A method and system for constant temperature control of the smelting stage of an electric arc furnace flat bath

Abstract

The present application belongs to the technical field of metallurgical engineering, and discloses a temperature constant control method and system for the flat bath smelting stage of an electric arc furnace. The mechanism model and the partition machine learning model are fused to realize accurate prediction and control of the bath temperature. The mechanism model is constructed using the energy balance principle, and the prediction error is corrected through the machine learning model. A gating network is used to dynamically adjust the weight of error correction according to real-time process parameters, so as to obtain a highly accurate mixed prediction temperature. At the control level, a hierarchical control architecture is adopted to decompose the temperature deviation to the bottom layer multivariable proportional-integral-derivative controller, the middle layer model predictive controller and the upper layer intelligent optimization algorithm, and the control results of the three layers are adaptively weighted and fused. The multi-model, hierarchical and adaptive control strategy significantly improves the accuracy and stability of the electric arc furnace temperature control, and optimizes the process consumption, thereby providing technical support for efficient and low-energy smelting process.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of metallurgical engineering, in particular to a temperature constant control method and system for flat bath smelting stage of electric arc furnace. BACKGROUND

[0002] The electric arc furnace is an important equipment for steel smelting. In the flat bath smelting stage, the stable control of the bath temperature has a decisive influence on the quality of the liquid steel, energy consumption control and production efficiency. The traditional temperature control of the electric arc furnace mainly relies on manual experience operation, which has the problems of low control precision, response lag and high energy consumption. The existing temperature control methods of the electric arc furnace mainly include: the method based on PID control, which has a simple structure, but for the electric arc furnace which is a complex system with multiple variables, strong coupling and nonlinearity, a single PID control cannot achieve satisfactory control effect. The method based on fuzzy control can handle certain nonlinear problems, but the parameter adjustment is complex and strongly dependent on expert experience. The method based on neural network has strong nonlinear mapping ability, but it needs a large amount of training data and has the problems of local convergence and overfitting. The method based on mechanism model has the advantage of clear physical meaning, but due to the complexity of the electric arc furnace system, the pure mechanism model often has the problem of large modeling error. The main problems existing in the prior art include: a single control algorithm cannot adapt to the complex characteristics of the electric arc furnace system; there is a lack of effective temperature prediction method, and the control is mainly passive response; the respective advantages of mechanism knowledge and data-driven method are not fully combined; the control strategy lacks hierarchy, and it is difficult to balance the requirements of fast response and long-term optimization. SUMMARY

[0003] The present application provides a temperature constant control method and system for flat bath smelting stage of electric arc furnace, which aims to solve the problems of low temperature control precision, large fluctuation and high energy consumption in the existing electric arc furnace smelting process, to realize the stable temperature control in the smelting stage, and to improve the smelting efficiency and product quality.

[0004] The technical scheme of the present application is as follows: a temperature constant control method for flat bath smelting stage of electric arc furnace, comprising the following steps:

[0005] S1, constructing a temperature prediction mechanism model of electric arc furnace: based on the principle of energy balance, a mechanism model is established with smelting process parameters as input variables and electric arc furnace bath temperature as output variable; the smelting process parameters include scrap steel quantity, oxygen blowing quantity, carbon powder quantity, natural gas quantity, electric power consumption quantity and lime quantity;

[0006] S2, constructing a partition machine learning error correction model:

[0007] The actual smelting end temperature in the historical smelting data is obtained, and the error between the predicted temperature of the mechanism model and the actual end temperature is calculated ;

[0008] According to the error interval, the training data is divided into three parts: The training data of the first error correction model is used to train the first error correction model ; The training data of the second error correction model is used to train the second error correction model ; The training data of the third error correction model is used to train the third error correction model ;

[0009] Train the gating model , whose input is the current smelting process parameters, and the output is a probability vector , which is used to indicate the weight of the three error correction models in the current state; the gating model adopts a neural network classifier based on Softmax output layer;

[0010] S3, real-time temperature prediction:

[0011] Real-time acquisition of current smelting process parameters, input into the mechanism model, and get the mechanism predicted temperature ;

[0012] Input the current smelting process parameters into the three error correction models, respectively, to get the error prediction value ;

