Method for predicting, regulating and controlling temperature of blast furnace molten iron

Abstract

The invention relates to a blast furnace molten iron temperature prediction and regulation method, and belongs to the technical field of blast furnace ironmaking production methods. According to the technical scheme, the method comprises the steps that relevant parameters of the blast furnace molten iron temperature are collected, preprocessed, normalized and subjected to feature engineering processing, and a KAN-LSTM-Attention molten iron temperature prediction model is constructed; relevant regulation and control rule knowledge of the blast furnace molten iron temperature is collected, and a rule base containing fuzzy judgment rules and operation rules is constructed in combination with production data and expert experience; and dynamic optimization of the rule base is realized through a fuzzy control method, a molten iron temperature decision model is constructed, fuzzy reasoning and defuzzification operation are performed based on a predicted difference value and related parameters, and an optimal control strategy is selected to realize closed-loop regulation and control of the molten iron temperature. The method has the advantages that the model structure is simplified, deployment feasibility is improved, and accurate control over the temperature of the molten iron in the iron and steel industry is achieved.

Description

Technical Field

[0001] This invention relates to a method for predicting and controlling the temperature of molten iron in a blast furnace, belonging to the technical field of blast furnace ironmaking production methods. Background Technology

[0002] Hot metal temperature is a key indicator of the thermal state of a blast furnace, and its changes directly reflect the heat transfer and energy balance trends within the furnace. Blast furnace smelting is an extremely complex metallurgical process conducted under high temperature, high pressure, and closed conditions, characterized by large time delays, multiphase states, strong coupling, nonlinearity, and a lack of internal information. Constant transfer of heat, mass, and momentum occurs between the solid, liquid, and gas phases. The complex blast furnace ironmaking process generates a large volume of dispersed data, resulting in low overall data utilization. Solving this problem would be of significant value to many aspects of blast furnace ironmaking, including quality control, operational optimization, cost reduction and efficiency improvement, equipment protection, and safe production. In recent years, with the rapid development of artificial intelligence technology, its application in hot metal temperature prediction has become a research hotspot. Summary of the Invention

[0003] The purpose of this invention is to provide a method for predicting and controlling the temperature of molten iron in blast furnaces. This method constructs a self-learning control system based on a prediction model and fuzzy control theory. Through fuzzification processing, rule formulation, and dynamic reasoning, closed-loop temperature control is achieved. An iterative optimization mechanism is introduced to improve control adaptability, simplify the model structure, and enhance deployment feasibility. This avoids the problem of high dependence on high-precision sensor data and the difficulty in timely on-site parameter acquisition in three-dimensional unsteady models with gas flow as the core parameter adjustment mechanism. This method achieves precise control of molten iron temperature in the steel industry and effectively solves the aforementioned problems in the background technology.

[0004] The technical solution of this invention is: a method for predicting and controlling the temperature of molten iron in a blast furnace, comprising the following steps: (1) Collect, preprocess, normalize and feature engineering the relevant parameters of blast furnace hot metal temperature, use Pearson correlation coefficient and decision tree to screen key parameters, construct KAN-LSTM-Attention hot metal temperature prediction model, and realize the early judgment of temperature change trend. (2) Collect knowledge of the control rules related to the temperature of molten iron in blast furnace, and construct a rule base containing fuzzy judgment rules and operation rules by combining production data and expert experience; (3) The rule base is dynamically optimized by using fuzzy control method, a decision model for molten iron temperature is constructed, and fuzzy reasoning and defuzzification are performed based on the prediction difference and related parameters to select the optimal control strategy to achieve closed-loop control of molten iron temperature.

[0005] In step (1), parameters related to blast furnace molten iron temperature are selected by combining relevant literature on molten iron temperature prediction.

