An adaptive hybrid intelligent prediction method and system for arc furnace smelting endpoint parameters

Abstract

The present application belongs to the field of intelligent control and prediction technology of metallurgical industry process, and discloses a kind of electric arc furnace smelting end point parameter adaptive hybrid intelligent prediction method and system. Obtain electric arc furnace smelting process data;Build a double-driven prediction system containing mechanism model and data-driven model;Computer mechanism model prediction value and actual measurement value Determination coefficient, statistical effective historical data amount;Build a machine learning prediction model, and select modeling algorithm according to effective historical data amount;Based on determination coefficient and effective historical data amount, select hybrid prediction strategy;According to the hybrid prediction strategy, predict the output end point carbon content prediction value and end point temperature prediction value;Carry out prediction deviation threshold value judgment and execute control, adjust electric arc electric power, oxygen flow, feeding speed or cooling water flow.The present application realizes the accurate prediction and dynamic optimization control of electric arc furnace end point parameters, significantly improves the smelting quality, energy utilization rate and production stability.

Description

Technical Field

[0001] This invention relates to the field of intelligent control and prediction technology for metallurgical industrial processes, specifically to an adaptive hybrid intelligent prediction method and system for endpoint parameters of electric arc furnace smelting. Background Technology

[0002] Electric arc furnaces are crucial smelting equipment in modern steel production, offering advantages such as high flexibility, short production cycles, and the ability to utilize scrap steel resources. They are widely used in the production of high-quality and specialty steels. During electric arc furnace smelting, the final carbon content and final temperature are key control parameters determining whether the steel composition and temperature meet the tapping requirements. Excessive or insufficient final carbon content will affect the accuracy of the steel composition and the deoxidation and desulfurization effects; while deviations from the target final temperature may lead to overheating of the steel, accelerated furnace lining erosion, or incomplete melting, resulting in increased energy consumption and production instability. Therefore, achieving accurate prediction and intelligent control of electric arc furnace final parameters is a vital issue in the optimization and control of the smelting process.

[0003] Currently, the prediction of electric arc furnace endpoint parameters is mainly divided into two categories: mechanistic modeling and data-driven methods. Mechanistic modeling is based on metallurgical principles such as energy balance and material balance, and can reflect the physicochemical laws of the smelting process, possessing a certain degree of interpretability and universality. However, these models rely on numerous simplifying assumptions, making it difficult to accurately describe complex heat transfer, mass transfer, and redox reaction processes. They also exhibit poor adaptability to changes in operating conditions and furnace fluctuations, easily leading to systematic biases. Data-driven methods utilize historical production data and achieve empirical predictions through machine learning or deep learning models, demonstrating good nonlinear modeling and self-learning capabilities. However, these methods are highly dependent on the amount and quality of data, lack physical constraint mechanisms, and are prone to overfitting or prediction instability when samples are insufficient or operating conditions drift. Furthermore, the electric arc furnace production process is characterized by strong nonlinearity, multivariate coupling, and frequent operating condition fluctuations, making it difficult for a single mechanistic model or data model to achieve high-precision predictions under all operating conditions. Some studies have attempted to use a simple weighted or cascaded fusion of mechanistic models and data models, but they lack data quality assessment and strategy adaptation mechanisms, resulting in unstable prediction performance under different data scales and mechanistic model accuracy conditions, making it difficult to achieve model adaptation and online updates.

[0004] In summary, the existing electric arc furnace endpoint parameter prediction technology still has the following shortcomings: (1) the mechanism model has structural simplification errors, making it difficult to capture hidden influencing factors; (2) the data-driven model relies on a large number of high-quality samples and has limited generalization ability; (3) there is a lack of dynamic strategy adjustment mechanism for data scale and model accuracy; and (4) the prediction results and execution control have not been able to achieve closed-loop linkage.

[0005] Therefore, there is an urgent need to propose an adaptive hybrid intelligent prediction method and system for the endpoint parameters of electric arc furnace smelting that takes into account both the interpretability of the mechanism and the adaptability of data-driven processes, and has the ability to adaptively integrate multiple strategies and achieve closed-loop control, so as to improve the intelligence level of the smelting process and the accuracy of endpoint control. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention proposes an adaptive hybrid intelligent prediction method and system for the endpoint parameters of electric arc furnace smelting. It combines mechanistic models and data-driven models for dual-drive prediction, enabling high-precision intelligent prediction and adaptive control of the endpoint carbon content and endpoint temperature of the electric arc furnace.

[0007] This invention aims to propose solutions to the following problems existing in the endpoint control of electric arc furnaces: 1. Mechanism models are limited by idealized assumptions, making it difficult to accurately describe complex heat and mass transfer processes and nonlinear characteristics of equipment; 2. Pure data-driven models rely on large sample training, are highly sensitive to data quality and operating condition stability, and have limited generalization ability; 3. Fluctuations in raw materials and differences in operation between different smelting batches lead to large prediction errors in endpoint temperature and endpoint carbon content; 4. Existing prediction systems lack adaptive mechanisms and closed-loop control capabilities, and cannot dynamically adjust the prediction method according to data quality and mechanism accuracy.

[0008] The technical solution of the present invention is as follows: An adaptive hybrid intelligent prediction method for the endpoint parameters of electric arc furnace smelting, comprising the following steps:

[0009] S1. Data Acquisition: Acquire data on the electric arc furnace smelting process, including raw material composition, raw material addition amount, auxiliary material composition, auxiliary material addition amount, power supply system parameters, gas supply system parameters, water cooling system parameters, and smelting time.

[0010] S2. Construct a dual-drive prediction system that includes both mechanistic models and data-driven models;

[0011] S3. Conduct a comprehensive data quality assessment, including: validating the mechanistic model using historical production data, and calculating the coefficient of determination between the model's predicted values ​​and actual measured values. Statistical volume of valid historical data ;

[0012] S4. Construct a machine learning prediction model based on the amount of effective historical data. Select a modeling algorithm:

[0013] S5, Based on the coefficient of determination and effective historical data volume Select a hybrid forecasting strategy; based on the selected hybrid forecasting strategy, predict the final carbon content and final temperature values.

[0014] S6. Based on the output of S5, perform prediction deviation threshold judgment and execution control, adjust the electric arc power, oxygen blowing flow rate, feeding speed or cooling water flow rate to achieve convergence of the endpoint parameters to the target value.

[0015] The dual-drive prediction system includes:

[0016] A mechanistic model for the endpoint carbon content is constructed based on the principle of material balance. The endpoint carbon content is calculated according to the following equation:

[0017]

[0018] In the formula: For the first The quality of carbon-containing raw materials, For the first The percentage of carbon content in a carbon-containing raw material. This represents the number of types of carbon-containing raw materials; For the first Material quality for each carbon loss pathway For the first The percentage of carbon content in each carbon loss pathway The number of carbon loss pathways; The endpoint is the quality of molten steel; among which, carbon-containing raw materials include scrap steel, molten iron, direct reduced iron, scrap iron, carbon powder, and carbon raisers; carbon loss pathways include oxidative decarburization, dust carry-over, and slag absorption;

[0019] A mechanistic model for the endpoint temperature is constructed based on the principle of energy balance. The endpoint temperature is calculated according to the following equation:

[0020]

[0021] In the formula: The initial temperature; For total input energy, The energy required for the material to heat up and undergo phase change. Energy loss through various pathways; The average specific heat capacity of molten steel; where, This includes energy from electric arc heating, heat from chemical reactions, heat from fuel combustion, and energy from waste preheating. This includes the latent heat of fusion of the material, the sensible heat of the material's temperature rise, and the heat effect of the slag-forming reaction; This includes heat loss from the furnace bottom and top, heat loss carried away by flue gas, heat loss carried away by cooling water, and electrode radiation loss.

[0022] Construct a data-driven machine learning prediction model, which is trained based on historical production data.

[0023] The specific calculation formula for the mechanism model of the endpoint carbon content is as follows:

[0024]

[0025] In the formula: , , , These are the quality of scrap steel, pig iron, carbon raiser, and recycled materials, respectively. , , , These represent the percentage of carbon content in scrap steel, pig iron, carbon refining agent, and recycled materials, respectively. For oxygen blowing volume, Oxygen density, For decarbonization efficiency, The molar mass ratio of carbon to oxygen; For the quality of smoke and dust, The percentage of carbon content in the smoke and dust; For the quality of slag, This represents the percentage of carbon content in the slag.

