the pig iron to be desulfurized, the initial sulfur content of the pig iron to be desulfurized, and the target sulfur content of the pig iron to be desulfurized, and detecting the type of the allocated desulfurizing agent (2, 3, 4), the amount of the allocated desulfurizing agent (2, 3, 4), and the actual sulfur content reached after performing the pig iron desulfurization (1), for performing the pig iron desulfurization (1); and Optimize future pig iron desulfurization (1) based on the predictions of a single model (8), wherein the optimization is determined by comparing the predictions of the single model (8) to determine the treatment mode and the amount of desulfurizing agent (2, 3, 4) added in the treatment mode.

2. The method according to claim 1, wherein, The desulfurizing agents (2, 3, 4) are selected from magnesium, calcium carbide and / or lime.

3. The method according to claim 2, wherein, The method considers the following processing modes: lime single injection, calcium carbide single injection, lime-magnesium co-injection, calcium carbide-magnesium co-injection, and calcium carbide-magnesium co-injection with upstream and downstream lime single injection.

4. The method according to any one of claims 1 to 3, wherein, The standalone model (8) was constructed into a regression model.

5. The method according to any one of claims 1 to 4, wherein, The method takes into account the maximum possible addition rate and / or possible allocation time of several desulfurizing agents (2, 3, 4) during the optimization process.

6. The method according to any one of claims 1 to 5, wherein, The standalone model (8) takes into account the availability, quantity and / or cost of multiple desulfurizing agents (2, 3, 4) during the optimization process.

7. The method according to any one of claims 1 to 6, wherein, The method determines the amount of desulfurizing agent (2, 3, 4) required to achieve the target sulfur content for each treatment mode by predicting a separate model (8) of a certain amount of pig iron to be desulfurized with a specific initial sulfur content, and in particular selects or recommends the treatment mode that is convertible and has the lowest cost based on the amount of desulfurizing agent (2, 3, 4) available.

8. The method according to any one of claims 1 to 7, wherein, When the training (9) creates a separate model (8) for each of the multiple processing modes, only the data of pig iron desulfurization (1) that achieves the target sulfur content with a predetermined accuracy are considered.

9. The method according to any one of claims 1 to 8, wherein, When the training (9) is performed on a separate model (8) created for each of the multiple processing modes, only the data from a certain number of recent processing sessions are considered.

10. The method according to any one of claims 1 to 9, wherein, After the training (9) creates a separate model (8) for each of the multiple processing modes, the trained model (8) is compared with the previous model (8) for each processing mode.

11. The method according to claim 10, wherein, The comparison is based on data from a certain number of recent processing sessions and / or on a generally generated input dataset.

12. The method according to claim 11, wherein, The recently processed data is divided into a training dataset and a test dataset, wherein the training dataset is used to train (9) a separate model (8) created for each of the multiple processing modes, and the test dataset is used to compare the trained model (8) with the previous model (8) for each processing mode.

13. The method according to any one of claims 10 to 12, wherein, If the corresponding comparison improves the prediction accuracy, then the previous model (8) is replaced with the trained model (8).

14. The method according to claim 13, wherein, The corresponding previous model (8) is replaced by the corresponding trained model (8) only after manual release.

15. The method according to claim 13, wherein, The corresponding previous model (8) is automatically replaced by the corresponding trained model (8).

16. The method according to any one of claims 12 to 15, wherein, After replacement, the corresponding previous model (8) is stored, in particular along with the timestamp.

17. The method according to any one of claims 1 to 16, wherein, The created individual model (8) is initialized and trained based on historical data or data from comparable pig iron desulfurization (1).

18. The method according to any one of claims 1 to 17, wherein, The method also takes into account the chemical analysis of the pig iron to be desulfurized.

19. The method according to any one of claims 1 to 18, wherein, Multiple models (8) are generated for each processing mode, wherein the models (8) are constructed as neural networks and are different from each other in terms of their hyperparameters.

20. The method according to claim 19, wherein, The hyperparameters are selected from: number of layers, number of nodes, number of subsets for k-fold cross-validation, number of training cycles, allowable deviation between sulfur content and target sulfur content to account for processing during the training phase, learning rate, and momentum.

21. A system for optimizing desulfurization of pig iron (1), wherein, The desulfurization of the pig iron is achieved by distributing one or more desulfurizing agents (2, 3, 4) into the pig iron pot (5) or torpedo car, wherein several processing modes are available depending on the type of desulfurizing agent (2, 3, 4) distributed, and wherein the system is configured to perform the method of the invention according to any one of claims 1 to 20.

22. A computer program comprising, when a computer executes the program, causing the computer to perform a method according to any one of claims 1 to 20.

Images