Converter blowing carbon oxygen balance inverse solution and process parameter optimization method

Abstract

The present application relates to converter steelmaking technical field, especially a kind of converter blowing carbon oxygen balance inverse solution and process parameter optimization method, construct carbon oxygen balance inverse solution model, inverse solution is process parameter initial value for target endpoint parameter;Establish multi-constraint dynamic parameter optimization model in stages, respectively establish the constraint condition and objective function corresponding to each stage, generate the optimal dynamic process parameter curve of each stage with time variation;Establish real-time feedback closed-loop control mechanism, dynamically trigger inverse solution recalculation and parameter optimization iteration.The present application has the advantages that: mechanism-data hybrid driving strategy is used, carbon oxygen reaction thermodynamics and kinetics are used as constraint boundary, PSO algorithm is used to optimize the initial weight and bias of BP neural network, effectively solve the problems of multiple solutions, local optimum and slow convergence;Ensure that the inverse solution result meets the nature of metallurgical reaction, avoid the reaction too fast or too slow caused by inverse solution parameter, realize more accurate and stable inverse solution from process parameter to target state.

Description

Technical Field

[0001] This invention relates to the field of converter steelmaking technology, and in particular to a method for reverse carbon-oxygen balance solution and process parameter optimization in converter blowing. Background Technology

[0002] Converter blowing is a key process in modern steel production, with its core being the thermodynamic and kinetic balance of the carbon-oxygen reaction within the molten pool. This process directly determines the precision of controlling carbon and temperature at the converter's final stage, while also affecting steel feedstock consumption, slag-forming agent consumption, smelting cycle, CO2 emission levels, and the final quality of the steel. It is the core link restricting efficient, low-carbon, and precise converter production.

[0003] Currently, traditional converter blowing control methods still face numerous technical bottlenecks, making it difficult to meet the demands of modern steel production for high precision, low cost, low energy consumption, and low carbon emissions. Specifically: (1) The carbon-oxygen balance inverse solution method has defects: the traditional carbon-oxygen balance inverse solution method mainly relies on pure mechanism model or pure data-driven model. Among them, the pure mechanism model cannot adapt to the complex working condition fluctuations in actual production; the pure data-driven model has a large deviation in the inverse solution results when the process parameters fluctuate, and may even lead to control failure.

[0004] (2) Lack of adaptive adjustment capability of process parameters: In traditional methods, process parameters are set using fixed curves, which cannot be adaptively adjusted according to the differences in carbon-oxygen reaction characteristics in each blowing stage. This can easily lead to production problems such as overblowing, re-drying, and splashing, affecting the stability of smelting.

[0005] (3) Single optimization objective: Existing technologies only focus on the single objective of the endpoint carbon temperature double hit rate, which cannot take into account the multi-objective synergistic optimization of steel stability, smelting efficiency, production cost and low carbon emissions.

[0006] (4) Insufficient robustness under complex working conditions: When faced with complex working conditions such as fluctuations in molten iron composition, changes in furnace type, increase in furnace age, and equipment wear, the control accuracy of the existing technology drops significantly and the robustness is insufficient.

[0007] In existing technologies, most techniques involving carbon-oxygen balance control in converter blowing are based on endpoint prediction using machine learning or parameter calculation using traditional analytical methods. For example, CN108251591B discloses a top-bottom combined blowing converter production process control method using an LSTM system, which only achieves forward prediction of the endpoint state and does not involve intelligent inverse solution of carbon-oxygen balance. CN106119458B discloses a converter steelmaking process cost control method and system based on BP neural network, which relies on genetic algorithm to complete offline optimization of smelting static process parameters, but does not combine inverse solution technology to achieve accurate solution of carbon-oxygen balance, nor does it consider the dynamic optimization needs of the entire smelting process. In existing academic literature, some studies have proposed applying intelligent algorithms to converter carbon-oxygen reaction modeling, but these studies only remain at the level of theoretical simulation research and have not formed a complete set of process parameter optimization schemes that can be implemented. They have also failed to solve key engineering problems such as multi-solution screening and real-time adaptation of dynamic constraints in intelligent inverse solution calculation.

[0008] In summary, existing technologies have not achieved a deep integration of carbon-oxygen balance inverse solution, staged multi-constraint dynamic optimization, and real-time closed-loop control, and cannot fundamentally solve the above-mentioned technical defects, making it difficult to adapt to the high-quality production development needs of modern converters that are intelligent, efficient, and low-carbon. Summary of the Invention

[0009] The purpose of this invention is to provide a method for inverse carbon-oxygen balance solution and process parameter optimization in converter blowing. By constructing a mechanism-data hybrid driven inverse carbon-oxygen balance solution model, a precise and stable inverse solution from the target endpoint parameters to the process parameters is achieved. Combining a staged multi-constraint dynamic optimization model with the NSGA-III algorithm, an optimal dynamic process parameter curve adapted to the characteristics of the entire blowing cycle is generated. A real-time closed-loop feedback control mechanism is established to dynamically correct process deviations. Ultimately, this replaces manual experience-based operation, achieving precise, stable, and automatic control of the entire converter blowing process, improving the endpoint carbon-temperature double hit rate, and reducing steel material consumption, slagging agent consumption, and CO2 emissions.

[0010] To achieve the above objectives, the present invention provides the following technical solution: A method for reverse carbon-oxygen balance solution and process parameter optimization in converter blowing includes: S1. Collect data from the entire converter blowing process and perform data preprocessing; S2. Construct a carbon-oxygen balance inverse solution model. The carbon-oxygen balance inverse solution model includes a mechanism constraint framework and a BP-PSO hybrid inverse solution model, and inversely solves the target endpoint parameters into the initial values ​​of process parameters. S3. Establish a phased multi-constraint dynamic parameter optimization model. Based on the carbon-oxygen reaction characteristics of each stage in the initial, middle and late stages of converter blowing, establish the corresponding constraints and objective functions for each stage. Based on the initial values ​​of process parameters obtained from the inverse solution, use the NSGA-III multi-objective optimization algorithm to perform phased iterative optimization and generate the optimal dynamic process parameter curves of each stage as time changes. S4. Establish a real-time feedback closed-loop control mechanism to dynamically trigger inverse solution recalculation and parameter optimization iteration based on the deviation between online detection data and the target. S5. The optimized process parameters are sent to the basic automation system in the form of digital instructions to control the automatic operation of oxygen lances, oxygen supply valves, bottom blowing valves and slagging agent feeding devices.

[0011] In S1, the data for the entire converter blowing process includes: S11, Initial conditions for molten iron: Initial carbon content Initial silicon content Initial manganese content Initial sulfur content Initial phosphorus content Initial temperature molten iron weight ; S12, Real-time status of the molten pool: Real-time carbon content Real-time oxygen activity Real-time temperature , furnace gas composition and Volume fraction; S13, Process Parameters: Oxygen supply flow rate Oxygen lance position Bottom blowing volume Timing and dosage of slag-forming agent addition ; S14, Equipment Status: Furnace volume ratio , Oxygen lance aperture Furnace age .

[0012] In S2, the carbon-oxygen balance inverse solution model includes a mechanistic constraint framework and a BP-PSO hybrid inverse solution model. The mechanistic constraint framework is used to limit the input and output boundary conditions of the inverse solution model, while the BP-PSO hybrid inverse solution model performs inverse solution within the boundary conditions. S21. The mechanistic constraint framework includes thermodynamic constraints and kinetic constraints on the carbon-oxygen reaction: (1) Thermodynamic constraints on the carbon-oxygen reaction are used to limit the output boundary of the inverse solution model, and the expression is: , ; in: The carbon content of the molten pool is expressed as % (%). Oxygen content in the molten pool, in % % It is carbon monoxide gas; The temperature of the molten pool is expressed in Kelvin (K). Standard Gibbs free energy, expressed in J / mol; 22100 and 38.3 are empirical thermodynamic coefficients for the carbon-oxygen reaction.

[0013] (2) Carbon-oxygen reaction kinetic constraints, used to limit the rate range of the carbon-oxygen reaction, are expressed as: ; in: This represents the carbon oxidation rate, expressed as % / min. The reaction rate constant is positively correlated with temperature, and its value ranges from 0.02 to 0.08 min. -1 ; This refers to the partial pressure of CO, measured in atm. This is the equilibrium constant for the carbon-oxygen reaction, which varies with smelting temperature. S22, BP-PSO hybrid inverse solution model: using a BP neural network as the solver and employing the PSO algorithm to optimize the initial weights and thresholds of the BP neural network; A backpropagation (BP) neural network consists of an input layer, hidden layers, and an output layer. The input layer contains the target endpoint parameters; Hidden layers are used to perform nonlinear transformations on the input features; The output layer contains the initial values ​​of the process parameters; The activation function used is the Sigmoid function, and the output formula is as follows: ; in: This is the output value of the kth node in the output layer, i.e., the initial value of the process parameters; This is the input value of the i-th node in the input layer, i.e., the target endpoint parameter; The weights from the i-th node in the hidden layer to the k-th node in the output layer; This is the bias of the k-th node in the output layer; This is the Sigmoid activation function.