[0026] The specific calculation formula for the mechanism model of the endpoint temperature is as follows:

[0027]

[0028] In the formula: The initial temperature. For power conversion efficiency, For input electrical power, For smelting time; This represents the total heat released by the oxidation reaction; Heat of fuel combustion To assist in heating; For the first The latent heat of fusion of a molten material This represents the number of types of materials being melted. For the first The quality of the heated material For the first The specific heat capacity of the heated material For the first Temperature rise of the heated material This represents the number of types of materials being heated; This is due to the heat effect of the slag-forming reaction; For the furnace wall area, The heat transfer coefficient of the furnace wall, The average temperature inside the furnace. The ambient temperature; For flue gas quality, For the specific heat capacity of flue gas, The flue gas temperature; The cooling water removes heat. Emissivity of the electrode surface The Stefan-Boltzmann constant is... The electrode surface area. This represents the electrode temperature.

[0029] The hybrid prediction strategy is specifically as follows:

[0030] when and When this happens, feature enhancement strategy A is activated;

[0031] when and When this occurs, error compensation strategy B is activated;

[0032] when and When this happens, activate dynamic weighting strategy C;

[0033] when When this happens, the pure data-driven strategy D is enabled.

[0034] The feature enhancement strategy A is as follows: the mechanism model outputs the mechanism prediction value; the mechanism prediction value is used as a derived feature and merged with the original production data feature to form an enhanced feature set, which is used to train and build a machine learning prediction model; the prediction value of the machine learning prediction model trained under the enhanced feature set is used as the final prediction value.

[0035] A1: The mechanistic prediction value of the endpoint carbon content was obtained by calculating using a mechanistic model. and endpoint temperature mechanism prediction value ;

[0036] A2: Will and As a derived feature, it is related to the features of the original production data. Merging to form an enhanced feature set ;

[0037] A3: Input the enhanced feature set into the machine learning prediction model, and output the final predicted value of the endpoint carbon content. and final predicted value of endpoint temperature ;

[0038] The error compensation strategy B is as follows: use the mechanism model to output the mechanism prediction value; use the machine learning prediction model to learn the systematic bias of the mechanism model and output the error compensation value; and add the mechanism prediction value and the error compensation value to obtain the final prediction value.

[0039] B1: Mechanistic prediction of endpoint carbon content obtained using a mechanistic model. and endpoint temperature mechanism prediction value ;

[0040] B2: Construct a machine learning prediction model to predict the error of the mechanism model based on the characteristics of the original production data, and output the predicted value of carbon content error. and temperature error prediction value ;

[0041] B3: Calculate the final predicted values ​​of the endpoint parameters:

[0042]

[0043] ;

[0044] The dynamic weighting strategy C: uses a mechanistic model to output mechanistic prediction values; trains and constructs a machine learning prediction model using the features of the original production data; and divides the historical production data into clusters according to a clustering algorithm. Each subgroup has a weighted computer model for each subgroup. and machine learning prediction model weights ,in Determine the cluster to which the new sample belongs. Endpoint parameter prediction value Mechanism model predictions Machine learning model predictions;

[0045] C1: Use the K-means clustering algorithm or the DBSCAN clustering algorithm to divide the historical production data into... Subgroup;

[0046] C2: For each subgroup The optimal weights for the mechanistic model and the machine learning model are determined through grid search or Bayesian optimization. The optimization objective is to minimize the root mean square error (RMSE) of the validation set, with the following constraints: and ;

[0047] C3: For each sample to be predicted, calculate its Euclidean distance to the center of each cluster and determine its cluster affiliation. ;

[0048] C4: Calculate the final predicted values ​​of the endpoint parameters:

[0049]

[0050]

[0051] In the formula: The carbon content value independently predicted by the machine learning prediction model. Temperature values ​​independently predicted by machine learning prediction models;

[0052] Pure data-driven strategy D: Train and build a machine learning prediction model using the features of the original production data; use the prediction value of the machine learning prediction model as the final prediction value.

[0053] The selection criteria for the modeling algorithm are as follows:

[0054] when At that time, the support vector regression algorithm was used;

[0055] when At that time, a few-sample learning algorithm based on tabular data was adopted;

[0056] when At that time, the gradient boosting decision tree algorithm is used;

[0057] when At that time, a deep neural network algorithm was used;

[0058] When using a deep neural network algorithm, the network structure includes an input layer, at least two hidden layers, and an output layer. The number of nodes in the hidden layers decreases sequentially. The activation function of the hidden layers is the ReLU function, and the output layer uses a linear activation function. The loss function is the mean squared error (MSE), and the optimizer is the Adam optimizer. The training process uses early stopping to prevent overfitting. Training stops when the validation set loss does not improve for 500 consecutive epochs. The total number of iterations is at least 2000.

[0059] Specifically, S6 is:

[0060] Calculate the prediction bias:

[0061]

[0062] For the predicted values ​​of machine learning prediction models, The target value for the machine learning prediction model;

[0063] Calculate threshold :

[0064]

[0065] in, To account for the allowable deviation in the process, To predict uncertainty, Let z be the standard deviation of process fluctuations, and z be the confidence coefficient.

[0066] when When this happens, execution control is triggered;

[0067] Calculate the optimal adjustment scheme:

[0068]

[0069] The difference between the target output and the current predicted output. For sensitivity matrix, For the adjustment vector of control variables, It is the regularization coefficient;

[0070] constraint:

[0071]

[0072] It is the lower limit of the control quantity. It is the current control quantity, It is the upper limit of the control quantity. It controls the lower limit of the increment. It controls the upper limit of the increment;

[0073] The obtained The data is sent to the execution end to adjust the electric arc power, oxygen flow rate, feeding speed, or cooling water flow rate, so as to achieve convergence of the endpoint parameters to the target value.

[0074] An adaptive hybrid intelligent prediction system for endpoint parameters in electric arc furnace smelting includes:

[0075] The data acquisition module is used to collect data on the electric arc furnace smelting process in real time, including raw material composition, raw material addition amount, auxiliary material composition, auxiliary material addition amount, power supply system parameters, gas supply system parameters, water cooling system parameters, and smelting time.

[0076] The mechanism calculation module includes a carbon content mechanism calculation unit and a temperature mechanism calculation unit; the carbon content mechanism calculation unit calculates the endpoint carbon content mechanism prediction value based on the material balance principle, and the temperature mechanism calculation unit calculates the endpoint temperature mechanism prediction value based on the energy balance principle.

[0077] The data modeling module is used to train machine learning prediction models based on historical production data and adaptively select modeling algorithms according to the data scale.

[0078] The quality assessment module is used to evaluate the accuracy of computer model predictions. Statistical historical data volume N;

[0079] The strategy decision-making module is used to predict accuracy based on the mechanistic model. The following strategies are used to adaptively select feature enhancement strategies based on the amount of historical data N: A) error compensation strategy, B) dynamic weighting strategy, C) or pure data-driven strategy.

[0080] The layered fusion module includes:

[0081] The feature enhancement execution unit is used to implement feature enhancement strategy A, which combines the mechanism prediction value as a derived feature with the original production data feature and inputs it into the machine learning prediction model.

[0082] The error compensation execution unit is used to implement error compensation strategy B, which learns and corrects the systematic bias of the mechanism model through machine learning prediction model;

[0083] The dynamic weighted execution unit includes a working condition identification subunit, a weight query subunit, and a weighted fusion subunit; the working condition identification subunit determines the working condition category of the sample to be predicted, the weight query subunit retrieves the optimal fusion weight for the corresponding working condition, and the weighted fusion subunit calculates the weighted average of the predicted values ​​from the computer model and the machine learning prediction model.

[0084] A pure data-driven execution unit that uses only machine learning prediction models for prediction;

[0085] The results output module is used to output the predicted values ​​of the endpoint carbon content and the endpoint temperature, and provides a prediction confidence assessment.

[0086] The model update module is used to update the machine learning prediction model online using the production data of the new furnace, and to re-evaluate the prediction accuracy of the mechanistic model and the amount of historical statistical data, and dynamically adjust the hybrid prediction strategy.

[0087] The execution control module receives the deviation between the predicted result and the target value, calculates the threshold, and determines the triggering condition. When |Δ| exceeds the threshold, the execution control module calculates the optimal control adjustment amount Δu based on the sensitivity matrix or optimization algorithm, and outputs it to the PLC system or DCS system after safety limiting and smoothing. The execution control module controls at least one of the following actuators: arc power, oxygen blowing flow rate, feeding speed, and cooling water flow rate, to realize automatic adjustment and closed-loop correction of the smelting process.