[0013] The gating model outputs the probability , and calculates the weighted error prediction value: ; When , trigger the Argmax hard classifier to select the corresponding error correction model or adopt the conservative reduction strategy;

[0014] Calculate the hybrid predicted temperature: ;

[0015] S4, hierarchical control of smelting process parameters:

[0016] Set the target temperature of the arc furnace flat pool smelting stage ;

[0017] Calculate the temperature deviation ;

[0018] Use the hierarchical control architecture to process the temperature deviation, including:

[0019] Bottom multivariable proportional-integral-derivative controller: decompose into each smelting process parameter dimension, input into the corresponding bottom multivariable proportional-integral-derivative controller, and get the basic adjustment amount ;

[0020] Middle model predictive controller: based on And the current smelting process parameter state, under the constraint condition, the predicted adjustment amount is optimized and calculated ;

[0021] Upper intelligent optimization algorithm: genetic algorithm is adopted to And Fusion optimization, output ;

[0022] Three layers of result weighted fusion, get the final adjustment amount:

[0023]

[0024] , , Respectively, the weight is adjusted according to the self-adaptive adjustment rule: when , Increase by 5% each time; when , Increase by 5% each time; when the arc furnace flat pool smelting stage temperature constant control system continuously and stably runs Start interval, Gradually increase by 5% each time; wherein, Is , Is , and satisfies ;

[0025] S5, closed loop execution and termination condition: according to Adjust the smelting process parameters; when the temperature prediction after implementation Enter Interval, closed loop control is terminated, until the next start; start interval is .

[0026] The arc furnace temperature prediction mechanism model adopts the following energy balance equation:

[0027]

[0028] Wherein, Is the predicted temperature, Is the current temperature, Is the heat generated by electric energy input; Is the heat generated by natural gas combustion; Is the heat generated by carbon powder oxidation reaction; Is the heat released by impurity element oxidation in scrap steel melting process; Is the reaction heat consumption, Is the heat loss, Is the pool quality, Cp is the specific heat capacity, is the time step.

[0029] The first error correction model is trained by support vector regression; the second error correction model is trained by random forest regression; and the third error correction model is trained by neural network regression.

[0030] The control equation of the bottom-layer multivariable proportional-integral-derivative controller is:

[0031]

[0032] wherein, is the PID output of the m-th smelting process parameter, are proportional coefficient, integral coefficient and derivative coefficient respectively, is the corresponding temperature deviation component. The middle-layer model predictive controller adopts the following optimization objective function:

[0033]

[0034]

[0035] The constraint condition is:

[0036]

[0037] wherein, is the predicted output, is the reference trajectory, is the control input, is the input increment, is the prediction time domain, is the control time domain, is the weight matrix.

[0038] The genetic algorithm is specifically:

[0039] Individual coding: the smelting process parameter adjustment amount is coded as a real number vector;

[0040] Fitness function:

[0041]

[0042] wherein, is the process consumption cost, is the control stability index, is the weight coefficient;

[0043] Genetic operation: real number crossover and Gaussian mutation are adopted;

[0044] Elitist reservation: the optimal individual of each generation is reserved. ​

[0045] The method also includes an online learning step:

[0046] Collect new smelting data in real time;

[0047] When the accumulated data reaches a threshold, retrain the error correction model With the gating model ;

[0048] Use a sliding window to keep the data up-to-date;

[0049] If the cross-validation performance improves, replace the original error correction model and the gating model.

[0050] The processing of the temperature deviation also includes feedforward control: according to the scrap composition and process requirements, predict the future temperature change trend, generate a feedforward compensation amount, and superimpose it with the feedback control amount calculated from the temperature deviation To achieve feedforward-feedback composite control.

[0051] An electric arc furnace flat bath smelting stage temperature constant control system applies an electric arc furnace flat bath smelting stage temperature constant control method, comprising:

[0052] Data acquisition module: real-time acquisition of smelting process parameters and temperature data;

[0053] Mechanism model calculation module: based on energy balance, calculate the mechanism prediction temperature ;

[0054] Partition machine learning module: based on error correction model and gating model Output ;

[0055] Temperature prediction module: calculate ;

[0056] Hierarchical control module: including PID control submodule, MPC submodule and genetic algorithm optimization submodule;

[0057] Execution control module: adjust the corresponding smelting process parameters;

[0058] Termination control module: when Enter Pause the closed loop until the next start.