[0014] The PSO algorithm is used to optimize the initial weights and threshold of a BP neural network. With bias The iterative formula is as follows: ; in: Let be the velocity of the i-th particle in the j-th dimension during the (t+1)-th iteration; Let be the velocity of the i-th particle in the j-th dimension during the t-th iteration; , For learning factors; , A random number within the interval [0,1]; Let be the individual historical optimal position of the i-th particle in the j-th dimension during the t-th iteration; Let be the globally optimal position of the group in the j-th dimension during the t-th iteration; For the first The current position of a particle in the j-th dimension during the t-th iteration; For inertia weights, a linear decreasing strategy is adopted, and the expression is: ; in: Let be the inertia weight at the t-th iteration; The initial inertia weight is set to 1.2. The final inertia weight is set to 0.8. This represents the maximum number of iterations, with a value of 100. This represents the current iteration number, with a value ranging from 1 to 100.

[0015] In S3, the objective function of the NSGA-III multi-objective optimization algorithm is: ; in: Optimize function values ​​for multiple objectives; ~ These are the weighting coefficients, which are 0.3, 0.2, 0.2, 0.15, and 0.15 respectively. The endpoint carbon temperature double hit rate, with a value ranging from 0% to 100%; The blowing cycle is expressed in minutes and ranges from 10 to 16. This refers to the consumption of steel materials, expressed in kg / t, with a value range of 1050–1100. The value represents the consumption of slag-forming agent, expressed in kg / t, and ranges from 20 to 60. This represents CO2 emissions, expressed in tons of steel (t / t), with a range of 1.5 to 2.9.

[0016] In S3: Initial stage constraints include: heating rate of 80-100℃ / min, silicon oxidation rate ≥90%, manganese oxidation rate ≥70%, and slag basicity of 1.2-1.5; The mid-term constraints include: carbon oxidation rate of 0.03–0.06% / min, slag foaming degree of 0.8–1.2, splashing probability ≤5%, and bottom blowing gas volume of 5–20 m³ / min; The constraints for the later stage include: carbon content deviation ±0.01%, temperature deviation ±5℃, final oxygen activity 300~600ppm, and overblowing rate ≤2%.

[0017] In S4, establishing a real-time feedback closed-loop control mechanism includes: Real-time monitoring: Sampling frequency is 1 time / 30 seconds to obtain real-time carbon content. Real-time oxygen activity Real-time temperature ; Deviation calculation: ; ; ; in: The value range is for carbon content deviation. ; Real-time carbon content, in % % The target carbon content is expressed as % (%). For temperature deviation, the value range is: ; This is the real-time temperature, in °C. The target temperature is expressed in °C. This is due to oxygen activity deviation; Real-time oxygen activity; The target oxygen activity.

[0018] Deviation judgment: If the deviation is within the threshold, keep the current process parameters unchanged; If the deviation exceeds the threshold, the feedback correction process will begin. Feedback and corrections: The deviation signal is fed back to the carbon-oxygen balance inverse solution model and the staged optimization model to recalculate and output the corrected process parameters.

[0019] In S4, dynamic triggering of inverse solution recalculation and parameter optimization iteration includes: Real-time monitoring of molten pool status parameters and calculation of deviations from target values; When any deviation exceeds the set threshold, the inverse solution recalculation signal is triggered; Using the current molten pool state parameters as new input conditions, re-execute the inverse carbon-oxygen balance model in S2 to update the initial values ​​of process parameters; Using the updated process parameters as the initial population, the phased multi-constraint dynamic parameter optimization model in S3 is re-executed, and the corrected optimal process parameter curve is output. The revised process parameters are then sent to the basic automation system for execution.

[0020] Compared with the prior art, the beneficial effects of the present invention are: 1. A mechanism-data hybrid driving strategy is adopted, using the thermodynamics and kinetics of the carbon-oxygen reaction as the constraint boundary. The PSO algorithm is used to optimize the initial weights and biases of the BP neural network, effectively solving the problems of multiple solutions, local optima, and slow convergence in traditional inverse kinematics methods; through thermodynamic constraints ( <0) Ensure that the reverse solution results conform to the essence of metallurgical reaction. By using kinetic constraints (carbon oxidation rate 0.02~0.08min⁻¹), the reverse solution parameters are avoided from causing the reaction to be too fast or too slow, thus achieving a more accurate and stable reverse solution from the target state to the process parameters. 2. The converter blowing process was divided into three stages: initial stage (0-4 min), middle stage (4-10 min), and later stage (10-16 min). Constraints were established for each stage: initial stage: heating rate 80-100℃ / min, silicon oxidation rate ≥90%, manganese oxidation rate ≥70%, slag basicity 1.2-1.5; middle stage: carbon oxidation rate 0.03-0.06% / min, slag foaming degree 0.8-1.2, splashing probability ≤5%; later stage: carbon content deviation ±0.01%, temperature deviation ±5℃, final oxygen activity 300-600ppm, overblowing rate ≤2%. Through staged constraints, the overblowing rate in Example 1 was only 1.2%, with no back-drying or splashing phenomena throughout the process, which was significantly better than Comparative Example 1 (overblowing rate 6.8%, with back-drying and splashing). 3. Using the NSGA-III multi-objective optimization algorithm, a five-objective optimization function was constructed, including the endpoint carbon-temperature double hit rate, blowing cycle, steel material consumption, slagging agent consumption, and CO2 emissions. The weight coefficients were 0.3, 0.2, 0.2, 0.15, and 0.15, respectively. Taking a 180t converter as an example, Example 1 achieved a carbon-temperature double hit rate of 96%, a blowing cycle of 13min, steel material consumption of 1052kg / t, slagging agent consumption of 52kg / t (a 10.3% reduction compared to the traditional method), and CO2 emissions of 2.65t / t. All five indicators were better than those in Comparative Examples 1 to 3, achieving multi-objective synergistic optimization. 4. A closed-loop control mechanism covering the entire process of "real-time monitoring → deviation calculation → deviation judgment → feedback correction" has been established, with a sampling frequency of once every 30 seconds, acquiring data in real time. , , When the deviation exceeds the threshold ( ±0.01%, When the temperature reaches ±5℃, the inverse kinematics recalculation and parameter optimization iteration are triggered. In Example 2, when the silicon content of the molten iron increases from 0.5% to 0.7%, the inverse kinematics model automatically adjusts the constraint threshold, and the closed-loop control corrects the abnormal carbon oxidation rate in real time. In Example 3, when the furnace volume ratio drops to 0.7 (compact furnace type), the confidence interval is tightened from ±5% to ±3%, and the closed-loop system corrects the temperature deviation (+6℃) in real time. This invention has a strong adaptability to complex working conditions such as fluctuations in molten iron composition and changes in furnace type. Comparative Example 3 (pure mechanistic model) shows inverse kinematics deviation when the molten iron composition fluctuates slightly, which fully demonstrates the robustness advantage of this invention. 5. The phased and dynamically optimized process parameters (dynamic curve of oxygen supply flow rate, dynamic curve of oxygen lance position, dynamic curve of bottom blowing gas volume, timing and amount of slagging agent addition) are sent to the basic automation system in the form of digital commands to control the automatic operation of oxygen lance, oxygen supply valve, bottom blowing valve and slagging agent feeding device. The whole process does not require manual intervention, which replaces the traditional manual experience operation (comparative example 1 carbon temperature double hit rate is only 84%), and greatly improves production stability and consistency. 6. The BP-PSO hybrid inverse model is adopted, which is mainly data-driven and supplemented by mechanism constraints, without relying on precise mechanism model parameters. Compared with the pure mechanism model (Comparative Example 3), the present invention does not require precise thermodynamic and kinetic parameter calibration, and the model maintenance cost is low. Compared with the pure data-driven model (Comparative Example 2), the present invention ensures the metallurgical rationality of the inverse solution results through mechanism constraints, and avoids large deviations in the data-driven model when the operating conditions fluctuate. Attached Figure Description

[0021] Figure 1 It is a process for reverse solving of carbon-oxygen balance and optimization of process parameters in converter blowing. Detailed Implementation

[0022] The present invention will now be described in detail with reference to the accompanying drawings, but it should be noted that the implementation of the present invention is not limited to the following embodiments.

[0023] The following embodiments are implemented based on the technical solution of the present invention, providing detailed implementation methods and specific operation processes. However, the scope of protection of the present invention is not limited to the following embodiments. Unless otherwise specified, the methods used in the following embodiments are conventional methods.

[0024] Example 1:

[0025] See Figure 1 A method for reverse carbon-oxygen balance solution and process parameter optimization in converter blowing includes: S1. Collect data from the entire converter blowing process and perform data preprocessing; A data acquisition system for the entire converter blowing process was established to comprehensively collect various key data during the smelting process. This embodiment uses a 180t top-and-bottom combined blowing converter to smelt ordinary carbon steel as the test object, and the endpoint control targets are: carbon content 0.05%~0.08%, temperature 1620~1650℃, and oxygen activity 300~600ppm.

[0026] S11. The content to be collected includes: (1) Initial conditions for molten iron: Initial carbon content =4.5%, initial silicon content =0.5%, initial manganese content =0.1%, initial sulfur content =0.03%, initial phosphorus content =0.08%, initial temperature =1350℃, molten iron weight =180t.