[0088] The execution control module includes:

[0089] The deviation monitoring subunit is used to calculate the predicted deviation Δ and determine the threshold in real time.

[0090] The optimized solution sub-unit is used to solve for the optimal adjustment scheme Δu;

[0091] The safety limiting subunit is used to perform limiting, rate limiting, and smoothing.

[0092] The instruction issuing subunit is used to send control instructions to the electric furnace control system through the industrial communication interface;

[0093] The feedback monitoring subunit is used to collect post-execution operating parameters and update the model to achieve closed-loop control.

[0094] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0095] Integrating complementary advantages: Combining physical mechanism models with data-driven models, taking into account both theoretical consistency and flexible adaptability;

[0096] Adaptive prediction strategy: Dynamically select the optimal fusion method based on the accuracy of the mechanism model and the data scale, significantly improving the reliability of prediction;

[0097] Algorithm self-selection mechanism: Automatically matches algorithm type based on data volume to achieve intelligent evolution throughout the entire life cycle;

[0098] Closed-loop self-tuning control: The prediction results are linked with the execution control, and it has the functions of deviation identification, optimal adjustment and feedback update to realize automatic process adjustment;

[0099] Online learning capability: The system can update itself over time and with the accumulation of data, adapting to fluctuations in raw material composition and changes in operating conditions;

[0100] Highly applicable to industrial applications: The modular architecture facilitates its adoption on electric arc furnace production lines of varying sizes and supports integration with MES and DCS systems.

[0101] In summary, the adaptive hybrid intelligent prediction method and system for electric arc furnace smelting endpoint parameters proposed in this invention constructs an integrated intelligent system of "mechanism-data fusion-control closed loop," realizing accurate prediction and dynamic optimization control of electric arc furnace endpoint parameters, significantly improving smelting quality, energy utilization rate, and production stability, and has broad prospects for industrial application. Attached Figure Description

[0102] Figure 1 A schematic diagram of an adaptive hybrid intelligent prediction method for end-point parameters in electric arc furnace smelting;

[0103] Figure 2 This is a schematic diagram of an adaptive hybrid intelligent prediction system for the end-point parameters of electric arc furnace smelting. Detailed Implementation

[0104] The adaptive hybrid intelligent prediction method for the endpoint parameters of electric arc furnace smelting proposed in this invention integrates an adaptive hybrid prediction system that combines a mechanistic model with data-driven learning. Through the synergistic effect of the interpretability of the mechanistic model and the nonlinear expressive ability of the machine learning prediction model, it achieves high-precision and robust prediction of the endpoint parameters of electric arc furnace smelting. Combined with adaptive control of prediction deviation, it achieves precise control of the endpoint composition and temperature.

[0105] The data acquisition system collects multi-source data on the electric arc furnace smelting process in real time, covering the following dimensions:

[0106] Raw material information: composition and amount of scrap steel, molten iron, direct reduced iron, carbon raiser, auxiliary materials, etc.;

[0107] Operating parameters: power supply voltage, current, power factor, oxygen blowing flow rate, furnace gas composition, water cooling system temperature and flow rate, etc.

[0108] Smelting process parameters: smelting time, furnace temperature change, flue gas temperature, slag temperature, furnace top temperature, etc.

[0109] All data, after undergoing time-series synchronization, outlier removal, and missing value imputation, serve as the input basis for subsequent mechanism and data-driven models.

[0110] A carbon content model at the endpoint is established based on the principle of material balance. Taking into account carbon input, carbon loss and molten steel mass, a calculable energy and material conservation equation is formed to reflect the dynamic changes of carbon elements in the smelting process.

[0111] A final temperature model is established based on the principle of energy balance, incorporating input energy, melting energy requirements, and heat loss through multiple pathways into the calculation, thereby achieving quantitative simulation of the smelting thermal process.

[0112] Both types of mechanistic models can be customized and corrected according to process parameters, and have good interpretability and universality.

[0113] Machine learning algorithms (including support vector regression, gradient boosting decision trees, deep neural networks, etc.) are used to learn implicit relationships from historical production data that cannot be explicitly described by mechanistic models, such as operating habits, raw material fluctuations, and instrument delays.

[0114] This model obtains a nonlinear mapping relationship between input features and endpoint parameters through training, providing compensation or fusion information for hybrid prediction.

[0115] Mechanistic models provide physical consistency constraints, while data models offer flexible nonlinear expressive capabilities. The two work together in a hierarchical fusion manner to achieve a balance between prediction accuracy and robustness.

[0116] R², the accuracy index of a mechanistic model, measures the goodness of fit between the predicted results of a mechanistic model and the actual measured values.

[0117] Data sufficiency index N: The number of samples with available historical data.

[0118] The system dynamically selects from four strategies based on the evaluation results of R² and N:

[0119] 1. Feature enhancement strategy A: When the mechanism model has high accuracy and sufficient data, the mechanism output is used as a derived feature to enhance the feature space of the machine learning model;

[0120] 2. Error Compensation Strategy B: When the mechanistic model has high accuracy but the data is limited, the systematic error of the mechanistic model learned by the machine learning model is corrected.

[0121] 3. Dynamic weighting strategy C: When both the mechanism model and the data model are reliable, different working condition categories are identified by clustering and weighting coefficients are calculated to achieve adaptive weighted fusion of working conditions;

[0122] 4. Pure data-driven strategy D: When the reliability of the mechanistic model is low, prediction is made solely by relying on the data-driven model.

[0123] This adaptive selection mechanism ensures optimal prediction performance under any data and operating conditions.

[0124] The optimal modeling algorithm is automatically matched based on the amount of historical data N: This algorithm selection mechanism realizes the adaptive intelligence of the data-driven part and avoids model degradation caused by human intervention.

[0125] The system outputs the predicted values ​​of carbon content and temperature at the endpoint, and performs bias and confidence level assessments.

[0126] By calculating the prediction bias The system determines the adaptive threshold based on the predicted uncertainty. When the deviation exceeds the threshold, the system automatically triggers the control mechanism, uses a constraint optimization algorithm to calculate the optimal adjustment amount Δu, and performs closed-loop regulation on the execution quantities such as arc power, oxygen blowing flow rate, feeding speed, or cooling water flow rate, so that the endpoint parameters converge to the target value.

[0127] After accumulating a certain number of production data (e.g., 2000 heats), the system automatically performs model retraining and performance evaluation, including: mechanism model parameter correction; data distribution drift detection and feature re-screening; and hybrid strategy weight update.

[0128] This enables the prediction system to continuously self-optimize and accumulate knowledge, allowing the model to maintain high accuracy and stability over a long period. Specific implementation method one:

[0130] This embodiment provides an adaptive hybrid intelligent prediction method for endpoint parameters in electric arc furnace smelting. This method can adaptively select the optimal mechanism-data hybrid prediction strategy based on data quality, overcoming the limitations of traditional single prediction methods and significantly improving the prediction accuracy of endpoint carbon content and temperature.

[0131] like Figure 1 As shown, the prediction method in this embodiment includes the following steps:

[0132] Step S1: Data Acquisition. Real-time acquisition of smelting process data is performed from the electric arc furnace distributed control system (DCS) and production execution system (MES), specifically including:

[0133] (1) Raw material composition data: carbon content of scrap steel =0.15%-0.45%, pig iron carbon content =3.5%-4.5%, carbon content of carbon raiser =85%-98%, carbon content of recycled materials =0.05%-0.8%, and the content of elements such as silicon, manganese, phosphorus, and sulfur in each raw material;

[0134] (2) Raw material addition data: scrap steel quality =80-120 tons, pig iron quality =0-40 tons, carbon raiser quality =0-500 kg, remelting material quality =0-10 tons;

[0135] (3) Auxiliary material composition and addition amount: mass of lime and CaO content, mass of dolomite and MgO content, mass of fluorite, mass of lightly calcined magnesia balls;

[0136] (4) Power supply system parameters: input power, smelting time, cumulative power consumption, current, voltage, power factor;

[0137] (5) Gas supply system parameters: oxygen blowing volume, oxygen flow rate, nitrogen flow rate, argon flow rate;

[0138] (6) Water cooling system parameters: cooling water flow rate, inlet water temperature, outlet water temperature, and heat carried away by the cooling water Q_cooling water;

[0139] (7) Other process parameters: number of times the furnace door is opened, foam slag retention time, ambient temperature T_ambient = 15-35℃.