[0059] The beneficial effects of the present application: by combining the mechanism model based on energy balance and the partition machine learning error correction model, the physical law of smelting process parameters and the nonlinear mode of historical data are comprehensively utilized, the temperature prediction accuracy is significantly improved, the prediction error is controlled within 95%, which is better than the traditional single model method. Adopting hierarchical control architecture, combining the fast response of the bottom layer multivariable proportional-integral-derivative controller, the dynamic optimization of the middle layer model predictive controller (MPC) and the global search ability of the upper layer intelligent optimization algorithm, the temperature fluctuation of the flat bath smelting stage is controlled within the target temperature range, ensuring the stability of the smelting process. Through accurate temperature control and process parameter optimization adjustment, the energy waste of excessive heating or cooling is reduced, the consumption of electric energy, oxygen and carbon powder is reduced, and the overall energy consumption is reduced by 10%-15%, which significantly improves the economic benefit of the electric arc furnace. The partition error correction model According to the error interval, the soft weighting mechanism and the hard classifier strategy of the gating model are combined to effectively deal with the error distribution change under different smelting conditions, improve the adaptability and robustness of the model to complex working conditions. Through the online learning step, new data is collected in real time and the error correction model and the gating model are updated regularly, and the sliding window mechanism is used to maintain the timeliness of the data, ensuring that the system can adapt to dynamic factors such as scrap composition changes and equipment aging, and maintain control performance for a long time. The introduction of feedforward control can predict the temperature change trend according to the scrap composition and process requirements, and adjust the process parameters in advance, combined with feedback control, further improve the response speed and control accuracy of the system, and reduce temperature overshoot and oscillation. The hierarchical control architecture optimizes PID, MPC and genetic algorithm, considers temperature deviation, process constraints and control stability, and outputs the optimal process parameter adjustment, which improves smelting efficiency and product quality. The modular design of the control system is convenient for hardware deployment and supports function extension (such as adding process parameters or control strategies), which is suitable for electric arc furnaces of different sizes and types, and has wide industrial application prospects. The reasonable design of closed-loop control and termination condition, combined with 10s~30s start interval, balances the calculation load and control accuracy, avoids equipment fatigue caused by frequent adjustment, and improves operation stability and equipment safety. By reducing energy consumption and optimizing resource utilization, carbon emissions and waste gas generation are reduced, which meets the requirements of green manufacturing and sustainable development, and provides technical support for energy saving and emission reduction in the steel industry. BRIEF DESCRIPTION OF DRAWINGS

[0060] Figure 1 A flat bath smelting stage temperature constant control method for an electric arc furnace.

[0061] Figure 2 A flat bath smelting stage temperature constant control system for an electric arc furnace. DETAILED DESCRIPTION

[0062] A temperature constant control method for the flat bath smelting stage of an electric arc furnace combines a mechanism model and a machine learning error correction model for temperature prediction, and adopts a hierarchical control architecture to realize accurate adjustment of process parameters, including the following steps:

[0063] S1, a temperature prediction mechanism model of an electric arc furnace is constructed: based on the principle of energy balance, a mechanism model is established with smelting process parameters as input variables and electric arc furnace bath temperature as output variables, the smelting process parameters including scrap steel quantity, oxygen blowing quantity, carbon powder quantity, natural gas quantity, power consumption quantity and lime quantity. Specifically, the mechanism model adopts the following energy balance equation:

[0064] wherein, is the predicted temperature, is the current temperature, is the heat generated by electric energy input; is the heat generated by natural gas combustion; is the heat generated by carbon powder oxidation reaction; is the heat released by impurity element oxidation in the scrap steel melting process; is the reaction heat consumption, is the heat loss, is the bath mass, is the specific heat capacity, is the time step. The mechanism model realizes dynamic prediction of the bath temperature by quantifying the input heat (such as arc heat, chemical reaction heat) and loss heat (such as radiation, convection heat loss).