[0027] (2) Real-time status of the molten pool: Real-time carbon content Real-time oxygen activity Real-time temperature , furnace gas composition and Volume fraction, sampling frequency is 1 time / 30 seconds; (3) Process parameters: Oxygen supply flow rate (Range 35000~42000 m³ / h), Oxygen Lance Position (Range 1.2~2.0m), bottom blowing air volume (Range 5~20 m³ / min), Timing and Dosage of Adding Slag-Forming Agents (Lime, Dolomite) ; (4) Equipment status: Furnace volume ratio =0.85, oxygen lance aperture =50mm, furnace age =500 furnaces.

[0028] S12, Data Preprocessing: Outliers were removed using the 3σ criterion: the mean μ and standard deviation σ of each parameter were calculated, and data points exceeding the range of μ ± 3σ were removed. Missing values ​​were filled using linear interpolation: linear interpolation was performed on valid data from adjacent time points. The data were standardized to the [0,1] interval using the min-max normalization method. ; in: The data is after normalization; This represents the minimum value of the original data, with units consistent with the specific parameters. This is the raw data; the units are consistent with the specific parameters. This represents the maximum value of the original data, with the unit consistent with the specific parameter.

[0029] Eliminate the influence of dimensions and construct a standardized carbon-oxygen reaction dataset.

[0030] S2. Construct a carbon-oxygen equilibrium inverse solution model; The carbon-oxygen equilibrium inverse model comprises a mechanistic constraint framework and a hybrid BP-PSO model. Using the thermodynamics and kinetics of the carbon-oxygen reaction as constraints, it leverages the PSO algorithm to optimize the initial weights and thresholds of the BP neural network. This addresses the problems of multiple solutions, local optima, and slow convergence in traditional inverse models, achieving a precise and stable inverse solution from the target state to process parameters. S21, Mechanistic Constraint Framework: Based on the thermodynamic and kinetic laws of the carbon-oxygen reaction in a converter, a core constraint formula is established to ensure that the inverse solution result conforms to the essence of the metallurgical reaction. The formula is as follows: (1) Thermodynamic constraints on the carbon-oxygen reaction are used to limit the output boundary of the inverse solution model, and the expression is: , ; in: The carbon content of the molten pool is expressed as % (%). Oxygen content in the molten pool, in % % It is carbon monoxide gas; The temperature of the molten pool is expressed in Kelvin (K). Standard Gibbs free energy, expressed in J / mol; 22100 and 38.3 are empirical thermodynamic coefficients for the carbon-oxygen reaction.

[0031] when When <0, the carbon-oxygen reaction can proceed spontaneously.

[0032] (2) Carbon-oxygen reaction kinetic constraint, used to limit the rate range of the carbon-oxygen reaction, to avoid the reverse solution parameter causing the carbon oxidation rate to be too fast or too slow, leading to problems such as back-drying and splashing. The expression is: ; in: This represents the carbon oxidation rate, expressed as % / min. The reaction rate constant is positively correlated with temperature, and its value ranges from 0.02 to 0.08 min. -1 ; This refers to the partial pressure of CO, measured in atm. This is the equilibrium constant for the carbon-oxygen reaction, which varies with smelting temperature. S22, BP-PSO hybrid inverse solution model: using a BP neural network as the solver, and employing the PSO algorithm to optimize the initial weights and thresholds of the BP neural network, thereby improving the model's convergence speed and inverse solution accuracy; (1) BP neural network structure: The input layer contains the target endpoint parameters, including the target carbon content. =0.06%, target oxygen activity =500ppm, target endpoint temperature =1635℃.

[0033] The hidden layer has 12 nodes. After multiple tests comparing the number of nodes, the 12-node method showed the lowest error, stable convergence, and no obvious overfitting, while balancing accuracy and efficiency.

[0034] The output layer contains the initial values ​​of process parameters, including oxygen supply flow rate. Oxygen lance position Bottom blowing volume The activation function used is the Sigmoid function: the output formula is as follows: ; in: This is the output value of the kth node in the output layer, i.e., the initial value of the process parameters; This is the input value of the i-th node in the input layer, i.e., the target endpoint parameter; The weights from the i-th node in the hidden layer to the k-th node in the output layer; This is the bias of the k-th node in the output layer; The Sigmoid activation function maps any real number to the interval (0, 1), and is used to map the output value to a reasonable range of process parameters.

[0035] (2) The PSO algorithm is used to optimize the initial weights of the BP neural network. With bias To avoid the BP neural network getting trapped in local optima, the iterative formula is as follows: ; in: Let be the velocity of the i-th particle in the j-th dimension during the (t+1)-th iteration; Let be the velocity of the i-th particle in the j-th dimension during the t-th iteration; , =1.5 is the learning factor (acceleration coefficient), which represents the step size for the particle to learn towards the individual optimum and the global optimum, respectively; , A random number within the interval [0,1], used to introduce random perturbations; Let be the individual historical optimal position of the i-th particle in the j-th dimension during the t-th iteration; Let be the globally optimal position of the group in the j-th dimension during the t-th iteration; For the first The current position of a particle in the j-th dimension during the t-th iteration; The inertia weight is used to balance global exploration and local development capabilities, employing a linear decreasing strategy, and its expression is: ; in: Let be the inertia weight at the t-th iteration; The initial inertia weight is set to 1.2. The final inertia weight is set to 0.8. This represents the maximum number of iterations, with a value of 100. This represents the current iteration number, with a value ranging from 1 to 100.

[0036] (3) The inverse solution process is as follows: Target endpoint parameter: target carbon content =0.06%, target oxygen activity =500ppm, target endpoint temperature The optimized BP-PSO hybrid model was input at 1635℃, and the calculation process is as follows: The first step is to normalize the input data, mapping the input data to the [0,1] interval; ; in: The original value of the i-th input parameter ( , , ); This is the minimum value of the i-th input parameter, with the unit consistent with the specific parameter. This represents the maximum value of the i-th input parameter, with the unit consistent with the specific parameter. The normalized value of the i-th input parameter; The value range of each parameter is determined based on historical data statistics: 0.02%–0.12%, after normalization ; 200~800ppm, after normalization ; 1600~1680℃, after normalization ; The second step is to calculate the output layer node values ​​through forward propagation. The input to the j-th node (j=1,...12) in the hidden layer is: ; in: This is the input value of the j-th node in the hidden layer; The weights from the i-th node in the input layer to the j-th node in the hidden layer; This is the bias of the j-th node in the hidden layer; The hidden layer output is: ; in: This represents the output value of the j-th node in the hidden layer. Use the Sigmoid activation function; The negative of the natural constant e Power of; Taking the optimized partial weights as an example: Node j=1; =2.5×0.40+1.8×0.50+0.5×0.4375-1.2=0.91875, =0.714; Node j=2; =0.8×0.40+2.2×0.50+1.5×0.4375-0.5=1.57625, =0.829; The remaining hidden layer nodes are calculated in the same way to obtain the hidden layer output vector h=[0.714,0.829,…] (12 values).

[0037] Step 3: Calculate the normalized value of the output layer; The input to the k-th node of the output layer (k=1,2,3, corresponding to oxygen supply flow rate, oxygen gun position, and bottom blowing volume respectively) is: ; The normalized value of the output layer is: ; in: This is the normalized output value of the k-th node in the output layer; Use the Sigmoid activation function; The negative of the natural constant e Power of; This is the input value of the k-th node in the output layer; The weights from the j-th node in the hidden layer to the k-th node in the output layer; This represents the output value of the j-th node in the hidden layer. This is the bias of the k-th node in the output layer.

[0038] The output layer weights and biases after PSO optimization are calculated as follows: Oxygen supply flow node (k=1): , ; Oxygen lance position node (k=2): , ; Bottom blowing volume node (k=3): , .

[0039] Step 4: Denormalize the output layer Denormalize the output layer normalized values ​​to the actual physical quantities: ; in: This is the actual value of the kth output parameter after inverse normalization, with the unit consistent with the specific parameter. This is the minimum value of the k-th output parameter, with the unit consistent with the specific parameter. This is the normalized value of the k-th output parameter; This represents the maximum value of the k-th output parameter, with the unit consistent with the specific parameter. The value range of each output parameter: Oxygen supply flow rate: 35,000–42,000 m³ / h =35000+0.35×(42000-35000)=37450m 3 / h; Oxygen lance position: 1.2–2.0 m =1.2 + 0.65 × (2.0 - 1.2) = 1.72 m; Bottom blowing volume: 5–20 m³ / min =5 + 0.40 × (20 - 5) = 11m 3 / min.

[0040] Step 5: Set the confidence interval; Based on the model's prediction error distribution on the validation set, the confidence interval for each output parameter is set to ±5%. Oxygen supply flow rate: 37450 ± 1872 m³ / h (35578~39322 m³ / h); Oxygen lance position: 1.72 ±0.086 m (1.634~1.806 m); Bottom blowing air volume: 11 ± 0.55 m³ / min (10.45~11.55 m³ / min).

[0041] This confidence interval can be dynamically adjusted according to the actual fluctuation of operating conditions.