[0140] All collected data underwent outlier detection and missing value processing, eliminating obviously erroneous data points, and processing missing values ​​using forward imputation or mean interpolation methods.

[0141] Step S2: Construction of the Dual-Driven Prediction Model. Construct a dual-drive prediction system comprising a mechanistic model and a data-driven model:

[0142] (1) Mechanism model for endpoint carbon content. Based on the principle of carbon element material balance, an endpoint carbon content calculation model is established. The input sources of carbon include various carbon-containing raw materials, and the output of carbon includes three pathways: oxidation decarbonization, dust removal, and slag absorption.

[0143] The general formula for calculating the endpoint carbon content is: ;

[0144] Where: n is the number of carbon-containing raw material types, in this embodiment n=4 (scrap steel, pig iron, carbon raiser, recycled material); p is the number of carbon loss pathways, in this embodiment p=3 (oxidation decarburization, flue dust, slag).

[0145] (2) Mechanism model of endpoint temperature;

[0146] Based on the principle of energy balance, a calculation model for the endpoint temperature is established. The overall energy balance equation of the system is: Input energy = Demand energy + Loss energy.

[0147] The general formula for calculating the endpoint temperature is: ;

[0148] in: The initial temperature of the scrap steel is usually taken as the ambient temperature of 15-35℃. The average specific heat capacity of molten steel is taken as 0.67-0.75 kJ / (kg·K).

[0149] Calculation methods for each energy term:

[0150] Input energy item: Arc heating energy: Among them, the power conversion efficiency Oxidation reaction is exothermic. The oxidation efficiency of each element Determined based on actual operating conditions; Heat of fuel combustion: If natural gas or coal gas is used for auxiliary heating; auxiliary heating: This includes waste heat recovery from exhaust gases, etc.

[0151] Energy demand item: Latent heat of fusion of material: Latent heat of molten scrap steel Latent heat of molten pig iron Sensible heat of material heating: Solid specific heat capacity Specific heat capacity of liquid ; Thermal effect of slag formation reaction: The combined effect of the endothermic decomposition of lime and the exothermic reaction of slag formation.

[0152] Energy loss item: Heat dissipation from furnace wall: furnace wall area heat transfer coefficient Average temperature inside the furnace Flue gas carries away heat: flue gas temperature flue gas specific heat capacity Cooling water removes heat: Electrode radiation loss: Electrode emissivity Stefan-Boltzmann constant , electrode temperature .

[0153] (3) Data-driven machine learning prediction model;

[0154] Build a machine learning model based on historical production data to learn implicit influencing factors that cannot be accurately described by the mechanism model, such as the lining erosion state, scrap quality fluctuations, operator habits, etc.

[0155] The input features include: original process parameters (35 - 50 dimensions), statistical features (such as feeding rate, power curve slope, etc., 10 - 15 dimensions), and time series features (such as the end-point parameters of the previous 5 furnace runs, 10 dimensions).

[0156] The output targets are the true measured values of the end-point carbon content and the end-point temperature.

[0157] Step S3: Comprehensive evaluation of data quality.

[0158] (1) Evaluation of the accuracy of the mechanism model;

[0159] Extract the production data of the most recent 6 months from the historical database as the validation set, use the mechanism model to predict all furnace runs, and calculate the coefficient of determination R² between the predicted value and the actual measured value:

[0160]

[0161] Where: is the actual measured end-point parameter, is the predicted value of the mechanism model, is the average of the actual values.

[0162] Calculate the end-point carbon content and of the end-point temperature respectively, and take the smaller value of the two as the overall accuracy index R² of the mechanism model.

[0163] The physical meaning of R²: R² > 0.7: The mechanism model has high prediction accuracy and can explain more than 70% of the variation; 0.3 < R² ≤ 0.7: The mechanism model has certain prediction ability but there are large systematic biases; R² ≤ 0.3: The mechanism model has weak prediction ability and the data-driven method is more reliable;

[0164] [[ID=四十九]](2) Evaluation of data sufficiency; the number N of available historical smelting furnace runs in the statistical database. Data quality requirements:

[0165] Data integrity: missing rate < 5%; data consistency: parameter fluctuation coefficient < 15% under the same working conditions; data timeliness: preferentially use data from the recent 12 months;

[0166] The impact of data volume N: N < 1000: small samples, complex models are prone to overfitting; 1000 ≤ N < 10000: medium samples, suitable for tree models; 10000 ≤ N < 20000: large samples, ensemble learning can be used; N ≥ 20000: large samples, deep neural networks can be trained.

[0167] In this embodiment, a steel plant accumulated historical data from 25,000 heats, and the mechanistic model performed well on the validation set. , ,Pick .

[0168] Step S4: Adaptively select a hybrid prediction strategy;

[0169] Based on the evaluation results of step S3 (R²=0.58, N=25000), the prediction strategy is selected according to the following decision tree: Decision rule: When R²>0.3 and N>20000, feature enhancement strategy A is enabled; when R²>0.3 and N≤20000, error compensation strategy B is enabled; when R²>0.7 and N>20000, dynamic weighting strategy C is enabled; when R²≤0.3, pure data-driven strategy D is enabled.

[0170] This embodiment satisfies condition 1 (0.58>0.3 and 25000>20000), therefore feature enhancement strategy A is selected.

[0171] Design principles of strategy selection:

[0172] Feature enhancement strategy A: When the mechanism model has a certain accuracy and there is sufficient data, the mechanism knowledge is used as the data-driven model for feature enhancement.

[0173] Error compensation strategy B: When the accuracy of the mechanistic model is generally poor but the data is insufficient, machine learning is used to compensate for the systematic bias of the mechanistic model.

[0174] Dynamic weighting strategy C: When the mechanism model is accurate and the data is sufficient, the fusion weight of the mechanism and data is dynamically adjusted for different working conditions;

[0175] Pure data-driven strategy D: When the accuracy of the mechanistic model is too low, it relies entirely on data-driven methods.

[0176] Step S5: Build a machine learning model;

[0177] Based on the historical data volume N=25000, and following the adaptive algorithm selection rules, a deep neural network algorithm is adopted.

[0178] (1) Network structure design;

[0179] Input layer: 55 neurons (35 raw features + 15 statistical features + 5 temporal features);

[0180] First hidden layer: 128 neurons, ReLU activation function; Second hidden layer: 64 neurons, ReLU activation function; Third hidden layer: 32 neurons, ReLU activation function;

[0181] Output layer: 2 neurons (carbon content, temperature), linear activation function;

[0182] The number of hidden layer nodes decreases in powers of 2, which is beneficial for the layer-by-layer abstraction of features.

[0183] (2) Training configuration loss function: mean squared error ;

[0184] Optimizer: Adam optimizer, initial learning rate 0.001, using a learning rate decay strategy; batch size: 64; total iterations: 3000 epochs; early stopping strategy: training stops when the validation set loss has not improved for 500 consecutive epochs to prevent overfitting. Regularization: L2 regularization, regularization coefficient λ=0.0001; data augmentation: 5% Gaussian noise is added to the input features to enhance model robustness.

[0185] (3) Dataset partitioning: Training set: 20,000 batches (80%);

[0186] Validation set: 2500 heats (10%);

[0187] Test set: 2500 heats (10%). Stratified sampling was used to ensure that the proportion of each operating condition type was consistent across the three datasets.

[0188] (4) Applicable Scenarios for Other Algorithms: Support Vector Regression (SVR): Used when N < 1000. The RBF kernel function is employed, and the penalty parameter C and kernel parameter are determined through cross-validation. SVR exhibits good generalization ability with small sample sizes.

[0189] TabNet (a few-shot learning algorithm for tabular data): Used when 1000 ≤ N < 10000. TabNet uses an attention mechanism for feature selection, which can automatically identify important features and is suitable for tabular data with a medium sample size.

[0190] Gradient Boosting Decision Trees (GBDT / XGBoost / LightGBM): Used when 10000≤N<20000. Tree models are insensitive to data distribution, require no complex feature engineering, have fast training speed, and perform well with medium to large sample sizes.