[0006] In step (1), the data is preprocessed, including handling duplicate values, missing values, and outliers. Based on the original time-series data, modeling is performed after data cleaning and feature filtering, including the following steps: First, use the duplicated function in the pandas library to process duplicate values ​​in the data. Based on the default rules or specified columns, identify duplicate rows in the dataset and return a sequence of boolean values. Use this sequence to filter out duplicate values ​​and process them accordingly. Secondly, the KNN algorithm is used to fill in missing values. If most of the k nearest neighbors of a sample x in the sample set belong to a certain category yLabel, then the sample x also belongs to the category yLabel and has the characteristics of samples of this category. Finally, outlier detection was performed by drawing a violin plot to see the initial distribution of outliers. Then, Lonely Forest was used to detect outliers and replace them with the mean. Finally, Gaussian blurring was used to perform secondary detection and processing on the data.

[0007] In step (1), the data is normalized using the extreme value method.

[0008] In step (1), key parameters are screened, and the original data is transformed into data that better represents the potential problem characteristics of the prediction model. This is achieved by selecting the most relevant features, extracting features, and constructing features. The construction of features is achieved by a dimensionality reduction algorithm. Finally, the parameters selected by the Pearson correlation coefficient and the decision tree algorithm are combined for comprehensive analysis, and the same parameters from the two methods are taken as the model input parameters.

[0009] In step (1), a KAN-LSTM-Attention molten iron temperature prediction model is constructed, which integrates the three models. The original data is first extracted by the KAN layer, then the long-term dependencies are mined by the LSTM layer, then the Attention layer focuses on key time steps, and finally the context vector is input into the fully connected layer to obtain the predicted value of molten iron temperature, as follows: First, KAN, as the first layer of the model, receives the raw molten iron temperature data. It uses kernel adaptive filtering to perform preliminary feature extraction and filtering on the data, removes noise and captures nonlinear relationships, and outputs high-quality features for subsequent model use. Secondly, LSTM is a classic recurrent neural network that excels at handling long-term dependencies in time series. It receives features extracted by KAN, learns in depth the time dynamics of molten iron temperature changes, and outputs the hidden state containing contextual information at each time step. Finally, the Attention mechanism dynamically weights the hidden states of the LSTM output, calculates the attention weights at each time step, and sums them to generate a context vector containing global understanding.

[0010] In step (2), the parameters selected from theoretical combustion temperature, furnace top temperature, heat load, water inlet temperature, oxygen enrichment flow rate, lower average pressure, total furnace pressure difference, oxygen enrichment pressure and valve seat temperature are used as inputs, and molten iron temperature is used as output. The first 70% of the data is used as the training set and the last 30% is used as the test set to construct the molten iron temperature prediction model.

[0011] In step (3), the control method adopts fuzzy control, which uses fuzzy set theory to convert the precise input quantity into a fuzzy quantity, then performs inference according to the pre-defined fuzzy rules, and finally converts the fuzzy output quantity obtained by inference into a precise quantity, thereby controlling the controlled object.

[0012] In step (3), the control strategy is as follows: First, the difference between the predicted result and the target molten iron temperature of 1500℃ is used as one of the input variables for fuzzy control. At the same time, the theoretical combustion temperature, oxygen enrichment flow rate, lower average pressure difference, whole furnace pressure difference, oxygen enrichment pressure, valve seat temperature, furnace top pressure, heat load, and parameters closely related to the molten iron temperature are selected as other input variables. Secondly, these input variables are fuzzified. Based on the actual range of change of each variable and its influence on the temperature of molten iron, the corresponding fuzzy subsets are determined, and a suitable membership function is selected for each fuzzy subset. Then, fuzzy rules are formulated by combining expert experience, actual production data and knowledge of molten iron temperature control. Furthermore, in the fuzzy inference stage, the Mamdani inference method is adopted, and the output variable represents the adjustment strategy for the molten iron temperature. The rule matching, rule activation and rule aggregation steps are performed in sequence. Finally, the fuzzy output is converted into a precise control quantity by the defuzzification method, namely the centroid method. Based on the feedback results, the parameters of the fuzzy control system, including the membership function and fuzzy rules, are corrected and optimized in real time.