[0042] S3. Establish a phased, multi-constraint dynamic parameter optimization model; Based on the carbon and oxygen physicochemical properties of the initial, middle and late stages of converter blowing, the blowing process is divided into three stages: initial stage (silicon-manganese oxidation period, 0-4 min), middle stage (main carbon oxidation period, 4-10 min), and late stage (endpoint control period, 10-16 min). For each stage, a corresponding objective function and constraint conditions are established. The NSGA-Ⅲ multi-objective optimization algorithm is used to iteratively optimize the initial values ​​of process parameters output by the inverse solution into the optimal dynamic curve of the entire converter blowing process. (1) Phased objectives and constraints; ① Initial stage (silicon-manganese oxidation period, 0-4 min): The objective is rapid slag formation to lay the foundation for subsequent carbon-oxygen reactions. The constraints focus on the heating rate, silicon-manganese oxidation rate, and slag basicity, as follows: Heating rate: =95 ℃ / min, controlled within the range of 80~100 ℃ / min; Silicon oxidation rate: =92%; Manganese oxidation rate: =75%; Slag alkalinity: =1.35, controlled within the range of 1.2~1.5.

[0043] in: To control the heating rate of the molten pool, ensure rapid slag formation in the initial stage while avoiding excessively rapid heating that could lead to splashing; , These are the oxidation rates of silicon and manganese, ensuring sufficient oxidation of silicon and manganese and reducing the burden of slag formation in the later stages; The slag alkalinity (the mass ratio of CaO to SiO2 in the slag) is used to control slag fluidity and desulfurization and dephosphorization effects.

[0044] ②Mid-term (main carbon oxidation period, 4–10 min): The objective is to ensure a stable carbon-oxygen reaction. The constraints focus on the carbon oxidation rate, the degree of slag foaming, and the risk of splashing. The specific constraints are as follows: Carbon oxidation rate: =0.045% / min, controlled within the range of 0.03~0.06% / min; Degree of slag foaming: =1.0, controlled within the range of 0.8~1.2; Splash probability: =3%≤5%; Bottom blowing volume: =15 m3 / min, controlled within the range of 5~20 m3 / min.

[0045] ③ Later stage (endpoint control period, 10-16 min): The objective is to precisely control the final carbon temperature. The constraints are carbon content deviation, temperature deviation, and overblowing rate, as detailed below: Carbon content deviation: =±0.005%, controlled within the range of ±0.01%; Temperature deviation: =±3℃, controlled within the range of ±5℃; endpoint oxygen activity: =500ppm, controlled within the range of 300~600ppm; Overblowing rate: =1.2%≤2%.

[0046] (2) NSGA-III multi-objective optimization algorithm; With quality, efficiency, cost, and low carbon as the four core objectives, a multi-objective optimization function is constructed. The initial values ​​of process parameters output by the inverse solution are iteratively optimized into the optimal dynamic parameter curves for each stage. A minimization objective system is established and adapted to the solution logic of the NSGA-Ⅲ algorithm. ① The multi-objective optimization function is defined as follows: (Ⅳ) in: Optimize function values ​​for multiple objectives; ~ These are the weighting coefficients, which are 0.3, 0.2, 0.2, 0.15, and 0.15 respectively, with a total weight of 1. The endpoint is the carbon temperature double hit rate, with a value range of 0 to 100%. The maximum hit rate is transformed into the minimum 1-hit rate through the inverse objective function, which is adapted to the minimization solution logic of the NSGA-Ⅲ algorithm. The blowing cycle is defined as 10–16 min. For steel consumption, the value ranges from 1050 to 1100 kg / t; The slag-forming agent consumption mainly includes lime, dolomite, etc., with a value range of 20-60 kg / t; The value represents CO2 emissions, ranging from 1.5 to 2.9 t / t steel.

[0047] ②Optimization process: The initial values ​​of the process parameters output from the inverse kinematics algorithm are used as the initial population for the NSGA-Ⅲ algorithm: Initial oxygen supply flow rate: 37450 m³ / h; Initial position of oxygen lance: 1.72 m; Initial bottom blowing air volume: 11 m³ / min.

[0048] Using (Ⅰ) to (Ⅲ) as constraints and objective function (Ⅳ) as the optimization objective, the number of iterations is set to 80 (which can be adjusted between 50 and 100 depending on the complexity of the working conditions) to ensure that the multi-objective optimization fully converges.

[0049] After iterative optimization using NSGA-Ⅲ, the optimal dynamic process parameter curves for each stage were obtained:

[0050] A phased dynamic adaptive optimization of process parameters was achieved.

[0051] S4. Establish a real-time feedback closed-loop control mechanism; dynamically trigger inverse solution recalculation and parameter optimization iteration based on the deviation between online detection data and the target; the real-time feedback closed-loop control mechanism includes: Real-time monitoring: Sampling frequency is 1 time / 30 seconds to obtain real-time carbon content. Real-time oxygen activity Real-time temperature ; Deviation calculation: ; ; ; in: The value range is for carbon content deviation. ; Real-time carbon content, in % % =0.06% is the target carbon content, in percentages (%). For temperature deviation, the value range is: ; This is the real-time temperature, in °C. =1635 is the target temperature, in °C; This is due to oxygen activity deviation; Real-time oxygen activity; =500ppm is the target oxygen activity.

[0052] Deviation judgment: Set deviation threshold , All real-time monitoring data throughout the process meet the requirements. Within ±0.005%, In ±3 Within.

[0053] If the deviation is within the threshold, keep the current process parameters unchanged; If the deviation exceeds the threshold, the feedback correction process will begin. Feedback and corrections: The deviation signal is fed back to the carbon-oxygen balance inverse solution model and the staged optimization model to recalculate the initial values ​​of the process parameters. Then, it is input into the staged multi-constraint dynamic optimization model for a second iteration to output new optimal process parameters, thereby achieving real-time correction of the parameters.

[0054] Through the aforementioned closed-loop control mechanism, the system can adapt to complex operating conditions such as fluctuations in molten iron and changes in furnace type in real time, ensuring a stable blowing process and precise endpoint control.

[0055] S5. Optimize instruction output and automatic execution; The dynamically optimized process parameters, refined in stages, are sent to the converter's basic automation system in the form of digital commands, specifically including: Oxygen supply flow rate dynamic curve: initial stage 39000 m³ / h → middle stage 38000 m³ / h → later stage 40000 m³ / h; Oxygen lance position dynamic curve: initial stage 1.8 m → middle stage 1.6 m → later stage 1.3 m; Bottom blowing volume dynamic curve: initial stage 10 m³ / min → middle stage 15 m³ / min → later stage 12 m³ / min; Timing and amount of slag-forming agent added: Initially add 1500 kg of lime and 500 kg of dolomite.

[0056] The basic automation system controls the automatic operation of oxygen lances, oxygen supply valves, bottom blowing valves, and slagging agent feeding devices, achieving fully automatic closed-loop precise control of the entire converter blowing process without manual intervention.

[0057] Example 2:

[0058] The test was conducted using a 180t top-and-bottom blown converter to smelt plain carbon steel. The endpoint control targets were: carbon content 0.05%~0.08%, temperature 1620~1650℃, and oxygen activity 300~600ppm. The molten iron charge was 180t, and the initial composition was as follows: =4.5%, =0.7%, =0.1%, equipment parameters are furnace volume ratio =0.85, oxygen lance orifice diameter 50mm. The core difference from Example 1 is that the initial silicon content of the molten iron is increased from 0.5% to 0.7%, while the other basic conditions remain the same.

[0059] A method for reverse carbon-oxygen balance solution and process parameter optimization in converter blowing includes: S1. Collect data from the entire converter blowing process and perform data preprocessing; Establish a data acquisition system for the entire converter blowing process to comprehensively collect various key data during the smelting process.

[0060] S11. The content to be collected includes: (1) Initial conditions for molten iron: Initial carbon content =4.5%, initial silicon content =0.7%, initial manganese content =0.1%, initial sulfur content =0.03%, initial phosphorus content =0.08%, initial temperature =1350℃, molten iron weight =180t.

[0061] (2) Real-time status of the molten pool: Real-time carbon content Real-time oxygen activity Real-time temperature , furnace gas composition and Volume fraction, sampling frequency is 1 time / 30 seconds; (3) Process parameters: Oxygen supply flow rate (Range 35000~42000 m³ / h), Oxygen Lance Position (Range 1.2~2.0m), bottom blowing air volume (Range 5~20 m³ / min), Timing and Dosage of Adding Slag-Forming Agents (Lime, Dolomite) ; (4) Equipment status: Furnace volume ratio =0.85, oxygen lance aperture =50mm, furnace age =500 furnaces.

[0062] S12, Data Preprocessing: Outliers were removed using the 3σ criterion: the mean μ and standard deviation σ of each parameter were calculated, and data points that exceeded the range of μ ± 3σ were removed.

[0063] Taking the silicon content in molten iron as an example, the historical data mean μ = 0.52%, and the standard deviation σ = 0.08%. After removing data points exceeding the range of 0.28% to 0.76%, this embodiment... =0.7% is within the valid range. Missing values ​​are filled using linear interpolation: linear interpolation is performed on valid data from adjacent time points. The data is standardized to the [0,1] interval using min-max normalization. ; Taking the carbon content of molten iron as an example: =4.5%, =4.0%, =5.0%, then =(4.5-4.0) / (5.0-4.0)=0.5. Eliminate the influence of dimensions and construct a standardized carbon-oxygen reaction dataset.