[0191] Using the mechanistic model established in step S2, the process parameters for the current furnace batch are calculated:

[0192] Assume the parameters for the current heat are: Scrap steel: , ;pig iron: , Carbon raiser: , ; Recycled materials: , Oxygen blowing volume: Input power: Smelting time: Other parameters were calculated by substituting typical values ​​into the carbon content mechanism model:

[0193] ;

[0194] Substitute into the temperature mechanism model for calculation (detailed calculations of each energy term omitted): ;

[0195] Model output: , ;

[0196] Step S7: Output the predicted values ​​of the endpoint parameters;

[0197] The system outputs the predicted endpoint parameters for this furnace cycle: Predicted endpoint carbon content: Predicted final temperature: ;

[0198] Predicted confidence intervals: carbon content [0.38%, 0.46%], temperature [1628℃, 1648℃];

[0199] Prediction confidence level: Based on the Bootstrap method, the prediction results with a confidence level of 92% are displayed in real time through the human-machine interface and sent to the upper control system.

[0200] Step S8: Prediction deviation threshold judgment and execution control;

[0201] (1) Calculate the prediction deviation;

[0202] Assume the target endpoint parameter for this furnace run is the target carbon content: Target temperature:

[0203] Prediction bias: (Slightly high); (Too low);

[0204] (2) Threshold calculation;

[0205] Carbon content threshold: allowable process deviation Prediction uncertainty Standard deviation of process fluctuation Confidence coefficient ;

[0206]

[0207]

[0208] Temperature threshold: Allowable deviation in process Prediction uncertainty Standard deviation of process fluctuation Confidence coefficient ;

[0209]

[0210] ;

[0211] (3) Trigger condition judgment;

[0212] Carbon content: No control was triggered.

[0213] temperature: No control was triggered.

[0214] The predicted deviations for this furnace run were all within the allowable range, so the system did not implement active control and continued normal smelting.

[0215] (4) Calculation of the optimal adjustment scheme (when triggering control);

[0216] If a certain batch of furnaces occurs In the case of this, the following control logic will be executed:

[0217] Assume the system has adjustable control variables. These represent the electric arc power, oxygen blowing flow rate, feeding speed, and cooling water flow rate, respectively.

[0218] Establish the sensitivity matrix S:

[0219]

[0220] The sensitivity matrix is ​​determined through a combination of mechanistic analysis and statistical data analysis, for example: <0: Increase oxygen blowing volume to reduce carbon content;

[0221] Increase electrical power to raise temperature;

[0222] Optimization objective: To increase cooling water flow and reduce temperature; Minimize weighted control bias and control cost.

[0223]

[0224] in: For the expected correction amount, To control cost weights.

[0225] Constraints: Range of control variables: ;

[0226] Single-step adjustment range: ;

[0227] For example: Electric power adjustment range: -5MW≤ΔP≤+5MW;

[0228] Oxygen flow rate adjustment range: The above optimization problem is solved using a quadratic programming solver (such as CVXOPT) to obtain the optimal control adjustment. .

[0229] (5) Issue control commands;

[0230] The calculated After safety limiting and smoothing, the data is sent to the PLC control system via OPCUA or ModbusTCP industrial communication protocol: If Then send the instruction: increase the electrode current by 15%; if If the command is sent to reduce the oxygen lance valve opening by 10%, the system will continuously monitor the changes in furnace parameters, forming a closed-loop feedback correction to ensure that the endpoint parameters converge to the target value. Specific Implementation Method Two:

[0232] This embodiment describes the application scenario of error compensation strategy B.

[0233] Suppose a small to medium-sized steel plant has accumulated historical data of N = 8000 heats, and the accuracy of the mechanistic model on the validation set is R² = 0.52.

[0234] According to the decision rule: R²>0.3 and N≤20000, choose error compensation strategy B.

[0235] Similar to Specific Implementation Method 1, the mechanistic model is used for calculation: =0.38%; =1625℃;

[0236] Construct a machine learning model (XGBoost algorithm is selected in this scenario, since N=8000∈[1000,10000)), and use the original input features X_the systematic error of the original prediction mechanism model.

[0237] Training objective: Carbon content error: ;

[0238] Temperature error: The model learns the error distribution pattern through training data from 6400 furnace runs, and the model outputs: ; ;

[0239] The advantages of error compensation strategy B are: when the amount of data is limited, it takes the mechanistic model as the main body and uses the data-driven method to learn only the systematic bias, which avoids the overfitting risk of the pure data-driven model, and at the same time corrects the inherent error of the mechanistic model. Specific implementation method three:

[0241] This embodiment describes the application scenario of dynamic weighting strategy C.

[0242] Suppose a large steel plant has accumulated historical data of N = 35,000 heats, and the accuracy of the mechanistic model on the validation set after parameter optimization is R² = 0.76.

[0243] According to the decision rule: R²>0.7 and N>20000, choose the dynamic weighted strategy C.

[0244] The implementation steps of dynamic weighting strategy C are as follows:

[0245] (C1) Working condition clustering;

[0246] The K-means clustering algorithm was used to divide the 28,000 heats of training data into K=5 subgroups: Cluster 1: High scrap ratio (scrap steel ratio > 85%), 5,600 heats; Cluster 2: High iron ratio (iron steel ratio > 25%), 4,200 heats; Cluster 3: Conventional conditions (scrap steel ratio 70-85%), 12,000 heats; Cluster 4: High power short process (P > 100MW, t < 45min), 3,800 heats; Cluster 5: Low power long process (P < 80MW, t > 65min), 2,400 heats. Clustering features included key process parameters such as scrap steel ratio, iron steel ratio, average power, smelting time, and oxygen blowing intensity.

[0247] (C2) Weight optimization;

[0248] For each cluster k, the optimal weight α is searched using Bayesian optimization on the corresponding validation data. k and β k The optimization objective is to minimize RMSE.

[0249] Optimization results: Cluster 1: α1=0.35, β1=0.65 (high scrap steel condition, large composition fluctuation, high data model weight); Cluster 2: α2=0.68, β2=0.32 (high-speed molten iron condition, accurate mechanism model description, high weight); Cluster 3: α3=0.52, β3=0.48 (conventional condition, balanced integration); Cluster 4: α4=0.45, β4=0.55 (high-power condition, complex energy distribution); Cluster 5: α5=0.61, β5=0.39 (long process condition, relatively accurate mechanism model); The weight lookup table is stored in the system database and accessed in real time.

[0250] (C3) Operating condition identification;

[0251] For a new furnace to be predicted, calculate its eigenvector and the Euclidean distance between it and the centers of each cluster:

[0252] Assume the current furnace characteristics are: scrap ratio = 78%, molten iron ratio = 18%, average power = 88MW, smelting time = 58min, and oxygen blowing intensity = 95Nm³ / min.

[0253] Calculate distance: (distance from cluster center 1); (2 distances from the center of cluster); (Minimum distance from cluster center 3); (4 distances from the cluster center); (5 distances from the center of the cluster);

[0254] This furnace batch is determined to belong to Cluster 3 (normal operating condition), and the corresponding weights are extracted: , .

[0255] (C4) Weighted fusion prediction;

[0256] Predictions from the logical model and the machine learning model, respectively:

[0257] Mechanism model: , ;

[0258] Machine learning models (independently trained deep neural networks): , ;

[0259] Weighted fusion:

[0260] ;

[0261] ;

[0262] The advantages of dynamic weighting strategy C are: when the mechanistic model is accurate and the data is sufficient, the fusion weights are dynamically adjusted according to the characteristics of different working conditions, giving full play to the advantages of mechanistic knowledge and data knowledge, and achieving accurate prediction. Specific implementation method four:

[0264] This embodiment describes the application scenario of the pure data-driven strategy D.

[0265] Suppose a newly built steel plant uses a new type of composite furnace charge. The parameters in the mechanism model (such as decarburization efficiency, energy conversion efficiency, etc.) are significantly different from those of the traditional process, resulting in the accuracy of the mechanism model on the validation set being R²=0.18.

[0266] According to the decision rule: R²≤0.3, choose the pure data-driven strategy D.

[0267] The implementation steps of a purely data-driven strategy D are as follows:

[0268] (D1) Data layer modeling;

[0269] Since the historical data volume is N=15000 furnace cycles, the LightGBM gradient boosting decision tree algorithm is adopted according to the algorithm selection rules in step S5.

[0270] Model configuration: Number of trees: 500; Maximum depth: 8; Learning rate: 0.05; Minimum number of samples per leaf node: 20; Feature sampling ratio: 0.8; Sample sampling ratio: 0.8; Input features: Only raw process parameters are used. (55 dimensions), excluding mechanistic predictions.

