Method for optimizing a pig iron desulfurization process

Abstract

The invention relates to a method for optimizing a pig iron desulfurization process (1), wherein the pig iron desulfurization process (1) is carried out by dispensing one or more desulfurization agents (2, 3, 4) into a pig iron ladle (5) or a torpedo car, wherein a plurality of treatment patterns are available which differ with respect to the type of desulfurization agents (2, 3, 4) dispensed. The optimization is based on creating a separate model (8) for each of the plurality of treatment patterns in order to optimize the added amounts of the desulfurization agents (2, 3, 4) for the associated treatment pattern, and training (9) the created separate models (8) for each of the plurality of treatment patterns after a predefined number of performed pig iron desulfurization processes (1) with the corresponding treatment pattern. Future pig iron desulfurization processes (1) are optimized on the basis of the predictions of the separate models (8), wherein the optimization determines a treatment pattern and the added amounts of the desulfurization agents (2, 3, 4) for the treatment pattern by means of a comparison of the predictions of the separate models (8). The invention also relates to a system and a computer program which are designed to carry out the method according to the invention.

Description

[0001] Process for optimizing pig iron desulfurization

[0002] The invention relates to a method for optimizing pig iron desulfurization, wherein the pig iron desulfurization is carried out by discharging one or more desulfurizing agents into a pig iron ladle or a torpedo car, wherein several treatment patterns are available which differ with regard to the type of desulfurizing agents discharged.

[0003] Pig iron desulfurization is typically located between the blast furnace and the converter. The desulfurizing agents magnesium (Mg), carbide (CaC2), and lime (CaO) are typically available in different bunkers. The desulfurizing agents can be characterized, for example, by their different desulfurizing effectiveness and material costs. The desulfurizing agents magnesium, carbide, and lime are arranged in descending order of desulfurizing effectiveness and material costs.

[0004] During treatment, the desulfurization agents are injected into the hot metal ladle or torpedo car via a powder lance. Depending on the treatment, one or more blowing phases may occur, in which either only carbide or lime (mono-injection) or a mixture of carbide or lime with magnesium (co-injection) is added. The respective treatment patterns—lime mono-injection, carbide mono-injection, lime-magnesium co-injection, and carbide-magnesium co-injection—take into account the required desulfurization work, for example, based on the hot metal weight, the starting and target sulfur content, as well as existing constraints regarding material availability and costs, maximum possible blowing rates, and specified release times.

[0005] Analytical and numerical process models are known from practice to support the operator in selecting the treatment pattern to be used and the quantities of desulfurizing agents to be added. These process models usually contain empirical approaches that describe the quantities of desulfurizing agents to be added as a function of several process variables, such as the amount of hot metal, hot metal temperature, starting sulfur content, and target sulfur content. Such empirical approaches naturally have limited accuracy, particularly in this case because plant-specific influences such as yield factors, which can change during operation, must also be taken into account. Ideally, a plant-specific adaptation of the process model based on historical treatments is carried out during initial commissioning. However, further model adaptations during ongoing operation often prove difficult.This may be due to one or more of the following reasons: a lack of transparency in the model used or the influences of individual model parameters, highly nonlinear relationships between the model's input and output variables, and the need for structured analysis of large amounts of data. As a result, some treatments involve the inappropriate addition of desulfurizing agents, which is cost-effective and / or necessitates post-treatment.

[0006] It is therefore an object of the invention to optimize the addition quantities of the individual desulfurizing agents in pig iron desulfurization so that the target sulfur content is achieved as reliably and cost-effectively as possible in each treatment.

[0007] The object is achieved according to the invention by a method according to claim 1. Advantageous embodiments of the method are defined in the dependent claims. The object is further achieved by a system according to claim 21 and a computer program according to claim 22.