[0064] S2. Construct a carbon-oxygen equilibrium inverse solution model; The carbon-oxygen balance inverse model includes a mechanistic constraint framework and a hybrid BP-PSO inverse model. During data preprocessing, a high silicon content (0.7%) was identified, and the carbon-oxygen balance inverse model automatically adjusted the constraint thresholds. S21, Mechanistic Constraint Framework: (1) Thermodynamic constraints of carbon-oxygen reaction: , ; Taking an initial silicon content of 0.7% in molten iron as an example, the calculation... = -22100 - 38.3×1623 ≈ -84200 J / mol < 0, the carbon-oxygen reaction can proceed spontaneously.

[0065] (2) Kinetic constraints of carbon-oxygen reaction: ; Taking the intermediate carbon oxidation rate as an example, where Take 0.05 min -1 [C] = 0.3% =1.2 atm, =450, calculated as follows ≈0.048% / min.

[0066] S22, BP-PSO hybrid inverse solution model: (1) BP neural network structure: The input layer contains the target endpoint parameters, including the target carbon content. =0.06%, target oxygen activity =500ppm, target endpoint temperature =1635℃.

[0067] The hidden layer has 12 nodes. After multiple tests comparing the number of nodes, the 12-node method showed the lowest error, stable convergence, and no obvious overfitting, while balancing accuracy and efficiency.

[0068] The output layer contains the initial values ​​of process parameters, including oxygen supply flow rate. Oxygen lance position Bottom blowing volume The activation function used is the Sigmoid function: the output formula is as follows: ; Input layer target carbon content Taking 0.06% as an example, after normalization =0.40, hidden layer output After weight calculation, the values ​​are passed to the output layer, and the oxygen supply flow rate node in the output layer is then processed. After mapping with the Sigmoid function, the normalized value is 0.42.

[0069] (2) The PSO algorithm is used to optimize the initial weights of the BP neural network. With bias To avoid the BP neural network getting trapped in local optima, the iterative formula is as follows: ; The inertia weight is used to balance global exploration and local development capabilities, employing a linear decreasing strategy, and its expression is: ; The PSO algorithm iterative formula combined with specific data: taking the 50th iteration as an example; =1.2-(1.2-0.8) / 100×50=1.0; set up =0.1, = =1.5, =0.3, =0.6, =1.2, 0, =1.5; but =1.0×0.1+1.5×0.3×(1.2-1.0)+1.5×0.6×(1.5-1.0)=0.1+0.09+0.45=0.64.

[0070] (3) The inverse solution process is as follows: Target endpoint parameter: target carbon content =0.06%, target oxygen activity =500ppm, target endpoint temperature The optimized BP-PSO hybrid model was input at 1635℃, and the calculation process is as follows: The first step is to normalize the input data, mapping the input data to the [0,1] interval; Same as Example 1.

[0071] The second step is to calculate the output layer node values ​​through forward propagation. Same as Example 1.

[0072] Step 3: Calculate the normalized value of the output layer; The input to the k-th node of the output layer (k=1,2,3, corresponding to oxygen supply flow rate, oxygen gun position, and bottom blowing volume respectively) is: ; The normalized value of the output layer is: ; The output layer weights and biases after PSO optimization are calculated as follows: The calculation process for the oxygen supply flow node (k=1) is as follows: =-0.62, then; ; The calculation process for the oxygen lance position node (k=2) is as follows: =-0.62, then; ; The calculation process for the bottom blowing volume node (k=3) is as follows: =-0.41, then: .

[0073] Step 4: Denormalize the output layer Denormalize the output layer normalized values ​​to the actual physical quantities: ; The value range of each output parameter: Oxygen supply flow rate: =35000+0.42×(42000-35000)=35000+2940=37940≈39500 m³ / h (after model correction); Oxygen lance position: 1.2–2.0 m =1.2+0.70×(2.0-1.2)≈1.85m (considering the high silicon content, the gun position should be appropriately increased); Bottom blowing volume: 5–20 m³ / min =5 + 0.38 × (20 - 5) ≈ 11m 3 / min.

[0074] Step 5: Set the confidence interval; The confidence interval for each output parameter is set to ±5%, which can be adjusted according to actual working conditions.

[0075] Through the above inverse solution process, the initial value of oxygen supply flow rate is 39500 m³ / h, the initial value of oxygen lance position is 1.85m, and the initial value of bottom blowing air volume is 11 m³ / min.

[0076] S3. Establish a phased, multi-constraint dynamic parameter optimization model; The blowing process is divided into three stages: initial stage (silicon-manganese oxidation period, 0-4 min), middle stage (main carbon oxidation period, 4-10 min), and final stage (endpoint control period, 10-16 min). (1) Phased objectives and constraints; ① Initial stage (silicon-manganese oxidation period, 0-4 min): The goal is to achieve rapid slag formation. Considering the high silicon content, the initial lime addition will be increased to 2000 kg, and the slag basicity will be appropriately increased.

[0077] Heating rate: =98 ℃ / min, controlled within the range of 80~100 ℃ / min; Silicon oxidation rate: =95%; Manganese oxidation rate: =78%; Slag alkalinity: =1.40, controlled within the range of 1.2~1.5.

[0078] ②Mid-term (main carbon oxidation period, 4–10 min): The goal is to ensure the smooth progress of the carbon-oxygen reaction.

[0079] Carbon oxidation rate: =0.048% / min, controlled within the range of 0.03~0.06% / min; Degree of slag foaming: =1.0, controlled within the range of 0.8~1.2; Splash probability: =3.5%≤5%; Bottom blowing volume: =15m 3 / min, controlled at 5~20 m 3 / min interval.

[0080] ③ Later stage (endpoint control period, 10-16 min): The goal is to precisely control the final carbon temperature.

[0081] Carbon content deviation: =±0.006%, controlled within the range of ±0.01%; Temperature deviation: =±4℃, controlled within the range of ±5℃; endpoint oxygen activity: =500ppm, controlled within the range of 300~600ppm; Overblowing rate: =1.5%≤2%.

[0082] (2) NSGA-III multi-objective optimization algorithm; ① The multi-objective optimization function is defined as follows: (Ⅳ) The initial values ​​of the process parameters output from the inverse kinematics solution (oxygen supply 39500 m³ / h, nozzle position 1.85 m, bottom blowing 11 m³ / min) were used as the initial population for the NSGA-Ⅲ algorithm, with 90 iterations. The optimal dynamic process parameter curves for each stage were obtained:

[0083] S4. Establish a real-time feedback closed-loop control mechanism; Real-time monitoring: Sampling frequency is 1 time / 30 seconds to obtain real-time carbon content. Real-time oxygen activity Real-time temperature ; Deviation calculation: ; ℃; ppm; Deviation detection: Set deviation threshold , All real-time monitoring data throughout the process meet the requirements. Within ±0.005%, In ±3 Within.

[0084] Feedback and correction process: Mid-term (around the 6th minute) real-time monitoring detected a slight anomaly in the carbon oxidation rate; measured value =0.028% / min, which is lower than the constraint limit of 0.03% / min. The deviation exceeds the threshold, triggering the feedback correction process. The deviation signal is fed back to the inverse carbon-oxygen equilibrium model; Based on the current molten pool state parameters ( =0.35%, =1580℃ Using 450ppm as the new input condition, the inverse S2 solution calculation is re-executed; The inverse kinematic model outputs the corrected process parameters: oxygen supply flow rate is adjusted to 37800 m³ / h, and gun position is adjusted to 1.58m; The corrected parameters are input into the phased optimization model for a second iteration. Output new optimal process parameters to bring the carbon oxidation rate back to the reasonable range of 0.035% / min.

[0085] Through the aforementioned closed-loop control mechanism, the system adapts in real time to the fluctuations in carbon-oxygen reaction under high-silicon molten iron conditions, ensuring the stability of the blowing process.

[0086] S5. Optimize instruction output and automatic execution; The dynamically optimized process parameters, refined in stages, are sent to the converter's basic automation system in the form of digital commands, specifically including: Oxygen supply flow rate dynamic curve: initial stage 39500 m³ / h → middle stage 38500 m³ / h → later stage 40500 m³ / h; Oxygen lance position dynamic curve: initial stage 1.85 m → middle stage 1.65 m → later stage 1.35 m; Bottom blowing volume dynamic curve: initial stage 10 m³ / min → middle stage 15 m³ / min → later stage 12 m³ / min; Timing and amount of slag-forming agent added: Initially add 2000 kg of lime and 600 kg of dolomite.

[0087] The basic automation system controls the automatic operation of each device, achieving fully automatic closed-loop precise control.

[0088] Implementation results:

[0089] This invention fully demonstrates its adaptability to high-silicon molten iron conditions. By automatically adjusting the constraint threshold through the inverse kinematics model, increasing the initial lime addition, and real-time correction through the closed-loop control mechanism, the stability and control accuracy of the blowing process are ensured. Example 3:

[0090] The test was conducted using a 180t top-and-bottom blown converter to smelt plain carbon steel. The endpoint control targets were: carbon content 0.05%~0.08%, temperature 1620~1650℃, and oxygen activity 300~600ppm. The molten iron charge was 180t, and the initial composition was as follows: =4.5%, =0.5%, =0.1%, equipment parameters are: oxygen lance orifice diameter 50mm. The core difference from Example 1 is that the converter furnace volume ratio is adjusted to... =0.7 (furnace type is relatively compact), all other basic conditions remain the same. A method for reverse carbon-oxygen balance solution and process parameter optimization in converter blowing includes: S1. Collect data from the entire converter blowing process and perform data preprocessing; Establish a data acquisition system for the entire converter blowing process to comprehensively collect various key data during the smelting process.