[0013] The beneficial effects of this invention are: by constructing a self-learning control system based on a predictive model and combined with fuzzy control theory, temperature closed-loop control is achieved through fuzzification processing, rule formulation and dynamic reasoning, and an iterative optimization mechanism is introduced to improve control adaptability, simplify the model structure, improve deployment feasibility, and avoid the problem of strong dependence on high-precision sensor data and difficulty in timely on-site parameter acquisition in three-dimensional unsteady models with gas flow as the core parameter adjustment mechanism, thus achieving precise control of molten iron temperature in the steel industry. Attached Figure Description

[0014] Figure 1 This is a violin comparison chart of some parameters before and after data processing in this invention; Figure 2This is a comparison image of the standard wind speed of this invention after Gaussian blurring; Figure 3 This is a flowchart of the prediction model fusion process of the present invention; Figure 4 This is a graph showing the prediction results of the prediction model of this invention; Figure 5 This is a block diagram of the fuzzy controller of the present invention; Figure 6 This is a schematic diagram of the fuzzy control strategy of the present invention; Figure 7 This is a graph showing the temperature variation of molten iron under different control methods of the present invention. Detailed Implementation

[0015] To make the purpose, technical solutions, and advantages of the invention's embodiments clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only a small part of the embodiments of the present invention, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the protection scope of the present invention.

[0016] A method for predicting and controlling the temperature of molten iron in a blast furnace, comprising the following steps: (1) Collect, preprocess, normalize and feature engineering the relevant parameters of blast furnace hot metal temperature, use Pearson correlation coefficient and decision tree to screen key parameters, construct KAN-LSTM-Attention hot metal temperature prediction model, and realize the early judgment of temperature change trend. (2) Collect knowledge of the control rules related to the temperature of molten iron in blast furnace, and construct a rule base containing fuzzy judgment rules and operation rules by combining production data and expert experience; (3) The rule base is dynamically optimized by using fuzzy control method, a decision model for molten iron temperature is constructed, and fuzzy reasoning and defuzzification are performed based on the prediction difference and related parameters to select the optimal control strategy to achieve closed-loop control of molten iron temperature.

[0017] In step (1), parameters related to blast furnace molten iron temperature are selected by combining relevant literature on molten iron temperature prediction.

[0018] In step (1), the data is preprocessed, including handling duplicate values, missing values, and outliers. Based on the original time-series data, modeling is performed after data cleaning and feature filtering, including the following steps: First, use the duplicated function in the pandas library to process duplicate values ​​in the data. Based on the default rules or specified columns, identify duplicate rows in the dataset and return a sequence of boolean values. Use this sequence to filter out duplicate values ​​and process them accordingly. Secondly, the KNN algorithm is used to fill in missing values. If most of the k nearest neighbors of a sample x in the sample set belong to a certain category yLabel, then the sample x also belongs to the category yLabel and has the characteristics of samples of this category. Finally, outlier detection was performed by drawing a violin plot to see the initial distribution of outliers. Then, Lonely Forest was used to detect outliers and replace them with the mean. Finally, Gaussian blurring was used to perform secondary detection and processing on the data.

[0019] In step (1), the data is normalized using the extreme value method.

[0020] In step (1), key parameters are screened, and the original data is transformed into data that better represents the potential problem characteristics of the prediction model. This is achieved by selecting the most relevant features, extracting features, and constructing features. The construction of features is achieved by a dimensionality reduction algorithm. Finally, the parameters selected by the Pearson correlation coefficient and the decision tree algorithm are combined for comprehensive analysis, and the same parameters from the two methods are taken as the model input parameters.

[0021] In step (1), a KAN-LSTM-Attention molten iron temperature prediction model is constructed, which integrates the three models. The original data is first extracted by the KAN layer, then the long-term dependencies are mined by the LSTM layer, then the Attention layer focuses on key time steps, and finally the context vector is input into the fully connected layer to obtain the predicted value of molten iron temperature, as follows: First, KAN, as the first layer of the model, receives the raw molten iron temperature data. It uses kernel adaptive filtering to perform preliminary feature extraction and filtering on the data, removes noise and captures nonlinear relationships, and outputs high-quality features for subsequent model use. Secondly, LSTM is a classic recurrent neural network that excels at handling long-term dependencies in time series. It receives features extracted by KAN, learns in depth the time dynamics of molten iron temperature changes, and outputs the hidden state containing contextual information at each time step. Finally, the Attention mechanism dynamically weights the hidden states of the LSTM output, calculates the attention weights at each time step, and sums them to generate a context vector containing global understanding.