[0008] The invention relates to a method for optimising pig iron desulfurisation, wherein the pig iron desulfurisation is carried out by discharging one or more desulfurising agents into a pig iron ladle or a torpedo car, wherein several treatment patterns are available which differ in terms of the type of desulfurising agents discharged, comprising the steps:

[0009] Determining the weight of the pig iron to be desulfurized, the starting sulfur content of the pig iron to be desulfurized and the target sulfur content of the pig iron to be desulfurized;

[0010] Recording the type of desulfurizing agents released, the quantities of desulfurizing agents released and the sulfur content actually achieved after the pig iron desulfurization has been completed for the pig iron desulfurizations carried out;

[0011] Creating a separate model for each of the multiple treatment patterns to optimize the addition rates of desulfurizing agents for the respective treatment pattern, wherein the separate models are based on machine learning and each predicts the achievable sulfur contents for future hot metal desulfurizations and the required amount of desulfurizing agent to be added;

[0012] Training the created separate models for each of the plurality of treatment patterns after a predetermined number of pig iron desulfurizations have been carried out with the corresponding treatment pattern, wherein the training of the separate models is based on the determination of the weight of the pig iron to be desulfurized, the starting sulfur content of the pig iron to be desulfurized and the target sulfur content of the pig iron to be desulfurized and the recording of the type of desulfurizing agents released, the quantities of desulfurizing agents released and the actually achieved sulfur content after the pig iron desulfurization has been carried out for the pig iron desulfurizations carried out; and

[0013] Optimizing future hot metal desulfurization based on the predictions of the separate models, whereby the optimization determines a treatment pattern and the addition quantities of the desulfurizing agents for the treatment pattern by comparing the predictions of the separate models.

[0014] In a first step, the weight of the pig iron to be desulfurized, the starting sulfur content of the pig iron to be desulfurized and the target sulfur content of the pig iron to be desulfurized are determined for each pig iron desulfurization.

[0015] According to the method according to the invention, certain data are recorded during the conduct of pig iron desulfurizations, in particular the type of desulfurizing agent released, the quantities of desulfurizing agents released, and the actual sulfur content achieved after the pig iron desulfurization for the pig iron desulfurizations performed. Furthermore, the pig iron weight during the pig iron desulfurization is recorded. Historical data on the pig iron desulfurizations performed are thus collected, with the collected historical data relating to the parameters relevant to the pig iron desulfurization. The collected historical data is stored, for example, in a database.

[0016] For each of the multiple treatment patterns, a separate model is created to optimize the addition rates of the desulfurizing agents for the respective treatment pattern. The separate models are based on different data, as each model only considers the data relevant to the associated treatment pattern and neglects the data relating to the other treatment patterns. Furthermore, data is preferably not exchanged between the separate models. The separate models are based on machine learning and predict the achievable sulfur contents for future hot metal desulfurizations and the required amount(s) of desulfurizing agent to be added.

[0017] The separate models created according to the invention for each of the multiple treatment patterns are trained after a predetermined number of hot metal desulfurizations have been performed with the corresponding treatment pattern. The training of the separate models is based on the weight of the hot metal to be desulfurized, the starting sulfur content of the hot metal to be desulfurized, the target sulfur content of the hot metal to be desulfurized, and the recorded associated historical data, in particular the weight of the hot metal during the hot metal desulfurization, the type of desulfurizing agent released, the quantities of the desulfurizing agent released, and the sulfur content actually achieved after the hot metal desulfurization for the hot metal desulfurizations performed. The training can also take into account the optionally recorded weight of the hot metal during the hot metal desulfurization.When training the separate models, only data from completed hot metal desulfurization processes with the corresponding treatment pattern are considered, so that the different treatment patterns are strictly separated from each other and the models are also trained separately. This eliminates any mutual influence of the treatment patterns or models.

[0018] Using the created models, future hot metal desulfurization processes can be predicted and optimized. The result of the optimization is the determination of a treatment pattern and the corresponding addition quantities of desulfurizing agent for that treatment pattern. According to the invention, the optimization is carried out by comparing the predictions of the separate models for the different treatment patterns.

[0019] According to a suitable variant of the invention, the desulfurization agents are selected from: magnesium, carbide, and / or lime. Theoretically, desulfurization could be carried out with any substance that has a higher affinity for sulfur than iron itself (e.g., magnesium, cerium, calcium, sodium, etc.). For economic and technical reasons, the aforementioned desulfurization agents magnesium, carbide, and / or lime are preferred in practice. According to a variant of the invention, the process takes into account the following treatment patterns: lime mono-injection, carbide mono-injection, lime-magnesium co-injection, carbide-magnesium co-injection, and carbide-magnesium co-injection with upstream and downstream lime mono-injection.

[0020] In a preferred variant of the invention, the separate models are designed as regression models. A regression model is based on statistical analysis methods and describes the relationships between a dependent variable (predicted variable) and one or more independent variables (control variable or regressor). Regression models are used to quantitatively describe relationships and predict values ​​of the dependent variables. In particular, the invention relates to linear regression models. The regression models can be implemented, in particular, using machine learning.