[0065] S2, a partition machine learning error correction model is constructed:

[0066] (a) obtaining the actual smelting end point temperature in the historical smelting data, calculating the error between the mechanism model predicted temperature and the actual end point temperature to identify the system deviation of the mechanism model;

[0067] (b) dividing the training data according to the error interval: according to the training data of , a first error correction model is constructed for processing small error scenarios; according to the training data of , a second error correction model is constructed for medium error scenarios; according to the training data of , a third error correction model is constructed for large error scenarios.

[0068] The first error correction model is trained by support vector regression; the second error correction model is trained by random forest regression; and the third error correction model is trained by neural network regression.

[0069] (c) Training a gating model Gate, which takes smelting process parameters as input and outputs a probability vector , indicating the weights of selecting three error correction models under the current state. The gating model adopts a neural network classifier based on a Softmax output layer to achieve soft weighted fusion. The gating model realizes adaptive model selection by learning the error distribution in historical data.

[0070] S3, Real-time temperature prediction:

[0071] (a) Collecting current smelting process parameters in real time, including scrap steel quantity, oxygen blowing quantity, carbon powder quantity, natural gas quantity, power consumption quantity, and lime quantity;

[0072] (b) Inputting the current smelting process parameters into the mechanism model to obtain a mechanism predicted temperature as a basic temperature estimate;

[0073] (c) Inputting the smelting process parameters into , , , respectively, to obtain error prediction values , , to correct the potential bias of the mechanism model;

[0074] (d) The gating model outputs a probability , and calculates a weighted error prediction value: When (where is a preset threshold), a hard classifier is triggered to select the corresponding error correction model or adopt a conservative reduction strategy (such as reducing the learning rate or reverting to the default model) to ensure prediction robustness;

[0075] (e) Calculating a hybrid predicted temperature: to realize hybrid prediction of mechanism and data-driven, and improve the accuracy of temperature estimation.

[0076] S4, Hierarchical control of smelting process parameters:

[0077] (a) Setting a target temperature for the arc furnace flat bath smelting stage, which is determined according to smelting requirements and material characteristics;

[0078] (b) Calculating a temperature deviation as a control input;

[0079] (c) Processing the temperature deviation using a hierarchical control architecture, including: a bottom layer multivariable proportional-integral-derivative controller (PID): decomposing into each process parameter dimension, inputting into the corresponding PID controller to obtain a basic adjustment amount The control equation of the bottom PID controller is:

[0080]

[0081] wherein, is the PID output of the i-th process parameter, , , , are proportional, integral and differential coefficients respectively, is the corresponding temperature deviation component; the layer provides fast response.

[0082] The middle model predictive controller (MPC) optimizes and calculates the predicted adjustment amount under the constraint condition based on the current process parameter state and the predicted state. The middle model predictive controller adopts the following optimization objective function:

[0083]

[0084] Constraint condition:

[0085]

[0086] wherein, is the predicted output, is the reference trajectory, is the control input, is the input increment, is the prediction time domain, is the control time domain, is the weight matrix; the layer considers future dynamics and constraints to achieve optimal control.

[0087] The upper intelligent optimization algorithm fuses and optimizes the predicted state and the predicted adjustment amount by using the genetic algorithm, and outputs the optimized process parameter adjustment amount. The upper intelligent optimization algorithm includes: (a) Individual encoding: the smelting process parameter adjustment amount is encoded as a real number vector;

[0088] (b) Fitness function:

[0089]

[0090]

[0091] wherein, is the temperature error, is the process consumption cost, is the control stability index, , , are weight coefficients.​​​

[0092] (c) Genetic manipulation: Real number crossover and Gaussian mutation are adopted;

[0093] (d) Elitism: The best individual of each generation is reserved. This layer provides global optimization and handles complex nonlinear problems;

[0094] The final adjustment amount is obtained by three-layer result weighted fusion:

[0095]

[0096] The adaptive weight adjustment rule is: when , increases by 5% each time; when , increases by 5% each time; when the electric arc furnace flat pool smelting stage temperature constant control system has been continuously and stably running for N start intervals, gradually increases by 5% each time; wherein, preferably , preferably , and satisfies . In addition, the temperature deviation processing also includes feedforward control: according to the scrap composition and process requirements, the future temperature change trend is predicted, and the process parameters are adjusted in advance to realize feedforward-feedback composite control, further improving the response speed and stability.