[0271] (D2) Decision-making level predicted output;

[0272] Input the raw features of the current batch into the trained LightGBM model:

[0273] ;

[0274] Model output: =0.41%; =1642℃;

[0275] The advantage of pure data-driven strategy D is that when the mechanistic model is not applicable or its accuracy is too low, it relies entirely on the data-driven method to automatically learn the input-output mapping relationship from historical data, thus avoiding inaccurate mechanistic knowledge from interfering with the prediction results. Specific implementation method five:

[0277] This embodiment provides an adaptive hybrid intelligent prediction system for the end-point parameters of electric arc furnace smelting, used to implement the above prediction method.

[0278] like Figure 2 As shown, the adaptive hybrid intelligent prediction system for the endpoint parameters of electric arc furnace smelting includes the following modules:

[0279] The data acquisition module connects to the electric arc furnace DCS system and MES system via industrial Ethernet to collect smelting process data in real time.

[0280] (1) Hardware interface: OPCUA server: collects real-time process parameters from the DCS system (sampling frequency 1Hz);

[0281] Database interface: Reads batch batching data and raw material testing data from the MES system;

[0282] ModbusTCP: Collects material quantity data from weighing systems and gas flow meters;

[0283] (2) Data preprocessing: Outlier detection: Outlier data points are identified using the 3σ principle and box plot method;

[0284] Missing value imputation: linear interpolation is used for continuous variables, and forward imputation is used for discrete variables;

[0285] Data synchronization: Aligning data from different sampling frequencies to a unified timestamp;

[0286] Feature engineering: The data acquisition module calculates derived features such as feeding rate, power change rate, and oxygen blowing intensity. The processed data is then stored in the time series database (InfluxDB) for subsequent modules to access.

[0287] The mechanism calculation module contains two calculation units:

[0288] (1) Carbon content mechanism calculation unit;

[0289] Implementing the carbon balance model in step S2 specifically includes: a carbon input calculation subroutine that iterates through each carbon-containing raw material and accumulates the carbon input amount;

[0290] Carbon loss calculation subroutine: Oxidative decarbonization amount = f(oxygen blowing amount, decarbonization efficiency);

[0291] Dust carry-away amount = f(molten steel volume, dust rate, dust carbon content);

[0292] Slag absorption = f(molten steel volume, slag ratio, carbon content in slag);

[0293] Carbon balance main program: Executes the material balance equation and outputs C_mechanism parameter database, which stores the typical carbon content of different raw material types, decarbonization efficiency curves under different process conditions, and other mechanism parameters.

[0294] (2) Temperature mechanism calculation unit;

[0295] Implementing the energy balance model in step S2 specifically includes: inputting the energy calculation subroutine: electrical energy input = f(power, time, electrical efficiency);

[0296] Chemical reaction heat = f(oxygen blowing rate, element content, reaction enthalpy); Fuel combustion heat = f(fuel type, fuel quantity, calorific value); Energy demand calculation subroutine: Latent heat of fusion = f(mass of each material, heat of fusion); Sensible heat of temperature rise = f(mass of each material, specific heat capacity, temperature rise); Slagging heat effect = f(slagging agent type, slagging dosage); Energy loss calculation subroutine: Furnace wall heat dissipation = f(furnace wall area, heat transfer coefficient, temperature difference, time); Heat carried away by flue gas = f(flue gas volume, flue gas temperature, specific heat capacity); Heat carried away by cooling water = f(water flow rate, inlet and outlet water temperature difference); Electrode radiation = f(electrode area, electrode temperature);

[0297] Energy balance main program: Executes the energy balance equation and outputs... The mechanism calculation module is written in C++, with fast calculation speed (single calculation <10ms) and supports real-time online prediction.

[0298] 3. The data modeling module is responsible for training and updating machine learning models:

[0299] (1) Training data preparation: Extract historical furnace data from the time series database;

[0300] Data cleaning: Remove abnormal furnace runs (such as power outages, equipment malfunctions, etc.);

[0301] Feature standardization: Z-score standardization is used to process continuous variables;

[0302] Dataset partitioning: Training set / Validation set / Test set = 8:1:1

[0303] (2) Algorithm adaptive selection:

[0304] Based on the amount of data N, select a suitable modeling algorithm from the algorithm library: SVR algorithm library (Scikit-learn): suitable for N<1000;

[0305] TabNet algorithm library (PyTorch-TabNet): suitable for 1000≤N<10000; XGBoost / LightGBM algorithm library: suitable for 10000≤N<20000; TensorFlow deep learning library: suitable for N≥20000;

[0306] (3) Hyperparameter optimization:

[0307] The optimal combination of hyperparameters is automatically found using grid search or Bayesian optimization (Optuna framework), with the optimization objective being the validation set RMSE.

[0308] (4) Model training: Use the training set data to fit the model;

[0309] Monitor model performance on the validation set to prevent overfitting;

[0310] Evaluate the model's generalization ability on the test set;

[0311] (5) Model storage:

[0312] The trained model is serialized and stored in ONNX or Pickle format for easy cross-platform deployment.

[0313] The data modeling module is deployed on a GPU server and implemented using the Python + TensorFlow / PyTorch framework.

[0314] 4. The quality assessment module periodically (weekly) evaluates the performance of the mechanistic model and the data-driven model:

[0315] (1) Mechanism model accuracy assessment: Extract actual production data from the most recent 1000 furnace batches;

[0316] Backtesting predictions are performed using mechanistic models;

[0317] Calculate the coefficient of determination R² for carbon content and temperature. C and R² T ;

[0318] The smaller value is taken as the overall accuracy R² of the mechanistic model;

[0319] (2) Data sufficiency assessment: Count the number N of historical furnaces that meet the quality standards in the statistical database;

[0320] Assess the temporal distribution and operational condition coverage of the data;

[0321] Identify areas of sparse data and indicate the need for supplementary data;

[0322] (3) Model performance monitoring: Calculate the prediction error distribution: MAE, RMSE, MAPE;

[0323] Plot a scatter plot of predicted vs. actual values ​​to check for systematic bias;

[0324] The results of the analysis of the differences in prediction accuracy under different working conditions are stored in a relational database (MySQL) and a visual report is generated for engineers to view.

[0325] 5. The strategy decision-making module automatically selects the optimal prediction strategy based on the quality assessment results:

[0326] (1) Decision-making logic:

[0327] if R²>0.7 and N>20000: strategy="dynamic weighting strategy C"; elif R²>0.3 and N>20000: strategy="feature enhancement strategy A"; elif R²>0.3 and N<=20000: strategy="error compensation strategy B"; else: strategy="pure data driven strategy D".

[0328] (2) Strategy switching: Strategy switching is triggered when the data quality assessment results change significantly;

[0329] Before switching strategies, A / B testing is conducted to compare the predictive performance of the old and new strategies.

[0330] After confirming that the new strategy is superior to the old strategy, officially switch to it and record the logs.

[0331] (3) Artificial intervention:

[0332] The system offers a manual mode, allowing engineers to force the specification of prediction strategies based on experience.

[0333] The strategy decision-making module is implemented using a rules engine, which supports flexible adjustment of strategy configuration.

[0334] 6. The layered fusion module executes the corresponding fusion prediction logic based on the output of the strategy decision module:

[0335] (1) Feature enhancement execution unit (strategy A) obtains from mechanism calculation module and ;

[0336] The mechanism prediction values ​​are concatenated with the original features to form an enhanced feature set;

[0337] Call the augmented feature model trained by the data modeling module;

[0338] Output the final predicted value.

[0339] (2) The error compensation execution unit (strategy B) obtains the error from the mechanism calculation module. and ;

[0340] Calling the error compensation model for prediction and ;

[0341] The final predicted value is calculated as: Mechanism value + Error compensation value.

[0342] (3) The dynamically weighted execution unit (strategy C) contains three sub-units:

[0343] a) Operating condition identification subunit: Extracts key features of the current furnace batch (scrap ratio, molten iron ratio, power, time, etc.);

[0344] Calculate the Euclidean distance to the center of each cluster;

[0345] Determine the cluster to which it belongs, k*Euclidean.

[0346] b) Weight Query Sub-unit: Query clusters from the weight database corresponding and ;

[0347] Return the weights of the mechanistic model and the weights of the data model;

[0348] c) Weighted fusion sub-unit: Obtain independent predictions from the mechanistic model and the data model respectively;

[0349] Calculate the weighted average based on the weights.

[0350] (4) Pure data-driven execution unit (strategy D);

[0351] Prediction is made by directly calling the data model, without using the mechanistic model.