[0021] According to a variant of the invention, the method takes into account the maximum possible addition rates and / or possible release times for the multiple desulfurizing agents during optimization. This ensures that the selected treatment pattern can be technically implemented in practice.

[0022] According to a further variant of the invention, the separate models consider the availability, available quantity, and / or cost of the multiple desulfurizing agents during optimization. Therefore, if the required quantities of desulfurizing agent are not available for a treatment pattern, the corresponding model does not provide a result that is considered during optimization. By considering the costs of the desulfurizing agents, the corresponding material costs can be determined for the individual treatment patterns, which differ in terms of the desulfurizing agents used.

[0023] In a suitable variant of the invention, the method uses the predictions of the separate models for a specific quantity of pig iron to be desulfurized with a specific starting sulfur content to determine the required quantities of desulfurizing agents for each treatment pattern to achieve the target sulfur content. In particular, the treatment pattern that is feasible based on the available quantities of desulfurizing agents and associated with the lowest costs is subsequently selected or suggested.

[0024] According to an advantageous variant of the invention, when training the created separate models for each of the multiple treatment patterns, only data from completed hot metal desulfurizations in which the target sulfur content was achieved with a predetermined accuracy are taken into account. The predetermined accuracy can thus be used to define which hot metal desulfurizations are taken into account when training the models. Only hot metal desulfurizations with sufficient accuracy regarding the target sulfur content are taken into account. This ensures that the hot metal desulfurization predicted by the model has a comparable accuracy regarding the target sulfur content. For training, in particular, those hot metal desulfurizations that cannot be identified as outliers based on the descriptive statistics criteria are taken into account.

[0025] According to a preferred variant of the invention, when training the created separate models for each of the multiple treatment patterns, only the data from a specific number of the most recently performed treatments are taken into account. In this way, hot metal desulfurizations from the more recent past are taken more into account in the model training. As a result, slowly changing conditions during hot metal desulfurization, in particular changes to the hot metal desulfurization plant, are taken into account in the model training. Data from older hot metal desulfurizations with outdated plant conditions are neglected according to this variant, and only more recent data on hot metal desulfurizations are taken into account in the model training. In a variant according to the invention, after training the created separate models for each of the multiple treatment patterns, the trained models are compared with the previous models for each treatment pattern.After a model has been trained, the trained model is compared with the corresponding previous model. This ensures that the trained model is at least as good as the previous model.

[0026] According to an expedient variant of the invention, the comparison is based on the data of a specific number of the most recently performed treatments and / or on the basis of generically generated input data sets. In particular, the comparison of the trained model with the previous model is carried out for a specific number of random, generically generated input data sets within predefined intervals (Monte Carlo simulations).

[0027] According to a particularly advantageous variant of the invention, the data from the most recently performed treatments are divided into training data sets and test data sets. The training data sets are used to train the created separate models for each of the multiple treatment patterns, and the test data sets are used to compare the trained models with the previous models for each treatment pattern. This is a so-called k-fold cross-validation. For example, the invention uses a factor of 5, whereby the data set is divided into 4 training data sets and 1 test data set. The training and / or test data sets are expediently normalized.

[0028] In a suitable variant of the invention, the previous models are replaced by the trained models if the respective comparison provides an improvement in the prediction accuracy.

[0029] According to a preferred variant of the invention, the corresponding previous model is replaced by the corresponding trained model only after manual approval. This is particularly useful if the trained model does not show better or even worse comparison results than the previous model.

[0030] According to an alternative variant of the invention, the corresponding previous model is automatically replaced by the corresponding trained model. This occurs particularly if the trained model delivers clearly better comparison results than the previous model.

[0031] In a practical variant of the invention, the corresponding previous model is saved after the replacement, in particular together with a timestamp. Thus, the previous models can be accessed at any time.

[0032] According to a variant of the invention, the separate models created are trained for initialization on the basis of historical data or data from comparable pig iron desulfurizations.

[0033] According to an advantageous variant of the invention, the process further takes into account the chemical analysis of the pig iron to be desulfurized. This further improves the optimization of the pig iron desulfurization process.

[0034] In a preferred variant of the invention, several models are generated for each treatment pattern, with the models being designed as neural networks. Thus, several models are available for each treatment pattern, with the model that delivers the best results being selected for pig iron desulfurization.