[0097] S5, closed loop execution and termination condition: according to adjust the process parameters to realize real-time regulation and control of the smelting process; when the temperature prediction after implementation enters the interval, the closed loop control is terminated until the next start; the start interval is 10s-30s to balance the calculation load and control accuracy.

[0098] The method further includes an online learning step: (a) collecting new smelting data in real time; (b) when the cumulative data reaches a threshold, retraining the error correction model , , and the gate model Gate; (c) using a sliding window to maintain data timeliness to ensure that the model adapts to process changes; (d) if the cross-validation performance improves, replace the original model to realize adaptive updating of the model.

[0099] An electric arc furnace flat pool smelting stage temperature constant control system includes the following modules:

[0100] Data acquisition module: real-time acquisition of process parameters and temperature data, including scrap quantity, oxygen blowing quantity, carbon powder quantity, natural gas quantity, power consumption quantity, lime quantity, and molten pool temperature measurement value;

[0101] Mechanism model calculation module: based on energy balance calculation , dynamic temperature prediction is carried out using the above energy balance equation;

[0102] Partition machine learning module: based on , , And the gate output of the model Gate , the partition processing and weighted fusion of error correction are realized;

[0103] Temperature prediction module: calculate , provide hybrid temperature prediction;

[0104] Hierarchical control module: including PID control submodule for bottom layer multivariable PID control; MPC submodule for middle layer model predictive control; genetic algorithm optimization submodule for upper layer intelligent optimization; This module processes temperature deviation and realizes three-layer weighted fusion adjustment;

[0105] Execution control module: adjust the corresponding process parameters according to the final adjustment amount, such as adjusting the oxygen valve and electrode power through the actuator;

[0106] Termination control module: when Enter Pause the closed loop until the next start (10s~30s).

[0107] The electric arc furnace flat pool smelting stage temperature constant control system realizes the hardware deployment of the method through modular design, supports online learning and feedforward control extension.

[0108] Embodiment 1

[0109] This embodiment describes a hierarchical control method based on a mechanism-machine learning hybrid model for realizing temperature constant control in the flat pool smelting stage of a 120-ton electric arc furnace.

[0110] 1. Implementation process overview

[0111] The method flow of this embodiment is shown in Figure 1 First, a mechanism model for electric arc furnace temperature prediction based on the principle of energy balance is established; second, a partition machine learning error correction model is constructed using historical data to correct the prediction error of the mechanism model; then, the mechanism model and the error correction model are fused to perform real-time temperature prediction; then, a hierarchical control architecture (bottom layer PID, middle layer MPC, upper layer intelligent optimization algorithm) is used to process temperature deviation in stages, and the outputs of each layer are weighted and fused to obtain the final process parameter adjustment amount; finally, the control instruction is executed and the termination condition is judged.

[0112] 2. Detailed steps and parameter settings

[0113] Step S1: Constructing the mechanism model of EAF temperature prediction;

[0114] According to the principle of energy balance, the differential equation of the EAF bath temperature is established. The mechanism model equation used in this embodiment is as follows:

[0115]

[0116] wherein, is the predicted temperature, is the current temperature, is the heat generated by the input of electric energy; is the heat generated by the combustion of natural gas; is the heat generated by the oxidation reaction of carbon powder; is the heat released by the oxidation of impurity elements in the scrap steel melting process; is the heat consumed by the reaction, is the heat loss. Model parameter setting: bath mass : 80t is assumed. Specific heat capacity : is . Time step : 10 seconds is set.

[0117] Step S2: Constructing a partitioned machine learning error correction model;

[0118] This step aims to correct the system error of the mechanism model due to the simplifying assumption.

[0119] (a) Obtain data and calculate error: extract 1000 furnace times of process data (electricity consumption, oxygen blowing, carbon powder, etc.) and corresponding actual end point temperature in the flat bath stage from the historical database. Use the mechanism model of step S1 to predict these historical furnace times, and calculate the error between the predicted temperature of the mechanism model and the actual end point temperature: . is the temperature prediction error, i.e. the difference between the predicted temperature of the mechanism model and the actual measured temperature. is the model predicted temperature. is the actual measured temperature.