[0091] S11. The content to be collected includes: (1) Initial conditions for molten iron: Initial carbon content =4.5%, initial silicon content =0.5%, initial manganese content =0.1%, initial sulfur content =0.03%, initial phosphorus content =0.08%, initial temperature =1350℃, molten iron weight =180t.

[0092] (2) Real-time status of the molten pool: Real-time carbon content Real-time oxygen activity Real-time temperature , furnace gas composition and Volume fraction, sampling frequency is 1 time / 30 seconds; (3) Process parameters: Oxygen supply flow rate (Range 35000~42000 m³ / h), Oxygen Lance Position (Range 1.2~2.0m), bottom blowing air volume (Range 5~20 m³ / min), Timing and Dosage of Adding Slag-Forming Agents (Lime, Dolomite) ; (4) Equipment status: Furnace volume ratio =0.7, oxygen lance aperture =50mm, furnace age =500 furnaces.

[0093] S12, Data Preprocessing: Outliers are removed using the 3σ criterion: the mean μ and standard deviation σ of each parameter are calculated, and data points exceeding the range of μ ± 3σ are removed. Taking the furnace capacity ratio as an example, the historical data mean μ = 0.85 and standard deviation σ = 0.05; data points exceeding the range of 0.70 to 1.00 are removed in this embodiment. =0.7 is retained at the boundary and marked as a special working condition. Missing values ​​are filled using linear interpolation: linear interpolation is performed on valid data from adjacent time points. The data is standardized to the [0,1] interval using the min-max normalization method. ; Eliminate the influence of dimensions and construct a standardized carbon-oxygen reaction dataset.

[0094] Taking molten iron temperature as an example: =1350℃, =1300℃, =1400℃, but =(1350-1300) / (1400-1300)=0.5.

[0095] Eliminate the influence of dimensions and construct a standardized carbon-oxygen reaction dataset.

[0096] S2. Construct a carbon-oxygen equilibrium inverse solution model; The carbon-oxygen balance inverse model includes a mechanistic constraint framework and a hybrid BP-PSO inverse model. Considering the relatively small furnace volume ratio ( =0.7 (compact furnace type), the confidence interval is tightened from ±5% to ±3% in the inverse kiln stage to ensure that the process parameters are adapted to the compact furnace type and to avoid splashing.

[0097] S21, Mechanistic Constraint Framework: (1) Thermodynamic constraints of carbon-oxygen reaction: , ; Taking the furnace volume ratio of 0.7 in this embodiment as an example, the calculation is as follows: = -22100 - 38.3×1623 ≈ -84200 J / mol < 0, (2) Kinetic constraints of carbon-oxygen reaction: ; Taking the intermediate carbon oxidation rate as an example, where Take 0.045 min⁻¹, =0.35%, =1.1 atm, =460, calculated as follows ≈0.040% / min.

[0098] S22, BP-PSO hybrid inverse solution model: (1) BP neural network structure: The input layer contains the target endpoint parameters, including the target carbon content. =0.06%, target oxygen activity =500ppm, target endpoint temperature =1635℃.

[0099] The hidden layer has 12 nodes. After multiple tests comparing the number of nodes, the 12-node method showed the lowest error, stable convergence, and no obvious overfitting, while balancing accuracy and efficiency.

[0100] The output layer contains the initial values ​​of process parameters, including oxygen supply flow rate. Oxygen lance position Bottom blowing volume The activation function used is the Sigmoid function: the output formula is as follows: .

[0101] (2) The PSO algorithm is used to optimize the initial weights of the BP neural network. With bias To avoid the BP neural network getting trapped in local optima, the iterative formula is as follows: ; The inertia weight is used to balance global exploration and local development capabilities, employing a linear decreasing strategy, and its expression is: .

[0102] (3) The inverse solution process is as follows: Target endpoint parameter: target carbon content =0.06%, target oxygen activity =500ppm, target endpoint temperature The optimized BP-PSO hybrid model was input at 1635℃, and the calculation process is as follows: The first step is to normalize the input data, mapping the input data to the [0,1] interval; Same as Example 1.

[0103] The second step is to calculate the output layer node values ​​through forward propagation. Same as Example 1.

[0104] Step 3: Calculate the normalized value of the output layer; The input to the k-th node of the output layer (k=1,2,3, corresponding to oxygen supply flow rate, oxygen gun position, and bottom blowing volume respectively) is: ; The normalized value of the output layer is: ; The output layer weights and biases after PSO optimization are calculated as follows: Oxygen supply flow node (k=1). ,but ; Oxygen lance position node (k=2). ,but ; Bottom blowing volume node (k=3). ,but .

[0105] Step 4: Denormalize the output layer; Denormalize the output layer normalized values ​​to the actual physical quantities: ; The value range of each output parameter: Oxygen supply flow rate: 35,000–42,000 m³ / h =35000+0.34×(42000−35000)≈37500m 3 / h; Oxygen lance position: 1.2–2.0 m =1.2+0.60×(2.0−1.2)≈1.70m; Bottom blowing volume: 5–20 m³ / min =5 + 0.35 × (20 − 5) ≈ 10m 3 / min.

[0106] Step 5: Set the confidence interval; Considering the higher requirements for parameter accuracy in compact furnace designs, the confidence interval has been tightened from ±5% to ±3%. Oxygen supply flow rate: 37500 ± 1125 m³ / h (36375~38625 m³ / h); Oxygen lance position: 1.70 ± 0.051 m (1.649~1.751 m); Bottom blowing air volume: 10 ± 0.3 m³ / min (9.7~10.3 m³ / min).

[0107] Through the above inverse solution process, the initial values ​​of oxygen supply flow rate are output as 37500 m³ / h, oxygen lance position as 1.70m, and bottom blowing volume as 10 m³ / min, with a confidence interval of ±3%.

[0108] S3. Establish a phased, multi-constraint dynamic parameter optimization model; The blowing process was divided into three stages: initial stage (silicon-manganese oxidation period, 0–4 min), middle stage (main carbon oxidation period, 4–10 min), and final stage (endpoint control period, 10–16 min). Considering the sensitivity of the compact furnace type to the degree of foaming, the foaming constraint of the slag was strengthened.

[0109] (1) Phased objectives and constraints; ① Initial stage (silicon-manganese oxidation period, 0-4 min): The goal is to achieve rapid slag formation, while controlling the heating rate to avoid splashing.

[0110] Heating rate: =88 ℃ / min, controlled within the range of 80~100 ℃ / min (taking the lower value to adapt to compact furnace type); Silicon oxidation rate: =91%; Manganese oxidation rate: =73%; Slag alkalinity: =1.30, controlled within the range of 1.2~1.5.

[0111] ②Mid-term (main carbon oxidation period, 4–10 min): The goal is to ensure the smooth progress of the carbon-oxygen reaction and to strengthen the control of slag foaming.

[0112] Carbon oxidation rate: =0.040% / min, controlled within the range of 0.03~0.06% / min; Degree of slag foaming: =0.9, controlled within the range of 0.8~1.2 (take the lower value to avoid excessive foam causing splashing); Splash probability: =3.5%≤5%; Bottom blowing volume: =11m 3 / min, controlled at 5~20 m 3 / min range (adjusted to a lower value to enhance molten pool stirring and promote full carbon-oxygen reaction).

[0113] ③ Later stage (endpoint control period, 10-16 min): The goal is to precisely control the final carbon temperature.

[0114] Carbon content deviation: =±0.007%, controlled within the range of ±0.01%; Temperature deviation: =±4℃, controlled within the range of ±5℃; endpoint oxygen activity: =500ppm, controlled within the range of 300~600ppm; Overblowing rate: =1.4%≤2%.

[0115] (2) NSGA-III multi-objective optimization algorithm; ① The multi-objective optimization function is defined as follows: (Ⅳ) The initial values ​​of the process parameters output from the inverse kinematics solution (oxygen supply 37500 m³ / h, nozzle position 1.70 m, bottom blowing 10 m³ / min) were used as the initial population for the NSGA-Ⅲ algorithm, with 80 iterations. The optimal dynamic process parameter curves for each stage were obtained:

[0116] S4. Establish a real-time feedback closed-loop control mechanism; Real-time monitoring: Sampling frequency is 1 time / 30 seconds to obtain real-time carbon content. =0.055%, Real-time oxygen activity =480ppm, real-time temperature =1641℃; Deviation calculation: ; ℃; ppm; Deviation detection: Set deviation threshold , .

[0117] Feedback and correction process: (1) The deviation signal is fed back to the carbon-oxygen equilibrium inverse solution model; (2) Based on the current molten pool state parameters ( =0.055%, =1641℃ Using 480ppm as the new input condition, the inverse S2 solution calculation is re-executed; (3) The reverse model outputs the corrected process parameters: the oxygen supply flow rate is adjusted to 39800 m³ / h and the gun position is adjusted to 1.25 m; (4) Input the corrected parameters into the phased optimization model for a second iteration; (5) Output new optimal process parameters and shorten the blowing time by 0.3 min to 14.7 min to ensure that the final carbon temperature meets the standard.

[0118] Through the aforementioned closed-loop control mechanism, the system can maintain high control accuracy and good stability even under compact furnace operating conditions.