[0035] According to a convenient variant, the hyperparameters are selected from: number of layers, number of nodes, number of subsets of a k-fold cross-validation, number of training epochs, permissible deviation of the achieved sulfur content from the target sulfur content to consider a treatment for the training phase, learning rate and momentum.

[0036] The object is explained in more detail below using exemplary embodiments illustrated in the figures. They show: Fig. 1 shows a schematic view of a pig iron desulfurization process;

[0037] Fig. 2 is a schematic view of a first embodiment of the method according to the invention for optimizing pig iron desulfurization;

[0038] Fig. 3 is a flowchart of a training 9 of separate models 8 within the framework of the method according to the invention; and

[0039] Fig. 4

[0040] Fig. 1 shows a schematic view of a hot metal desulfurization plant 1. The hot metal desulfurization plant 1 is typically located between the blast furnace and the converter. Hot metal desulfurization 1 is carried out by discharging one or more desulfurizing agents 2, 3, 4 into a hot metal ladle 5. The individual desulfurizing agents 2, 3, 4 are stored in appropriate bunkers and discharged into the hot metal ladle 5 via associated pressure conveying vessels 6 and a conveying line 7. In particular, the desulfurizing agents 2, 3, 4 are introduced into the hot metal ladle 5 during treatment by powder injection via a blowing lance 11 connected to the conveying line 7.

[0041] For pig iron desulfurization, several treatment models are available, which differ in the type of desulfurizing agents 2, 3, and 4 released. Typically, the desulfurizing agents magnesium (Mg) 2, carbide (CaC2) 3, and lime (CaO) 4 are available in different bunkers. Desulfurizing agents 2, 3, and 4 can be characterized, for example, by their different desulfurization effectiveness and material costs. The order of the desulfurizing agents magnesium 2, carbide 3, and lime 4 represents a descending order of desulfurization effectiveness and material costs.

[0042] Depending on the treatment, one or more blowing phases may occur, in which either only Carbide 3 or Lime 4 (mono-injection) or a mixture of Carbide 3 or Lime 4 with Magnesium 2 (co-injection) is added. The respective treatment patterns, i.e., lime mono-injection, carbide mono-injection, lime-magnesium co-injection, and carbide-magnesium co-injection, take into account the required desulfurization work, for example, based on the pig iron weight, the starting and target sulfur content, as well as existing constraints regarding material availability and costs, maximum possible blowing rates, and specified discharge times.

[0043] Fig. 2 shows a schematic view of a first embodiment of the method according to the invention for optimizing hot metal desulfurization 1 in a steelworks with a blast furnace route. Hot metal desulfurization 1 is carried out by discharging one or more desulfurizing agents 2, 3, 4 into the hot metal ladle 5 or the torpedo car. Several treatment patterns are available, which differ in the type of desulfurizing agents 2, 3, 4 discharged, as described in detail, for example, in connection with Fig. 1. Therefore, reference is made to Fig. 1 regarding the general design of the hot metal desulfurization.

[0044] The inventive method for optimizing a pig iron desulfurization 1 is based on the creation of separate models 8 for each of the multiple treatment patterns to optimize the addition quantities of the desulfurizing agents 2, 3, 4 for the respective treatment pattern. The separate models 8 are based on machine learning and each predict the achievable sulfur contents for future pig iron desulfurizations 1 and the required quantities of desulfurizing agents 2, 3, 4 to be released. Fig. 2 shows a schematic view of a first embodiment of the inventive method at a time at which the separate models 8 have already been created and the method for optimizing the pig iron desulfurization 1 is being used. The creation of the separate models 8 can be regarded as an initialization step, and the created models 8 are subsequently trained as described below and improved with regard to prediction accuracy.The initialization is carried out, for example, on the basis of historical data or data from comparable pig iron desulfurizations 1 .

[0045] To optimize the desulfurization of pig iron 1, the weight of the pig iron to be desulfurized, the starting sulfur content (Sstart) of the pig iron to be desulfurized, and the target sulfur content (Sziei) of the pig iron to be desulfurized are determined. This information can also be provided by a higher-level production planning or process control system, which is equivalent to a determination within the meaning of the invention.