[0022] In step (2), the parameters selected from theoretical combustion temperature, furnace top temperature, heat load, water inlet temperature, oxygen enrichment flow rate, lower average pressure, total furnace pressure difference, oxygen enrichment pressure and valve seat temperature are used as inputs, and molten iron temperature is used as output. The first 70% of the data is used as the training set and the last 30% is used as the test set to construct the molten iron temperature prediction model.

[0023] In step (3), the control method adopts fuzzy control, which uses fuzzy set theory to convert the precise input quantity into a fuzzy quantity, then performs inference according to the pre-defined fuzzy rules, and finally converts the fuzzy output quantity obtained by inference into a precise quantity, thereby controlling the controlled object.

[0024] In step (3), the control strategy is as follows: First, the difference between the predicted result and the target molten iron temperature of 1500℃ is used as one of the input variables for fuzzy control. At the same time, the theoretical combustion temperature, oxygen enrichment flow rate, lower average pressure difference, whole furnace pressure difference, oxygen enrichment pressure, valve seat temperature, furnace top pressure, heat load, and parameters closely related to the molten iron temperature are selected as other input variables. Secondly, these input variables are fuzzified. Based on the actual range of change of each variable and its influence on the temperature of molten iron, the corresponding fuzzy subsets are determined, and a suitable membership function is selected for each fuzzy subset. Then, fuzzy rules are formulated by combining expert experience, actual production data and knowledge of molten iron temperature control. Furthermore, in the fuzzy inference stage, the Mamdani inference method is adopted, and the output variable represents the adjustment strategy for the molten iron temperature. The rule matching, rule activation and rule aggregation steps are performed in sequence. Finally, the fuzzy output is converted into a precise control quantity by the defuzzification method, namely the centroid method. Based on the feedback results, the parameters of the fuzzy control system, including the membership function and fuzzy rules, are corrected and optimized in real time.

[0025] In practical applications, firstly, by collecting, preprocessing, normalizing, and performing feature engineering on relevant parameters of blast furnace molten iron temperature, and using Pearson correlation coefficient and decision tree to screen key parameters, a KAN-LSTM-Attention molten iron temperature prediction model is constructed to achieve early prediction of temperature change trends. On the one hand, the overall structure is better able to cope with sudden changes in blast furnace conditions; on the other hand, it avoids the problems of weak adaptability to nonlinear disturbances in three-dimensional unsteady models with gas flow as the core parameter adjustment mechanism, as well as the limited modeling and generalization capabilities of ESN models. Secondly, by collecting knowledge of blast furnace molten iron temperature control rules, and combining production data and expert experience, a rule base containing fuzzy judgment rules and operation rules is constructed. Finally, the rule base is dynamically optimized through fuzzy control methods to construct a molten iron temperature decision model. Based on the prediction difference and relevant parameters, fuzzy inference and defuzzification operations are performed to select the optimal control strategy to achieve closed-loop control of molten iron temperature.

[0026] Preferably, the data analysis includes missing value processing, outlier processing, duplicate value processing, and normalization processing to clean the data.

[0027] Preferably, the molten iron prediction model includes model building and result analysis. The model building utilizes the KAN-LSTM-Attention model to comprehensively predict the molten iron temperature, and uses root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), symmetric mean absolute percentage error (SMAPE), and fitting accuracy (R2) as evaluation indicators to comprehensively assess the model's predictive performance.