[0120] (b) Divide the training data according to the error interval: the first error interval: when , the corresponding data is divided into set . The second error interval: when , the corresponding data is divided into set . The third error interval: when , the corresponding data is divided into set For each data set, an error correction model is trained respectively. In this embodiment, the first error correction model is established by using support vector regression algorithm, the model parameter is set as radial basis kernel function, the penalty coefficient C is 1.0, the slack variable is 0.1, and the kernel function parameter is in adaptive proportion form. The second error correction model is established by using random forest regression algorithm, the parameter is set as the number of trees 2000, the maximum depth is not limited, the minimum split sample number is 5, the minimum leaf node sample number is 2, the feature selection mode is square root mode, and the out-of-bag data score is enabled. The third error correction model is established by using neural network regression algorithm, the network structure includes an input layer, three full-connection hidden layers and an output layer, the number of hidden layer nodes is 128, 64 and 32 respectively, the activation function is ReLU, the random inactivation proportion between layers is set as 0.2, the number of output layer nodes is 1 and the linear activation is used; in the training process, the optimizer is Adam, the learning rate is 0.001, the loss function is mean square error, the batch size is 64, the maximum iteration round is 2000, and the early stopping mechanism is set to avoid overfitting.

[0121] (c) Training the gating model: the gating model uses a small neural network classifier, the input is the current smelting process features (power consumption, oxygen blowing, etc.), and the output is a Softmax vector of three probabilities . The gating model determines which error correction model should be used by learning the relationship between the current process state and the historical error interval. The network structure is set as: input layer (7 nodes corresponding to 7 input features), hidden layer 1 (64 nodes, ReLU activation), hidden layer 2 (32 nodes, ReLU activation), and output layer (3 nodes, Softmax activation).

[0122] Step S3: real-time temperature prediction;

[0123] (a) Real-time acquisition: real-time acquisition of process parameters such as power consumption, oxygen blowing, carbon powder, natural gas and lime amount every 10 seconds.

[0124] (b) Mechanism prediction: input the collected process parameters into the mechanism model to obtain the mechanism predicted temperature .

[0125] (c) Error prediction: input the process parameters into the three error correction models to obtain the respective error prediction values .

[0126] (d) Gating model weighted fusion: input the process parameters into the gating model to obtain the probability vector . Calculate the weighted error prediction value: .

[0127] (e) Mixed Prediction: The final mixed prediction temperature is: .

[0128] Step S4: Layered control of smelting process parameters;

[0129] Control objective: Set the target temperature for the flattened molten pool stage. .

[0130] Calculate temperature deviation: .

[0131] Layered control architecture:

[0132] Low-level multivariable PID controller: The parameters are broken down into four control variables: power consumption, oxygen blowing, toner, and natural gas, which serve as inputs to each PID controller. PID parameter settings: Power consumption PID Oxygen blowing toner ,natural gas Obtain the basic adjustment amount for each control variable. .

[0133] Mid-level Model Predictive Controller (MPC): Optimization Objective: . Indicates the current moment. Indicates the prediction step. For at any time The predicted first Step system output, For the first The reference output of the step. For the first Step-by-step control input prediction To control incremental forecasting, For the prediction time domain (30 seconds). To control the time domain (10 seconds). The output error weight matrix is... For the control quantity weight matrix, To control the incremental weight matrix, To constrain the increase in power consumption, To constrain oxygen blowing volume, The resulting prediction adjustment is used for optimization.

[0134] Upper-level intelligent optimization algorithm (genetic algorithm): Individual encoding: Real number encoding is used to encode the adjustment amounts of four process parameters—power consumption, oxygen blowing, carbon powder, and natural gas—into a vector. Fitness function: Weighting coefficients: Genetic operation: single-point real number crossover and Gaussian mutation. Elite preservation: the best 10% individuals are preserved in each generation; the optimization adjustment amount is obtained by optimization .

[0135] Weighted fusion: final adjustment amount Initial weight: Adaptive weight adjustment: when , increases by 5% each time. When the control system has been running stably for 100 start intervals in succession, , it gradually increases by 5% each time. Meanwhile, the conditions .