[0352] The layered fusion module is written in Python and provides prediction services to the outside world through a RESTful API.

[0353] 7. The results output module is responsible for displaying and transmitting the prediction results:

[0354] (1) Human-machine interface display: Real-time display of the predicted values ​​of the final carbon content and temperature;

[0355] Displays the prediction confidence interval and confidence level;

[0356] Comparison of visual mechanism predictions, data predictions, and fusion predictions;

[0357] Displays the currently used forecasting strategy;

[0358] (2) Data storage: Write the prediction results into the historical database;

[0359] Record information such as prediction time, predicted value, confidence interval, and strategy used;

[0360] (3) Interface output: The predicted value is sent to the upper control system via the OPCUA interface;

[0361] Supports industrial communication protocols such as ModbusTCP and Profinet;

[0362] Provides a WebService API for third-party systems to call;

[0363] (4) Alarm function: When the predicted value deviates from the target value by more than the threshold, an alarm is triggered;

[0364] Notify operators via audible and visual alarms and SMS messages;

[0365] The HMI interface module, which displays alarm information and suggested operation results, is developed using Qt / C++ and supports touchscreen operation.

[0366] 8. Model Update Module:

[0367] This module enables online adaptive updates of the model:

[0368] (1) Incremental learning trigger condition: The amount of new data reaches 2000 cycles;

[0369] The model's prediction accuracy continuously decreases beyond a set threshold (e.g., RMSE increases by 10%).

[0370] Significant changes have occurred in the process conditions (such as replacing the furnace lining or adjusting the batching strategy);

[0371] (2) Incremental training process: Extract the data from the 2000 newly added batches;

[0372] Combined with the original training set to form an updated training set;

[0373] Retrain the machine learning model (transfer learning can be used to speed up the training process);

[0374] Evaluate the performance of the updated model on the new validation set;

[0375] If performance improves, deploy the new model; otherwise, retain the original model.

[0376] (3) Mechanism model parameter correction: The actual distribution of parameters such as raw material composition and decarbonization efficiency in the new statistical data;

[0377] Update the average value and fluctuation range in the mechanistic model parameter database;

[0378] Re-evaluate the R² value of the mechanistic model;

[0379] (4) Dynamic strategy adjustment: Re-execute the strategy decision based on the updated R² and N values;

[0380] If the strategy changes, switch to the new strategy;

[0381] If the dynamic weighting strategy C is adopted, then the clustering and weight optimization are performed again;

[0382] (5) Model version management: A new model version number is generated with each update;

[0383] Retain the three most recent historical versions and support version rollback;

[0384] Record metadata such as training time, data volume, and performance metrics for each version. Model updates can be set to automatic mode (triggered periodically) or manual mode (executed after engineer confirmation).

[0385] 9. The execution control module implements automatic closed-loop control of the prediction deviation:

[0386] (1) Deviation monitoring subunit:

[0387] Real-time calculation of prediction deviation; calculation of dynamic threshold; determination of whether to trigger control;

[0388] (2) Optimize the solution of sub-units:

[0389] Establish the optimization problem: Decision variables: ;

[0390] Objective function: ;

[0391] Constraints: u min ≤u0+Δu≤u max ,Δu min ≤Δu≤Δu max Obtaining the sensitivity matrix S:

[0392] Method 1: Analytical differentiation through mechanistic model;

[0393] Method 2: Statistical regression using historical data;

[0394] Method 3: Online identification and real-time optimization of the solution: The CVXOPT library is used to solve the quadratic programming problem;

[0395] The calculation time is less than 50ms, which meets the requirements for real-time control.

[0396] (3) Safety limiting sub-unit:

[0397] Perform a security check on the optimization results:

[0398] (4) Instruction Issuance Subunit:

[0399] Convert control commands into a format recognizable by the actuator: Arc power adjustment: convert to electrode position or transformer tap command;

[0400] Oxygen flow rate adjustment: Converted to oxygen lance valve opening command;

[0401] Feeding speed adjustment: Converted to frequency command of the feeder inverter;

[0402] Cooling water flow rate adjustment: Converted into cooling water valve opening command and sent to PLC or DCS via industrial communication interface;

[0403] (5) Feedback Monitoring Subunit: Monitors the actual effect of control execution.

[0404] Changes in operating parameters within 5 minutes after data collection and execution;

[0405] Calculate the deviation between the actual correction amount and the expected correction amount;

[0406] Evaluate the control effect and feed it back to the optimization solution sub-unit;

[0407] The sensitivity matrix S is updated online to implement adaptive control closed-loop control logic.

[0408] Prediction deviation → Trigger judgment → Optimize solution → Safety check → Instruction issuance → Action execution → Effect monitoring → Model update → Loop.

Claims

1. An adaptive hybrid intelligent prediction method for endpoint parameters in electric arc furnace smelting, characterized in that, Includes the following steps: S1. Data Acquisition: Acquire data on the electric arc furnace smelting process, including raw material composition, raw material addition amount, auxiliary material composition, auxiliary material addition amount, power supply system parameters, gas supply system parameters, water cooling system parameters, and smelting time. S2. Construct a dual-drive prediction system that includes both mechanistic models and data-driven models; S3. Conduct a comprehensive data quality assessment, including: validating the mechanistic model using historical production data, and calculating the coefficient of determination between the model's predicted values ​​and actual measured values. Statistical volume of valid historical data ; S4. Construct a machine learning prediction model based on the amount of effective historical data. Select a modeling algorithm: S5, Based on the coefficient of determination and effective historical data volume Select a hybrid forecasting strategy; based on the selected hybrid forecasting strategy, predict the final carbon content and final temperature values. S6. Based on the output of S5, perform prediction deviation threshold judgment and execution control, adjust the electric arc power, oxygen blowing flow rate, feeding speed or cooling water flow rate to achieve convergence of the endpoint parameters to the target value. The hybrid prediction strategy is specifically as follows: when and When this happens, feature enhancement strategy A is activated; when and When this occurs, error compensation strategy B is activated; when and When this happens, activate dynamic weighting strategy C; when When this happens, enable the pure data-driven strategy D; The feature enhancement strategy A is as follows: the mechanism model outputs the mechanism prediction value; the mechanism prediction value is used as a derived feature and merged with the original production data feature to form an enhanced feature set, which is used to train and build a machine learning prediction model; the prediction value of the machine learning prediction model trained under the enhanced feature set is used as the final prediction value. A1: The mechanistic prediction value of the endpoint carbon content was obtained by calculating using a mechanistic model. and endpoint temperature mechanism prediction value ; A2: Will and As a derived feature, it is related to the features of the original production data. Merging to form an enhanced feature set ; A3: Input the enhanced feature set into the machine learning prediction model, and output the final predicted value of the endpoint carbon content. and final predicted value of endpoint temperature ; The error compensation strategy B is as follows: use the mechanism model to output the mechanism prediction value; use the machine learning prediction model to learn the systematic bias of the mechanism model and output the error compensation value; and add the mechanism prediction value and the error compensation value to obtain the final prediction value. B1: Mechanistic prediction of endpoint carbon content obtained using a mechanistic model. and endpoint temperature mechanism prediction value ; B2: Construct a machine learning prediction model to predict the error of the mechanism model based on the characteristics of the original production data, and output the predicted value of carbon content error. and temperature error prediction value ; B3: Calculate the final predicted values ​​of the endpoint parameters: ； ； The dynamic weighting strategy C: uses a mechanistic model to output mechanistic prediction values; trains and constructs a machine learning prediction model using the features of the original production data; and divides the historical production data into clusters according to a clustering algorithm. Each subgroup has a weighted computer model for each subgroup. and machine learning prediction model weights ,in Determine the cluster to which the new sample belongs. Endpoint parameter prediction value Mechanism model predictions Machine learning model predictions; C1: Use the K-means clustering algorithm or the DBSCAN clustering algorithm to divide the historical production data into... Subgroup; C2: For each subgroup The optimal weights for the mechanistic model and the machine learning model are determined through grid search or Bayesian optimization. The optimization objective is to minimize the root mean square error (RMSE) of the validation set, with the following constraints: and ; C3: For each sample to be predicted, calculate its Euclidean distance to the center of each cluster and determine its cluster affiliation. ; C4: Calculate the final predicted value of the endpoint parameter: ； ； In the formula: The carbon content value independently predicted by the machine learning prediction model. Temperature values ​​independently predicted by machine learning prediction models.