[0028] Preferably, the establishment of the fuzzy control includes a fuzzification interface, a knowledge base, a fuzzy inference engine, and a defuzzification interface. The fuzzification interface converts precise input data from the actual system into elements of a fuzzy set, that is, converts precise quantities into fuzzy linguistic variables. The knowledge base contains fuzzy control rules and related fuzzy data. Fuzzy control rules are summarized based on expert experience or actual operational data, describing the relationship between input and output variables. The fuzzy inference engine performs fuzzy logic inference based on the input fuzzy quantities and the fuzzy rules in the knowledge base to derive fuzzy output quantities. The defuzzification interface converts the fuzzy output quantities obtained from fuzzy inference into precise control quantities so that they can be applied to the actual controlled object.

[0029] Preferably, the fuzzy inference algorithm includes rule matching, rule activation, and rule aggregation. Rule matching involves meticulously comparing the fuzzified input variables with the rule antecedents in the rule base, calculating the activation degree of each rule using a specific calculation method (based on membership functions, taking the minimum value through an AND operation). Rule activation involves truncating or scaling the rule conclusions based on the activation degrees obtained from rule matching, and modifying the membership function of the fuzzy subset of the output variable. Rule aggregation in fuzzy control inference uses the maximum value method to merge the output activation rules, obtaining a final output fuzzy set, which is used for defuzzification to obtain precise control values. Example

[0030] This embodiment proposes a method for predicting, analyzing, and controlling blast furnace molten iron temperature based on KAN-LSTM-Attention, such as... Figures 1-7As shown, firstly, by collecting, preprocessing, normalizing, and performing feature engineering on relevant parameters of blast furnace molten iron temperature, and using Pearson correlation coefficient and decision tree to screen key parameters, a KAN-LSTM-Attention molten iron temperature prediction model is constructed to achieve early judgment of temperature change trends. This avoids the problems of complex image processing and high preprocessing costs associated with convolutional neural networks (CNNs). Furthermore, the image generation + CNN prediction chain is long, resulting in insufficient overall response speed and difficulty in adapting to second-level prediction requirements. This model does not use an image processing process and directly uses structured data (blast temperature, blast volume, pulverized coal injection, furnace pressure, etc.) as input. Secondly, by collecting knowledge of blast furnace molten iron temperature control rules, and combining production data and expert experience, a rule base containing fuzzy judgment rules and operation rules is constructed. Finally, the rule base is dynamically optimized through fuzzy control methods to construct a molten iron temperature decision model. Based on the prediction difference and relevant parameters, fuzzy inference and defuzzification operations are performed to select the optimal control strategy to achieve closed-loop control of molten iron temperature.

[0031] First, data collection was conducted. As shown in Table 1, based on relevant literature on molten iron temperature prediction, parameters related to furnace pressure, such as furnace body static pressure and overall furnace pressure difference, were selected. These directly reflect the resistance distribution of the gas flow inside the furnace, directly affecting gas utilization and heat transfer efficiency. Temperature-related parameters, such as top temperature and hot blast temperature, were selected to determine the heat input in the tuyeres area, directly affecting the sensible heat of the molten iron and the activity of the hearth. Parameters related to molten iron composition, such as Si and Ti, were selected. High Si content indicates sufficient heat in the hearth, while Ti reduction requires high temperatures. Combining these two factors allows for more accurate prediction of molten iron temperature. Cooling system parameters, such as inlet water temperature and furnace top cooling water pressure, were selected to reflect the thermal load status of the blast furnace lining.

[0032] Table 1 Initial parameters for predicting the molten iron temperature of blast furnace No. 6 in a steel plant