[0119] S5. Optimize instruction output and automatic execution; The dynamically optimized process parameters, refined in stages, are sent to the converter's basic automation system in the form of digital commands, specifically including: Oxygen supply flow rate dynamic curve: initial stage 38000 m³ / h → middle stage 37000 m³ / h → later stage 39000 m³ / h (corrected to 39800). Oxygen lance position dynamic curve: initial stage 1.75 m → middle stage 1.55 m → later stage 1.25 m (corrected to 1.25 m); Bottom blowing volume dynamic curve: initial stage 9 m³ / min → middle stage 11 m³ / min → later stage 10 m³ / min; Timing and amount of slag-forming agent added: Initially add 1500 kg of lime and 500 kg of dolomite.

[0120] The basic automation system controls the automatic operation of each device, achieving fully automatic closed-loop precise control.

[0121] Implementation results:

[0122] This invention fully demonstrates its adaptability to compact furnace conditions (smaller furnace volume ratio). By narrowing the confidence interval from ±5% to ±3% during the inverse kiln stage, strengthening slag foaming constraints, adjusting bottom blowing gas volume, and implementing a closed-loop control mechanism for real-time correction of temperature deviations, the stability and control accuracy of the blowing process under compact furnace conditions are ensured.

[0123] Comparative Example 1 This comparative example uses a 180t top-and-bottom blown converter to smelt plain carbon steel. The endpoint control targets are: carbon content 0.05%~0.08%, temperature 1620~1650℃, and oxygen activity 300~600ppm. The molten iron charge is 180t, and the initial composition is as follows: =4.5%, Si0=0.5%, =0.1%, equipment parameters are furnace volume ratio =0.85, oxygen lance orifice diameter 50mm. The basic conditions are consistent with those in Example 1.

[0124] Control methods used: Traditional converter control methods are used, without employing the carbon-oxygen balance inverse solution model, staged dynamic optimization, and real-time closed-loop control of this invention. The entire process relies on fixed process parameters combined with manual experience. Fixed oxygen supply flow rate: 38,000 m³ / h; Fixed oxygen lance position: 1.6 m; Fixed bottom blowing volume: 12 m³ / min; The amount of slag-forming agent added is set based on manual experience (1400 kg of lime and 500 kg of dolomite), without quantitative calculation. No data preprocessing or parameter optimization process; The final carbon temperature relies entirely on the operator's experience to judge, and is manually adjusted by observing the flame color at the furnace opening and taking samples for testing.

[0125] Implementation process: During the blowing process, the operator judges the endpoint based on experience. The first secondary gun test is conducted at the 12th minute, and the carbon content is measured to be 0.12% and the temperature to be 1610℃. Blowing continues until the 14th minute, when the second test is conducted, and the carbon content is measured to be 0.07% and the temperature to be 1630℃. The third test is conducted at the 15th minute, and the carbon content is measured to be 0.05% and the temperature to be 1640℃. The endpoint is then determined and blowing is stopped.

[0126] Implementation results:

[0127] Analysis and Conclusion: Traditional manual experience-based control methods have the following problems: The determination of the endpoint relies on the operator's experience, which is highly subjective and inconsistent. Fixed process parameters cannot adapt to the differences in carbon-oxygen reaction characteristics at different stages of blowing; The carbon temperature double hit rate is only 84%, the overblowing rate is as high as 6.8%, and the process stability is poor. Both the consumption of steel material and slag-forming agent are higher than those in the embodiments of the present invention; The re-drying and splashing phenomena that occurred during the blowing process pose safety hazards.

[0128] The overall performance indicators are far lower than those of the method of this invention.

[0129] Comparative Example 2 This comparative example uses a 180t top-and-bottom blown converter to smelt plain carbon steel. The endpoint control targets are: carbon content 0.05%~0.08%, temperature 1620~1650℃, and oxygen activity 300~600ppm. The molten iron charge is 180t, and the initial composition is as follows: =4.5%, =0.5% (fluctuating to 0.6%), equipment parameters are furnace capacity ratio =0.85, oxygen lance orifice diameter 50mm. The basic conditions are basically the same as in Example 1, but the silicon content of the molten iron fluctuates during the smelting process (from 0.5% to 0.6%).

[0130] Control methods used: The endpoint carbon temperature prediction is performed using only a single LSTM neural network, without implementing the carbon-oxygen balance inverse solution, staged dynamic optimization, and real-time closed-loop correction of this invention. The process parameters still use fixed curves. Fixed oxygen supply flow rate: 38,000 m³ / h; Fixed oxygen lance position: 1.6 m; Fixed bottom blowing volume: 12 m³ / min; Slag-forming agent addition amount: fixed curve (1400 kg lime, 500 kg dolomite); The final blowing time was adjusted only based on the LSTM prediction results, without any other optimization measures.

[0131] LSTM model configuration: Input characteristics: initial conditions of molten iron (C0, Si0, Mn0, T0), oxygen supply flow rate, oxygen lance position, and blowing time; Output: Endpoint carbon content, endpoint temperature; Training data: Historical production data from 500 furnaces; Network structure: 2-layer LSTM, 64 units per layer, fully connected layer output.

[0132] Implementation process: At the 8-minute mark, the LSTM model predicted a final carbon content of 0.065% and a final temperature of 1640℃ based on real-time data. When the silicon content in the molten iron fluctuated to 0.6%, the LSTM model failed to adapt to the changing operating conditions in a timely manner, and the prediction error gradually increased. By the 10-minute mark, the predicted carbon content error reached 0.02%, and the temperature error reached 8℃. An anomaly occurred in the mid-term carbon oxidation rate, with the measured value falling below the lower limit of 0.03% / min.

[0133] The operator stopped blowing at the 15th minute based on the LSTM prediction results, but the actual final carbon content was 0.09% (exceeding the target upper limit by 0.08%), and the actual final temperature was 1652℃ (exceeding the target upper limit by 1650℃).

[0134] Implementation results:

[0135] Analysis and Conclusion: The single LSTM prediction method has the following problems: Only the endpoint state prediction is achieved; the inverse solution of the carbon-oxygen balance is not realized. The process parameters still use fixed curves and cannot be dynamically adjusted based on the prediction results; When the composition of molten iron fluctuates, the prediction deviation increases significantly (temperature deviation reaches 8°C). It lacks multi-objective optimization capabilities and cannot balance efficiency, cost, and low carbon emissions. The carbon temperature double hit rate (87%) and overblowing rate (5.1%) are both lower than those of the embodiments of the present invention.

[0136] The control precision and stability are both inferior to the method of this invention.

[0137] Comparative Example 3 This comparative example uses a 180t top-and-bottom blown converter to smelt plain carbon steel. The endpoint control targets are: carbon content 0.05%~0.08%, temperature 1620~1650℃, and oxygen activity 300~600ppm. The molten iron charge is 180t, and the initial composition is as follows: =4.5%, =0.5% (fluctuating to 0.55%), equipment parameters are furnace capacity ratio =0.85, oxygen lance orifice diameter 50mm. The basic conditions are basically the same as in Example 1, but there are slight fluctuations in the silicon content of the molten iron.

[0138] Control methods used: The carbon-oxygen equilibrium inverse solution is performed using a pure mechanistic model, without incorporating the data-driven approach and PSO algorithm optimization of this invention, and without staged dynamic optimization. Fixed process parameters are calculated solely based on the carbon-oxygen reaction mechanism formula, without real-time closed-loop control.

[0139] Pure mechanistic model configuration: Based on the thermodynamic and kinetic equilibrium equations of the carbon-oxygen reaction, solve the following system of equations: Carbon-oxygen reaction equilibrium ; Material balance: Oxygen supply = Oxygen consumption from carbon oxidation + Oxygen consumption from iron oxidation + Oxygen consumption from the oxidation of other elements; Heat balance: Heat of chemical reaction = Heat carried away by furnace gas + Heat carried away by slag + Heat absorbed by molten steel + Heat loss.

[0140] Input the initial conditions and target endpoint parameters for molten iron, and output the fixed process parameters: Oxygen supply flow rate: 38200 m³ / h; Oxygen lance position: 1.62 m; Bottom blowing air volume: 11 m³ / min; Slag-forming agent addition: 1450 kg of lime and 520 kg of dolomite.

[0141] Implementation process: The pure mechanistic model outputs fixed process parameters based on theoretical calculations. During the blowing process, because the pure mechanistic model could not adapt to the slight fluctuations in the composition of the molten iron (SiO increased slightly from 0.5% to 0.55%), the inverse kinematic parameters deviated from the actual operating conditions. An abnormal carbon oxidation rate appeared in the middle stage, with the measured carbon oxidation rate dropping to 0.028% / min, below the reasonable lower limit of 0.03% / min, leading to a decrease in smelting efficiency. Operators attempted to manually intervene and adjust the oxygen supply flow rate, but due to a lack of quantitative guidance, the adjustment effect was unsatisfactory.

[0142] Implementation results:

[0143] Analysis and Conclusion: Pure mechanistic modeling methods have the following problems: (1) It cannot adapt to small fluctuations in the composition of molten iron, and the reverse solution parameters have large deviations; (2) Without data-driven correction, it is impossible to improve accuracy using historical data; (3) Without PSO algorithm optimization, the inverse solution result is prone to getting trapped in local optima; (4) Without phased dynamic optimization, it cannot adapt to the differences in carbon-oxygen reaction characteristics at each stage of blowing; (5) There is no real-time closed-loop control, and it cannot be corrected in time when the operating conditions fluctuate; (6) It is impossible to achieve multi-objective synergistic optimization of quality, efficiency, cost and low carbon.