[0046] Based on this information, the hot metal desulfurization 1 is optimized according to the invention by the separate models 8 providing a prediction for the achievable sulfur content and the required quantity(s) of desulfurizing agent(s) 2, 3, 4 to be delivered for the different treatment patterns. For example, the treatment pattern is selected that predicts the lowest sulfur content and / or is associated with the lowest costs in terms of the consumption of desulfurizing agent(s) 2, 3, 4.

[0047] The optimization of pig iron desulfurization 1 on the basis of the created models 8 expediently takes into account the maximum possible addition rates and / or possible release times for the plurality of desulfurization agents 2, 3, 4 as well as the available quantity and / or the costs of the plurality of desulfurization agents 2, 3, 4. The method according to the invention uses the predictions of the separate models 8 to determine the required quantities of desulfurization agents 2, 3, 4 for each treatment pattern in order to achieve the target sulfur content for a specific quantity of pig iron to be desulfurized with a specific starting sulfur content and, in particular, selects the treatment pattern which can be implemented on the basis of the available quantities of desulfurization agents 2, 3, 4 and which is associated with the lowest costs. After a successful orDuring a pig iron desulfurization 1, the type(s) of desulfurizing agents 2, 3, 4 released, the quantities of the desulfurizing agents 2, 3, 4 released, and the actually achieved sulfur content (S nde) are recorded. Optionally, the pig iron weight can also be recorded during the pig iron desulfurization 1, since this may differ from the determined or predetermined weight of the pig iron to be desulfurized. The collected data for a pig iron desulfurization 1, in particular the weight of the pig iron to be desulfurized, the starting sulfur content (Sstart) of the pig iron to be desulfurized and the target sulfur content (Sziei) of the pig iron to be desulfurized, the type(s) of desulfurizing agents 2, 3, 4 released, the quantities of the desulfurizing agents 2, 3, 4 released, and the actually achieved sulfur content (S nde), are stored in a database 10.

[0048] After a predetermined number of hot metal desulfurizations 1 have been performed with a corresponding treatment pattern, the associated model 8 is trained for this treatment pattern. Training is performed based on the information stored in the database 10, i.e., the determination of the weight of the hot metal to be desulfurized, the starting sulfur content of the hot metal to be desulfurized, and the target sulfur content of the hot metal to be desulfurized, as well as the recording of the type of desulfurizing agents 2, 3, 4 released, the quantities of the desulfurizing agents 2, 3, 4 released, and the actual sulfur content achieved after hot metal desulfurization 1 has been completed. The optionally recorded hot metal weight during hot metal desulfurization 1 can also be taken into account in the training 9 of the model 8.Each model 8 is trained separately, since the predetermined number of required hot metal desulfurizations 1 are achieved at different times for the individual treatment patterns. The hot metal desulfurizations 1 performed and the created models 8 each refer to a specific treatment pattern, so that the models 8 for different treatment patterns do not influence each other. When training the created separate models 8 for each of the multiple treatment patterns, only data from hot metal desulfurizations 1 performed in which the target sulfur content was achieved with a predetermined accuracy are considered. Therefore, hot metal desulfurizations 1 in which the target sulfur content was not achieved with a predetermined accuracy are neglected.Furthermore, the number of hot metal desulphurisations 1 taken into account for training 9 can be limited, in particular to a certain number of the last hot metal desulphurisations 1 carried out.

[0049] After a model 8 has been trained 9, the trained model 8 for the treatment pattern is compared with the previous model 8 for the treatment pattern. The comparison is performed, for example, based on the data from a specific number of the most recently performed treatments and / or based on generically generated input data sets. The previous model 8 is replaced by the trained model 8 if the comparison yields an improvement in prediction accuracy. The replacement occurs either after manual approval by a user or automatically. The replaced previous model 8 is conveniently stored, for example, in the database 10.

[0050] According to an advantageous variant of the invention, several models 8 are generated for each treatment pattern, wherein the models 8 are designed as neural networks. The hyperparameters are selected, for example, from the following: number of layers, number of nodes, number of subsets of a k-fold cross-validation, number of training epochs, permissible deviation of the achieved sulfur content from the target sulfur content to consider a treatment for the training phase, learning rate, and momentum.