[0033] Due to limitations in equipment operating conditions and interference from complex industrial environments, the raw data contains issues such as outlier redundancy, non-standard formatting, and logical conflicts. Preliminary processing is required to reduce invalid calculations and improve interpretability. The processing methods are as follows: First, duplicate value handling: Use the `duplicated` function from the pandas library to identify duplicate rows according to default rules or specified columns, returning a Boolean sequence for filtering. Second, missing value handling: Sensor malfunctions in blast furnace ironmaking can cause data loss. Missing points are filled using the nearest neighbor method or linear interpolation. For large amounts of missing data (above 60%), delete the data for that period and use the KNN algorithm to fill in the missing data. Third, outlier handling: First, draw a violin plot to observe the initial distribution, then use an isolation forest to detect outliers and replace them with the mean. This method constructs an isolation tree by randomly selecting features and splitting values ​​to divide the data space, calculates the average path length of data points to obtain an anomaly score, and sets a threshold to determine outliers. After processing, compare the data before and after using a violin plot. Figure 1 As shown, S101 represents the processed data, and S102 represents the unprocessed data. It's clear that most outliers have been processed, but a small number remain. Finally, a second detection process is performed using Gaussian blurring. This technique, based on a Gaussian function, maps the original data to a new feature space, reducing the impact of noise. Taking standard wind speed as an example... Figure 2 The results of outlier detection in standard wind speed data are shown. The blue image represents the detection results obtained using this method, and the orange image represents the detection results after secondary processing using this method. Figure 2 It can be seen that after Gaussian blurring, outliers are significantly reduced and data quality is significantly improved.

[0034] KAN, as the first layer of the model, receives the raw molten iron temperature data. It uses kernel adaptive filtering to perform initial feature extraction and filtering, removing noise and capturing nonlinear relationships, outputting high-quality features for subsequent model use. LSTM, a classic recurrent neural network, excels at handling long-term dependencies in time series. It receives the features extracted by KAN, deeply learns the temporal dynamics of molten iron temperature changes, and outputs hidden states containing contextual information at each time step. The Attention mechanism dynamically weights the hidden states output by LSTM, calculates the attention weights for each time step, and sums the weights to generate a context vector containing global understanding. The overall process is as follows: the raw data first undergoes feature extraction by the KAN layer, then the LSTM layer mines long-term dependencies, followed by the Attention layer focusing on key time steps, and finally, the context vector is input into the fully connected layer to obtain the predicted molten iron temperature. This fusion approach has significant advantages: KAN provides high-quality features to reduce the impact of noise, LSTM enhances temporal modeling capabilities, and Attention improves sensitivity to important information, enabling the KAN-LSTM-Attention model to accurately predict molten iron temperature, making it suitable for complex industrial scenarios. The KAN-LSTM-Attention model fusion flowchart is shown below. Figure 3As shown.

[0035] The prediction results of the KAN-LSTM-Attention model are shown in the figure below. Figure 4 As shown in the figure, S401 represents the actual value and S402 represents the predicted value. It can be seen that the model exhibits good performance in predicting molten iron temperature. The figure also shows that the model's predicted values ​​and the actual measured values ​​maintained a high degree of consistency throughout the process, with a significantly small prediction error. The mean absolute error (MAE) was controlled within ±5℃, indicating that the model has excellent modeling capabilities for steady-state processes.

[0036] To verify the performance of the KAN-LSTM-Attention model, this paper compares it with the traditional LSTM model and the LSTM-Attention model using the same training and test sets. The specific results are shown in Table 2.

[0037] Table 2 Comparison of prediction results from different models

[0038] In comparisons across multiple evaluation metrics, the KAN-LSTM-Attention model exhibits significant overall advantages. While its RMSE and MAE scores, measuring prediction error, are slightly higher than the LSTM and LSTM-Attention models (0.104 and 0.075 respectively), they still effectively reduce prediction error compared to the actual values ​​(RMSE=0.134, MAE=0.093).

[0039] Fuzzy control is a control method based on linguistic description and fuzzy logic. It achieves control of the controlled object through processes such as fuzzification of the input quantity (S501), fuzzy inference based on a knowledge base (S502), and defuzzification (S503). Fuzzy control utilizes fuzzy set theory to transform precise input quantities into fuzzy quantities, then performs inference according to pre-defined fuzzy rules, and finally converts the resulting fuzzy output quantity back into a precise quantity to control the controlled object. The block diagram of a fuzzy controller is shown below. Figure 5 As shown in the diagram. A schematic diagram of the fuzzy control strategy is shown below. Figure 6 As shown, the temperature and correlation coefficient are first obtained, the target temperature is set and the deviation is calculated. After input variable fuzzification, rule matching activation, rule aggregation, and defuzzification calculation, the output adjustment amount controls the water temperature. Finally, the process ends based on whether the deviation meets the requirements. This solves the problem that typical physical modeling iterative methods for calculating molten iron temperature only give a temperature prediction value and do not output suggested adjustment actions.