2. The adaptive hybrid intelligent prediction method for end-point parameters of electric arc furnace smelting according to claim 1, characterized in that, The dual-drive prediction system includes: A mechanistic model for the endpoint carbon content is constructed based on the principle of material balance. The endpoint carbon content is calculated according to the following equation: ； In the formula: For the first The quality of carbon-containing raw materials, For the first The percentage of carbon content in a carbon-containing raw material. This represents the number of types of carbon-containing raw materials; For the first Material quality for each carbon loss pathway For the first The percentage of carbon content in each carbon loss pathway The number of carbon loss pathways; The endpoint is the quality of molten steel; among which, carbon-containing raw materials include scrap steel, molten iron, direct reduced iron, scrap iron, carbon powder, and carbon raisers; carbon loss pathways include oxidative decarburization, dust carry-over, and slag absorption; A mechanistic model for the endpoint temperature is constructed based on the principle of energy balance. The endpoint temperature is calculated according to the following equation: ； In the formula: The initial temperature; For total input energy, The energy required for the material to heat up and undergo phase change. Energy loss through various pathways; The average specific heat capacity of molten steel; where, This includes energy from electric arc heating, heat from chemical reactions, heat from fuel combustion, and energy from waste preheating. This includes the latent heat of fusion of the material, the sensible heat of the material's temperature rise, and the heat effect of the slag-forming reaction; This includes heat loss from the furnace bottom and top, heat loss carried away by flue gas, heat loss carried away by cooling water, and electrode radiation loss. Construct a data-driven machine learning prediction model, which is trained based on historical production data.

3. The adaptive hybrid intelligent prediction method for end-point parameters of electric arc furnace smelting according to claim 2, characterized in that, The specific calculation formula for the mechanism model of the endpoint carbon content is as follows: ； In the formula: , , , These are the quality of scrap steel, pig iron, carbon raiser, and recycled materials, respectively. , , , These represent the percentage of carbon content in scrap steel, pig iron, carbon refining agent, and recycled materials, respectively. For oxygen blowing volume, Oxygen density, For decarbonization efficiency, The molar mass ratio of carbon to oxygen; For the quality of smoke and dust, The percentage of carbon content in the smoke and dust; For the quality of slag, This represents the percentage of carbon content in the slag. The specific calculation formula for the mechanism model of the endpoint temperature is as follows: In the formula: The initial temperature. For power conversion efficiency, For input electrical power, For smelting time; This represents the total heat released by the oxidation reaction; Heat of fuel combustion To assist in heating; For the first The latent heat of fusion of a molten material This represents the number of types of materials being melted. For the first The quality of the heated material For the first The specific heat capacity of the heated material For the first Temperature rise of the heated material This represents the number of types of materials being heated; This is due to the heat effect of the slag-forming reaction; For the furnace wall area, The heat transfer coefficient of the furnace wall, The average temperature inside the furnace. The ambient temperature; For flue gas quality, For the specific heat capacity of flue gas, The flue gas temperature; The cooling water removes heat. Emissivity of the electrode surface The Stefan-Boltzmann constant is... The electrode surface area. This represents the electrode temperature.

4. The adaptive hybrid intelligent prediction method for end-point parameters of electric arc furnace smelting according to claim 1, characterized in that, Pure data-driven strategy D: Train and build a machine learning prediction model using the features of the original production data; use the prediction value of the machine learning prediction model as the final prediction value.

5. The adaptive hybrid intelligent prediction method for end-point parameters of electric arc furnace smelting according to claim 1, characterized in that, The selection criteria for the modeling algorithm are as follows: when At that time, the support vector regression algorithm was used; when At that time, a few-sample learning algorithm based on tabular data was adopted; when At that time, the gradient boosting decision tree algorithm is used; when At that time, a deep neural network algorithm was used; When using a deep neural network algorithm, the network structure includes an input layer, at least two hidden layers, and an output layer. The number of nodes in the hidden layers decreases sequentially. The activation function of the hidden layers is the ReLU function, the output layer uses a linear activation function, the loss function is the mean squared error (MSE), and the optimizer is the Adam optimizer. The training process employs early stopping to prevent overfitting, stopping training when the validation set loss fails to improve for 500 consecutive epochs; the total number of iterations is at least 2000.

6. The adaptive hybrid intelligent prediction method for end-point parameters of electric arc furnace smelting according to claim 1, characterized in that, Specifically, S6 is: Calculate the prediction bias: For the predicted values ​​of machine learning prediction models, The target value for the machine learning prediction model; Calculate threshold : ； in, To account for the allowable deviation in the process, To predict uncertainty, Let z be the standard deviation of process fluctuations, and z be the confidence coefficient. when When this happens, execution control is triggered; Calculate the optimal adjustment scheme: ； The difference between the target output and the current predicted output. For sensitivity matrix, For the adjustment vector of control variables, It is the regularization coefficient; constraint: ； It is the lower limit of the control quantity. It is the current control quantity, It is the upper limit of the control quantity. It controls the lower limit of the increment. It controls the upper limit of the increment; The obtained The data is sent to the execution end to adjust the electric arc power, oxygen flow rate, feeding speed, or cooling water flow rate, so as to achieve convergence of the endpoint parameters to the target value.

7. A system employing the adaptive hybrid intelligent prediction method for the end-point parameters of electric arc furnace smelting as described in any one of claims 1-6, characterized in that, include: The data acquisition module is used to collect data on the electric arc furnace smelting process in real time, including raw material composition, raw material addition amount, auxiliary material composition, auxiliary material addition amount, power supply system parameters, gas supply system parameters, water cooling system parameters, and smelting time. The mechanism calculation module includes a carbon content mechanism calculation unit and a temperature mechanism calculation unit; the carbon content mechanism calculation unit calculates the endpoint carbon content mechanism prediction value based on the material balance principle, and the temperature mechanism calculation unit calculates the endpoint temperature mechanism prediction value based on the energy balance principle. The data modeling module is used to train machine learning prediction models based on historical production data and adaptively select modeling algorithms according to the data scale. The quality assessment module is used to evaluate the accuracy of computer model predictions. Statistical historical data volume N; The strategy decision-making module is used to predict accuracy based on the mechanistic model. The following strategies are used to adaptively select feature enhancement strategies based on the amount of historical data N: A) error compensation strategy, B) dynamic weighting strategy, C) or pure data-driven strategy. The layered fusion module includes: The feature enhancement execution unit is used to implement feature enhancement strategy A, which combines the mechanism prediction value as a derived feature with the original production data feature and inputs it into the machine learning prediction model. The error compensation execution unit is used to implement error compensation strategy B, which learns and corrects the systematic bias of the mechanism model through machine learning prediction model; The dynamic weighted execution unit includes a working condition identification subunit, a weight query subunit, and a weighted fusion subunit; the working condition identification subunit determines the working condition category of the sample to be predicted, the weight query subunit retrieves the optimal fusion weight for the corresponding working condition, and the weighted fusion subunit calculates the weighted average of the predicted values ​​from the computer model and the machine learning prediction model. A pure data-driven execution unit that uses only machine learning prediction models for prediction; The results output module is used to output the predicted values ​​of the endpoint carbon content and the endpoint temperature, and provides a prediction confidence assessment. The model update module is used to update the machine learning prediction model online using the production data of the new furnace, and to re-evaluate the prediction accuracy of the mechanistic model and the amount of historical statistical data, and dynamically adjust the hybrid prediction strategy. The execution control module receives the deviation between the predicted result and the target value, calculates the threshold, and determines the triggering condition. When |Δ| exceeds the threshold, the execution control module calculates the optimal control adjustment amount Δu based on the sensitivity matrix or optimization algorithm, and outputs it to the PLC system or DCS system after safety limiting and smoothing. The execution control module controls at least one of the following actuators: arc power, oxygen blowing flow rate, feeding speed, and cooling water flow rate, to realize automatic adjustment and closed-loop correction of the smelting process.

8. The system according to claim 7, characterized in that, The execution control module includes: The deviation monitoring subunit is used to calculate the predicted deviation Δ and determine the threshold in real time. The optimized solution sub-unit is used to solve for the optimal adjustment scheme Δu; The safety limiting subunit is used to perform limiting, rate limiting, and smoothing. The instruction issuing subunit is used to send control instructions to the electric furnace control system through the industrial communication interface; The feedback monitoring subunit is used to collect post-execution operating parameters and update the model to achieve closed-loop control.

Images