[0051] Fig. 3 shows a flowchart of a training 9 of separate models 8 within the framework of the method according to the invention. The abbreviations used in the flowchart are explained below:

[0052]

[0053] Selection of historical treatments for the dataset for model training:

[0054] For each treatment pattern, a data set Di is first composed of all historical treatments (new treatment, new, i — N ßehandiung, alle, i). The data set is then reduced to treatments where the target sulfur content Sziei was achieved with good accuracy, i.e. where | Sziei-Sende| / S Z iei < ABW holds. For the permissible deviation ABW, the values ​​{0.01, 0.02, 0.05, 0.1, 0.2, 0.5, 1.0, 1.5, 2.0} are iterated, resulting in different sizes of the resulting data set Di, red.

[0055] Subsets and complementary subsets of the dataset for testing and training:

[0056] To reduce the likelihood of overfitting the regression model 8, k-fold cross-validation is applied (here, k=5). In this procedure, a k-th subset K from Di,red is used successively as the test dataset and the set Di,red\K as the training dataset. The test and training datasets are normalized.

[0057] Subsequently, further iterations are performed using predefined values ​​for certain hyperparameters of the algorithm.

[0058] For each combination of ABW, division into test and training data sets, and hyperparameters, a model 8 configured as a neural network is initialized and trained 9. The input parameters are the chemical analysis, temperature, and quantity of the pig iron to be desulfurized. The output parameters are the quantities of desulfurizing agents 2, 3, and 4.

[0059] The number and size of the hidden layers can be specified as fixed values, or they can also be defined as hyperparameters over which the model is iterated. During each iteration, the resulting neural network—that is, the determined weights and the associated test error determined with the respective test dataset—is memorized. After all iterations have been completed, the neural network with the smallest test error is used as the new regression model. neu ,i for treatment pattern i.

[0060] Comparison of the new regression model 8 Mneu with the previously used Mait, i again a test error based on the non-reduced data set Di is used.

[0061] The sequence of steps according to the flowchart in Fig. 3 was applied as an example to a data set of 4898 treatments with ABW=1.1 according to treatment pattern 3 from real operator data. Model training 9 was carried out with the first 1000, 2000, and 3000 treatments, respectively, to simulate a successive model adaptation during operation. The specifications of the respective regression models 8 compared to the actually added amounts of lime 4 and magnesium 2 in kg are shown in Fig. 4. The columns from left to right contain the model results in ascending order of the number of treatments used for training 9. The comparisons shown include all treatments that were not considered in training 9 or testing the respective model 8. It can be seen that a successive model improvement is achieved from left to right. The respective error was calculated as follows:

[0062] List of reference symbols

[0063] 1 Pig iron desulfurization

[0064] 2 Desulfurizing agent (magnesium - Mg) 3 Desulfurizing agent (carbide - CaC2)

[0065] 4 Desulfurizing agents (lime - CaO)

[0066] 5 cast iron pan

[0067] 6 pressure conveying vessels

[0068] 7 Delivery line 8 Model

[0069] 9 Training

[0070] 10 Database

[0071] 11 Blow lance

Claims

Claims 1. A method for optimising pig iron desulfurisation (1), wherein the pig iron desulfurisation (1) is carried out by discharging one or more desulfurising agents (2, 3, 4) into a pig iron ladle (5) or a torpedo car, wherein several treatment patterns are available which differ in terms of the type of desulfurising agents (2, 3, 4) discharged, comprising the steps: Determining the weight of the pig iron to be desulfurized, the starting sulfur content of the pig iron to be desulfurized and the target sulfur content of the pig iron to be desulfurized; Recording the type of desulfurizing agents (2, 3, 4) released, the quantities of desulfurizing agents (2, 3, 4) released and the sulfur content actually achieved during the pig iron desulfurization (1), after the pig iron desulfurization has been carried out (1) for the pig iron desulfurizations carried out (1); Creating a separate model (8) for each of the plurality of treatment patterns to optimize the addition quantities of the desulfurizing agents (2, 3, 4) for the respective treatment pattern, wherein the separate models (8) are based on machine learning and each predict the achievable sulfur contents for future pig iron desulfurizations (1) and the required quantities of desulfurizing agents (2, 3, 4) to be delivered; Training (9) of the created separate models (8) for each of the plurality of treatment patterns after a predetermined number of pig iron desulfurizations (1) have been carried out with the corresponding treatment pattern, wherein the training (9) of the separate models (8) is based on the determination of the weight of the pig iron to be desulfurized, the starting sulphur content of the pig iron to be desulfurized and the target sulphur content of the pig iron to be desulfurized and the recording of the type of desulfurizing agents (2, 3, 4) released, the quantities of desulfurizing agents (2, 3, 4) released and the sulphur content actually achieved after the pig iron desulfurization (1) has been completed for the pig iron desulfurizations (1) carried out; and Optimizing future pig iron desulfurizations (1) based on the predictions of the separate models (8), wherein the optimization determines a treatment pattern and the addition quantities of the desulfurization agents (2, 3, 4) for the treatment pattern by comparing the predictions of the separate models (8).