[0040] To verify the control effect of the fuzzy controller on molten iron temperature, a molten iron temperature regulation experiment was conducted based on the prediction results of the KAN-LSTM-Attention model. During the experiment, the molten iron temperature was precisely controlled by dynamically adjusting key parameters such as oxygen enrichment flow rate and inlet temperature, with the target molten iron temperature set at 1500℃. The control effects of the fuzzy control system based on KAN-LSTM-Attention and the original PID control system were compared, and some results are shown below. Figure 7 As shown.

[0041] Experimental results show that when using the KAN-LSTM-Attention fuzzy control scheme, the fluctuation range of molten iron temperature near the target value is significantly reduced, controlled within the range of 1495℃ to 1505℃, while the temperature fluctuation range of the traditional PID control system is larger (1480℃-1520℃). This indicates that the deep learning-based fuzzy control strategy can more accurately maintain the molten iron temperature close to the set value. Figure 7 The error distribution shows that under the new control scheme, the temperature error of about 92% of the sampling points is controlled within ±5℃, which improves the temperature control accuracy by about 35% compared with the PID control system.

[0042] Compared with the prior art, the method of the present invention has the following advantages and beneficial effects: 1. In the method of this invention, KAN is based on the Kolmogorov-Arnold theorem to achieve dimensionality reduction of high-dimensional data and approximation of complex functions, LSTM solves the problem of long-term dependence of time series, and Attention focuses on key features. The three work together to improve the prediction robustness of the model under complex industrial conditions such as high temperature and high pressure. Compared with the traditional LSTM model, the prediction accuracy is improved by 4% (up to 84%), which reflects the technical breakthrough of cross-algorithm fusion.

[0043] 2. The performance of the model was verified through comparative experiments. The results show that the model achieves an R² score of 84%, which is 2% higher than LSTM-Attention, and the prediction error is controlled within ±5℃. Simultaneously, in the fuzzy control experiment, the molten iron temperature was precisely controlled within the range of 1495-1505℃, reducing the fluctuation range by 15℃ compared to traditional PID control, thus verifying the model's effectiveness in both theory and engineering practice.

Claims

1. A blast furnace hot metal temperature prediction and regulation method, characterized in that Comprise the following steps: (1) Collect, preprocess, normalize and feature engineering treatment on the related parameters of the blast furnace molten iron temperature, use the Pearson correlation coefficient and decision tree to screen the key parameters, build a KAN-LSTM-Attention molten iron temperature prediction model, realize the early judgment of temperature change trend; (2) Collect the related regulation knowledge of blast furnace molten iron temperature, combine production data and expert experience to build a rule base containing fuzzy judgment rules and operation rules; (3) Through fuzzy control method to realize the dynamic optimization of rule base, build molten iron temperature decision model, based on the prediction difference and related parameters, carry out fuzzy reasoning and defuzzification operation, select the optimal control strategy to realize the closed-loop regulation of molten iron temperature.

2. The blast furnace molten iron temperature prediction and regulation method according to claim 1, characterized in that: In step (1), combined with the related literature of molten iron temperature prediction, the parameters related to the blast furnace molten iron temperature are selected.

3. The blast furnace molten iron temperature prediction and regulation method according to claim 1, characterized in that: In step (1), the data is preprocessed, including duplicate value processing, missing value processing and outlier processing, based on the original time series data, after data cleaning and feature selection, the modeling includes the following steps: Firstly, use the duplicated function in pandas library to process the duplicate values in the data, identify the duplicate rows in the data set according to the default rule or specified column, return a boolean value sequence, according to which the duplicate values are filtered and processed accordingly; Secondly, use KNN algorithm to fill in the missing values, if the majority of the k nearest neighbors of a sample x belong to a certain class yLabel, then the sample x also belongs to the class yLabel and has the characteristics of this class of samples; Finally, draw a violin plot to detect the preliminary distribution of outliers, then use the isolation forest to detect outliers and replace them with the mean, and finally use the Gaussian fuzzy technique for secondary detection and processing.