[0136] Step S5: closed-loop execution and termination condition According to , the process parameters such as electrode lifting, oxygen blowing, carbon powder and natural gas are adjusted. Temperature prediction and control adjustment are performed again every 15 seconds (start interval). When the temperature prediction after implementation enters the interval, the closed-loop control is terminated, and the system enters the "maintenance" mode until the next start interval.

[0137] Example 2

[0138] This example describes a control system for implementing the method of Example 1, the hardware and software module configuration of which is as follows. This system is designed for a 120-ton electric arc furnace.

[0139] 1. System architecture: a distributed control architecture is adopted, mainly composed of an upper computer (industrial control computer), a data acquisition module, an execution control module, and various functional calculation modules, as shown in Figure 2 .

[0140] 2. Detailed module configuration and function

[0141] Data acquisition module: hardware: sensor array and data acquisition card. Function: real-time acquisition of electric arc furnace current, voltage, power data, calculation of power consumption. Real-time acquisition of oxygen blowing amount, natural gas amount through flow meter. Carbon powder amount, lime amount through weighing sensor or silo counter. Acquisition of molten pool temperature through high-precision thermocouple. All data are sent to the upper computer for processing at a frequency of 1 second.

[0142] Mechanism model calculation module: hardware: high-performance CPU in the upper computer. Software: mechanism model calculation program written in Python. Function: receive real-time data from the data acquisition module, calculate the mechanism prediction temperature every 10 seconds according to the energy balance equation in Example One.

[0143] Partitioned Machine Learning Module: Hardware: GPU in the host computer to accelerate model inference. Software: Trained and deployed machine learning models. Functionality: Error Correction Submodule: Contains three pre-trained Support Vector Regression (SVR) models for computation. The gated model submodule contains a pre-trained neural network classifier used to compute probability vectors. .

[0144] Temperature Prediction Module: Hardware: Host computer CPU. Software: Fusion algorithm program. Function: Calculates the final hybrid predicted temperature by combining the mechanistic prediction temperature with the weighted error prediction value. .

[0145] Hierarchical Control Module: Hardware: Host computer CPU. Software: Hierarchical control algorithm program. Function: PID control submodule: Based on the PID parameters set in Example 1, calculates the PID adjustment amounts for power consumption, oxygen blowing, toner, and natural gas in real time. The MPC submodule solves the MPC optimization problem in real time and calculates the predicted adjustment amount. Genetic Algorithm Optimization Submodule: This module uses a genetic algorithm to fuse and optimize the outputs of PID and MPC, resulting in... The fusion submodule: weights and fuses the outputs of the three layers according to adaptive weighting rules to obtain the final adjustment. .

[0146] Execution Control Module: Hardware: Industrial Programmable Logic Controller. Function: Receives the final adjustment amount from the hierarchical control module, converts it into actual control commands, and adjusts the corresponding process parameters through actuators (such as electrode lifting drive system, oxygen lance valve, carbon powder injection device, etc.).

[0147] Termination Control Module: Hardware: Host computer CPU. Software: Logic judgment program. Function: Checks the current predicted temperature every 15 seconds. Enter? The system enters a stable range. If the condition is met, a "pause closed loop" command is sent to the execution control module, and the system enters a hold state until the next startup interval begins a new cycle.

[0148] Online Learning Module: Hardware: Host computer CPU / GPU. Function: Real-time collection of new furnace data. When the accumulated data reaches 1000 furnaces, the online learning program is triggered. This program retrains the error correction model and the gating model using the latest data, and evaluates the performance of the new model through cross-validation. If the performance improvement exceeds a preset threshold (R... 2 If the improvement is 5%, the currently running model will be automatically replaced to ensure the system's adaptability and robustness.