[0144] The robustness is far lower than that of the mechanism-data hybrid driving mode of this invention.

[0145] A comprehensive comparison between Comparative Examples 1-3 and Examples 1-3 is shown below:

[0146] As can be seen from the table above, Embodiment 1 of the present invention is significantly superior to the three comparative examples in all key indicators, which fully demonstrates the technical advantages of the present invention in carbon-oxygen balance inverse solution, staged multi-constraint dynamic optimization and real-time closed-loop control.

Claims

1. A method for reverse solving of carbon-oxygen balance and optimization of process parameters in converter blowing, characterized in that, include: S1. Collect data from the entire converter blowing process and perform data preprocessing; the full process data includes initial hot metal conditions, real-time status of the molten pool, process parameters, and equipment status. S2. Construct a carbon-oxygen balance inverse solution model. The carbon-oxygen balance inverse solution model includes a mechanism constraint framework and a BP-PSO hybrid inverse solution model, and inversely solves the target endpoint parameters into the initial values ​​of process parameters. S3. Establish a phased multi-constraint dynamic parameter optimization model. Based on the carbon-oxygen reaction characteristics of each stage in the initial, middle and late stages of converter blowing, establish the corresponding constraints and objective functions for each stage. Based on the initial values ​​of process parameters obtained from the inverse solution, use the NSGA-III multi-objective optimization algorithm to perform phased iterative optimization and generate the optimal dynamic process parameter curves of each stage as time changes. S4. Establish a real-time feedback closed-loop control mechanism to dynamically trigger inverse solution recalculation and parameter optimization iteration based on the deviation between online detection data and the target. S5. The optimized process parameters are sent to the basic automation system in the form of digital instructions to control the automatic operation of oxygen lances, oxygen supply valves, bottom blowing valves and slagging agent feeding devices.

2. The method for reverse solving of carbon-oxygen balance and optimization of process parameters in converter blowing according to claim 1, characterized in that, In S1: The initial conditions of molten iron and the state of the equipment serve as the input boundaries of the carbon-oxygen balance inverse solution model. The real-time state of the molten pool is used for deviation calculation and correction triggering of real-time feedback closed-loop control. The process parameters are the output objects of the inverse solution and optimization. Specifically, this includes the following: S11, Initial conditions for molten iron: Initial carbon content Initial silicon content Initial manganese content Initial sulfur content Initial phosphorus content Initial temperature molten iron weight ; S12, Real-time status of the molten pool: Real-time carbon content Real-time oxygen activity Real-time temperature , furnace gas composition and Volume fraction; S13, Process Parameters: Oxygen supply flow rate Oxygen lance position Bottom blowing volume Timing and dosage of slag-forming agent addition ; S14, Equipment Status: Furnace volume ratio , Oxygen lance aperture Furnace age .

3. The method for reverse solving of carbon-oxygen balance and optimization of process parameters in converter blowing according to claim 1, characterized in that, In S2, the carbon-oxygen balance inverse solution model includes a mechanistic constraint framework and a BP-PSO hybrid inverse solution model. The mechanistic constraint framework is used to limit the input and output boundary conditions of the inverse solution model, and the BP-PSO hybrid inverse solution model performs inverse solution within the boundary conditions. S21. The mechanistic constraint framework includes thermodynamic constraints and kinetic constraints on the carbon-oxygen reaction: (1) Thermodynamic constraints on the carbon-oxygen reaction are used to limit the output boundary of the inverse solution model, and the expression is: ， ； in: The carbon content of the molten pool is expressed as % (%). Oxygen content in the molten pool, in % % It is carbon monoxide gas; The temperature of the molten pool is expressed in Kelvin (K). Standard Gibbs free energy, expressed in J / mol; 38.3 is the empirical thermodynamic coefficient for the carbon-oxygen reaction; (2) Carbon-oxygen reaction kinetic constraints, used to limit the rate range of the carbon-oxygen reaction, are expressed as: ； in: This represents the carbon oxidation rate, expressed as % / min. The reaction rate constant is positively correlated with temperature, and its value ranges from 0.02 to 0.08 min. -1 ; This refers to the partial pressure of CO, measured in atm. This is the equilibrium constant for the carbon-oxygen reaction, which varies with smelting temperature. S22, BP-PSO hybrid inverse solution model: using a BP neural network as the solver and employing the PSO algorithm to optimize the initial weights and thresholds of the BP neural network; A backpropagation (BP) neural network consists of an input layer, hidden layers, and an output layer. The input layer contains the target endpoint parameters; Hidden layers are used to perform nonlinear transformations on the input features; The output layer contains the initial values ​​of the process parameters; The activation function used is the Sigmoid function, and the output formula is as follows: ； in: This is the output value of the kth node in the output layer, i.e., the initial value of the process parameters; This is the input value of the i-th node in the input layer, i.e., the target endpoint parameter; The weights from the i-th node in the hidden layer to the k-th node in the output layer; This is the bias of the k-th node in the output layer; This is the Sigmoid activation function.

4. The method for reverse solving of carbon-oxygen balance and optimization of process parameters in converter blowing according to claim 3, characterized in that, The PSO algorithm is used to optimize the initial weights and threshold of the BP neural network. With bias The iterative formula is as follows: ； in: Let be the velocity of the i-th particle in the j-th dimension during the (t+1)-th iteration; Let be the velocity of the i-th particle in the j-th dimension during the t-th iteration; , For learning factors; , A random number within the interval [0,1]; Let be the individual historical optimal position of the i-th particle in the j-th dimension during the t-th iteration; Let be the globally optimal position of the group in the j-th dimension during the t-th iteration; For the first The current position of a particle in the j-th dimension during the t-th iteration; For inertia weights, a linear decreasing strategy is adopted, and the expression is: ； in: Let be the inertia weight at the t-th iteration; The initial inertia weight is set to 1.

2. The final inertia weight is set to 0.

8. This represents the maximum number of iterations, with a value of 100. This represents the current iteration number, with a value ranging from 1 to 100.

5. The method for reverse solving of carbon-oxygen balance and optimization of process parameters in converter blowing according to claim 1, characterized in that, In S3, the objective function of the NSGA-III multi-objective optimization algorithm is: ； in: Optimize function values ​​for multiple objectives; ~ These are the weighting coefficients, which are 0.3, 0.2, 0.2, 0.15, and 0.15 respectively. The endpoint carbon temperature double hit rate, with a value ranging from 0% to 100%; The blowing cycle is expressed in minutes and ranges from 10 to 16. This refers to the consumption of steel materials, expressed in kg / t, with a value range of 1050–1100. The value represents the consumption of slag-forming agent, expressed in kg / t, and ranges from 20 to 60. This represents CO2 emissions, expressed in tons of steel (t / t), with a range of 1.5 to 2.

9.

6. The method for reverse solving of carbon-oxygen balance and optimization of process parameters in converter blowing according to claim 1, characterized in that, In S3: The initial stage constraints include: heating rate of 80–100 ℃ / min, silicon oxidation rate ≥90%, manganese oxidation rate ≥70%, and slag basicity of 1.2–1.5; The mid-term constraints include: carbon oxidation rate of 0.03–0.06 % / min, slag foaming degree of 0.8–1.2, splashing probability ≤5%, and bottom blowing gas volume of 5–20 m³ / min; The constraints for the later stage include: carbon content deviation ±0.01%, temperature deviation ±5℃, final oxygen activity 300~600ppm, and overblowing rate ≤2%.

7. The method for reverse solving of carbon-oxygen balance and optimization of process parameters in converter blowing according to claim 1, characterized in that, In S4, the establishment of a real-time feedback closed-loop control mechanism includes: Real-time monitoring: Sampling frequency is 1 time / 30 seconds to obtain real-time carbon content. Real-time oxygen activity Real-time temperature ; Deviation calculation: ； ； ； in: The value range is for carbon content deviation. ; Real-time carbon content, in % % The target carbon content is expressed as % (%). For temperature deviation, the value range is: ; This is the real-time temperature, in °C. The target temperature is expressed in °C. This is due to oxygen activity deviation; Real-time oxygen activity; Target oxygen activity; Deviation judgment: If the deviation is within the threshold, keep the current process parameters unchanged; If the deviation exceeds the threshold, the feedback correction process will begin. Feedback and corrections: The deviation signal is fed back to the carbon-oxygen balance inverse solution model and the staged optimization model to recalculate and output the corrected process parameters.

8. The method for reverse solving of carbon-oxygen balance and optimization of process parameters in converter blowing according to claim 1, characterized in that, In S4, the dynamic triggering of inverse solution recalculation and parameter optimization iteration includes: Real-time monitoring of molten pool status parameters and calculation of deviations from target values; When any deviation exceeds the set threshold, the inverse solution recalculation signal is triggered; Using the current molten pool state parameters as new input conditions, re-execute the inverse carbon-oxygen balance model in S2 to update the initial values ​​of process parameters; Using the updated process parameters as the initial population, the phased multi-constraint dynamic parameter optimization model in S3 is re-executed, and the corrected optimal process parameter curve is output. The revised process parameters are then sent to the basic automation system for execution.

Images