4. The blast furnace molten iron temperature prediction and regulation method according to claim 1, characterized in that: In step (1), the data is normalized using the maximum and minimum method.

5. The method of claim 3, wherein the method further comprises: In step (1), the key parameters are selected, the original data is converted into data that can better represent the potential problem characteristics of the prediction model, which is realized by selecting the most relevant features, extracting features and constructing features, among which the constructed features are realized by dimensionality reduction algorithm; Finally, the parameters selected by Pearson correlation coefficient and decision tree algorithm are analyzed comprehensively, and the same parameters of the two methods are taken as the input parameters of the model.

6. The blast furnace molten iron temperature prediction and regulation method according to claim 1, characterized in that: In step (1), the KAN-LSTM-Attention molten iron temperature prediction model is built, the three models are fused, the original data is first extracted by KAN layer, then the long-term dependence is mined by LSTM layer, then the attention layer focuses on the key time steps, finally the context vector is input into the fully connected layer to get the molten iron temperature prediction value, as follows: Firstly, KAN as the first layer of the model, receives the original molten iron temperature data, uses kernel adaptive filtering to preliminarily extract features and filter data, removes noise and captures nonlinear relationship, outputs high-quality features for subsequent models to use; Secondly, LSTM is a classic recurrent neural network that is good at processing long-term dependencies of time series. It receives the features extracted by KAN, deeply learns the time dynamic characteristics of molten iron temperature change, and outputs the hidden state containing context information at each time step. Finally, the attention mechanism dynamically weights the hidden state output by LSTM, calculates the attention weight at each time step, and generates a context vector containing global understanding by weighted summation.

7. The blast furnace molten iron temperature prediction and regulation method according to claim 1, characterized in that: In step (2), the parameters screened by the theoretical combustion temperature, the top temperature of the furnace, the heat load, the inlet water temperature, the oxygen-enriched flow, the lower average pressure, the full furnace pressure difference, the oxygen-enriched pressure, and the valve seat temperature are used as inputs, and the molten iron temperature is used as output. The first 70% of the data is used as the training set, and the last 30% is used as the test set to build a molten iron temperature prediction model.

8. The blast furnace molten iron temperature prediction and regulation method according to claim 1, characterized in that: In step (3), the control method uses fuzzy control. The input precise quantity is converted into a fuzzy quantity using fuzzy set theory, then the fuzzy rules are inferred according to the pre-established fuzzy rules, and finally the fuzzy output quantity obtained by inference is converted into a precise quantity to control the controlled object.

9. The blast furnace molten iron temperature prediction and regulation method according to claim 1, characterized in that: In step (3), the control strategy is as follows: First, the difference between the prediction result and the target molten iron temperature of 1500℃ is selected as one of the input variables of fuzzy control, and the theoretical combustion temperature, oxygen-enriched flow, lower average pressure difference, full furnace pressure difference, oxygen-enriched pressure, valve seat temperature, furnace top pressure, heat load, and inlet water temperature are selected as other input variables closely related to molten iron temperature. Secondly, these input variables are fuzzified. According to the actual range of each variable and its impact on molten iron temperature, the corresponding fuzzy subsets are determined, and appropriate membership functions are selected for each fuzzy subset. Then, fuzzy rules are developed based on expert experience, actual production data, and knowledge of molten iron temperature control. Again, in the fuzzy reasoning link, the Mamdani reasoning method is used, and the output variable represents the adjustment strategy for molten iron temperature. The rule matching, rule activation, and rule aggregation steps are performed in turn. Finally, the fuzzy output is converted into a precise control quantity by the defuzzification method, i.e., the center of gravity method. According to the feedback result, the parameters of the fuzzy control system, including the membership function and the fuzzy rule, are modified and optimized in real time.

Images