2. Process according to claim 1, wherein the desulfurizing agents (2, 3, 4) are selected from: magnesium, carbide and / or lime.

3. The method of claim 2, wherein the method takes into account the following treatment patterns: lime mono-injection, carbide mono-injection, lime-magnesium co-injection, carbide-magnesium co-injection and carbide-magnesium co-injection with upstream and downstream lime mono-injection.

4. Method according to one of claims 1 to 3, wherein the separate models (8) are designed as regression models.

5. Method according to one of claims 1 to 4, wherein the method in the optimization the maximum possible addition rates and / or possible Release times for the several desulfurization agents (2, 3, 4) are taken into account.

6. Method according to one of claims 1 to 5, wherein the separate models (8) take into account the availability, the available quantity and / or the costs of the plurality of desulfurization agents (2, 3, 4) in the optimization.

7. Method according to one of claims 1 to 6, wherein the method, by means of the predictions of the separate models (8), determines the required quantities of desulfurizing agents (2, 3, 4) for each treatment pattern in order to achieve the target sulfur content for a specific quantity of pig iron to be desulfurized with a specific starting sulfur content and, in particular, selects or suggests the treatment pattern which is feasible on the basis of the available quantities of desulfurizing agents (2, 3, 4) and which is associated with the lowest costs.

8. Method according to one of claims 1 to 7, wherein, during training (9) of the created separate models (8) for each of the plurality of treatment patterns, only data from carried out pig iron desulfurizations (1) are taken into account in which the target sulfur content was achieved with a predetermined accuracy.

9. Method according to one of claims 1 to 8, wherein when training (9) the created separate models (8) for each of the plurality of treatment patterns, only the data of a certain number of the last treatments carried out are taken into account.

10. The method according to any one of claims 1 to 9, wherein after training (9) the created separate models (8) for each of the plurality of treatment patterns, the trained models (8) are compared with the previous models (8) for each treatment pattern.

11. Method according to claim 10, wherein the comparison is made on the basis of the data of a certain number of the last treatments performed and / or on the basis of generically generated input data sets.

12. The method according to claim 11, wherein the data of the last treatments performed are divided into training data sets and test data sets, wherein the training data sets are used to train (9) the created separate models (8) for each of the plurality of treatment patterns and the test data sets are used to compare the trained models (8) with the previous models (8) for each treatment pattern.

13. Method according to one of claims 10 to 12, wherein the previous models (8) are replaced by the trained models (8) if the respective comparison provides an improvement in the prediction accuracy.

14. The method according to claim 13, wherein the corresponding previous model (8) is replaced by the corresponding trained model (8) only after a 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. Method according to one of claims 12 to 15, wherein the corresponding previous model (8) is stored after the replacement, in particular together with a time stamp.

17. Method according to one of claims 1 to 16, wherein the created separate models (8) are trained for initialization on the basis of historical data or data from comparable pig iron desulfurizations (1).

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

19. Method according to one of claims 1 to 18, wherein a plurality of models (8) are generated for each treatment pattern, wherein the models (8) are designed as neural networks and differ from one another with regard to their hyperparameters.

20. The method of claim 19, wherein the hyperparameters are selected from: number of layers, number of nodes, number of subsets of a k-fold cross-validation, number of training epochs, allowable deviation of the achieved sulfur content from the target sulfur content to consider a treatment for the training phase, learning rate and momentum.

21. System for optimising pig iron desulfurisation (1), wherein the pig iron desulfurisation is carried out by discharging one or more desulfurising agents (2, 3, 4) into a pig iron ladle (5) or a torpedo car, wherein several treatment patterns are available which differ with regard to the type of desulfurising agents (2, 3, 4) discharged, wherein the system is designed to carry out the method according to the invention according to one of claims 1 to 20.

22. A computer program comprising instructions which, when executed by a computer, cause the computer to carry out the method according to any one of claims 1 to 20.

Images