Claims

1. A method for maintaining constant temperature during the smelting stage of an electric arc furnace flat-walled molten pool, characterized in that, Includes the following steps: S1. Constructing a temperature prediction mechanism model for electric arc furnaces: Based on the principle of energy balance, a mechanism model is established with smelting process parameters as input variables and electric arc furnace molten pool temperature as output variable; the smelting process parameters include scrap steel quantity, oxygen blowing quantity, carbon powder quantity, natural gas quantity, electricity consumption, and lime quantity; S2. Construct a partitioned machine learning error correction model: Obtain the actual smelting endpoint temperature from historical smelting data, and analyze the error between the computer model's predicted temperature and the actual endpoint temperature. ; Training data is divided according to error intervals: The training data was used to train and build the first error correction model. ; The training data was used to train the second error correction model. ; The training data is used to train the third error correction model. ; Training Gated Model Its input is the current smelting process parameters, and its output is a probability vector. , used to indicate the weights of the three error correction models in the current state; the gated model uses a neural network classifier based on the Softmax output layer; S3, Real-time Temperature Prediction: Real-time acquisition of current smelting process parameters, input into the mechanism model, and obtaining the mechanism-predicted temperature. ; The current smelting process parameters are input into three error correction models to obtain the predicted error values. ; Gated model output probability And calculate the weighted error prediction value: ;when When this occurs, the Argmax hard classifier is triggered to select the corresponding error correction model or adopt a conservative reduction strategy. Calculate the mixed predicted temperature: ; S4. Layered control of smelting process parameters: Set the target temperature for the flat-pool smelting stage in the electric arc furnace. ; Calculate temperature deviation ; The temperature deviation is processed using a hierarchical control architecture to obtain the final adjustment amount. ; S5. Closed-loop execution and termination conditions: According to... Adjusting smelting process parameters; temperature prediction after implementation. Enter When the interval is reached, the closed-loop control terminates until the next start; the start interval is... .

2. The method for maintaining constant temperature during the smelting stage of an electric arc furnace flat-walled molten pool according to claim 1, characterized in that, The hierarchical control architecture includes: Underlying multivariable proportional-integral-derivative controller: Decompose the parameters into individual smelting process parameters and input them into the corresponding underlying multivariable proportional-integral-derivative controller to obtain the basic adjustment amount. ; Mid-level model predictive controller: based on Based on the current smelting process parameters, optimize and predict the adjustment amount under constraints. ; Upper-level intelligent optimization algorithm: using genetic algorithm to optimize the upper-level intelligent optimization algorithm. and Integration and optimization, output ; The three-layer results are weighted and fused to obtain the final adjustment amount: , , These are the weights, and the weights are adjusted according to the adaptive adjustment rules: when hour, Increase, each time by 5%; when hour, Increase by 5% each time; when the temperature constant control system of the electric arc furnace flat molten pool smelting stage is continuously and stably operating. After one startup interval Gradually increase, with each increment being 5%; among which, for , for and satisfy .

3. The method for maintaining constant temperature during the smelting stage of an electric arc furnace flat-walled molten pool according to claim 1, characterized in that, The electric arc furnace temperature prediction mechanism model adopts the following energy balance equation: in, To predict temperature, The current temperature. The heat generated by electrical energy input; The heat generated by the combustion of natural gas; The heat generated by the oxidation reaction of the toner; This refers to the heat released during the oxidation of impurity elements in the process of melting scrap steel. The reaction consumes heat. For heat loss, For the quality of the molten pool, For specific heat capacity, For time step.

4. The method for maintaining constant temperature during the smelting stage of an electric arc furnace flat-walled molten pool according to claim 1, characterized in that, The first error correction model is trained using support vector regression; the second error correction model is trained using random forest regression; and the third error correction model is trained using neural network regression.

5. The method for maintaining constant temperature during the smelting stage of an electric arc furnace flat-walled molten pool according to claim 2, characterized in that, The control equations of the underlying multivariable proportional-integral-derivative controller are as follows: in, For the first PID output of various smelting process parameters These are the proportional coefficient, integral coefficient, and differential coefficient, respectively. This corresponds to the temperature deviation component.

6. The method for maintaining constant temperature during the smelting stage of an electric arc furnace flat-walled molten pool according to claim 2, characterized in that, The mid-level model predictive controller adopts the following optimization objective function: The constraints are as follows: in, To predict the output, For reference trajectory, To control the input, For input increment, To predict the time domain, To control the time domain, This is the weight matrix.

7. The method for maintaining constant temperature during the smelting stage of an electric arc furnace flat-walled molten pool according to claim 2, characterized in that, The genetic algorithm is specifically as follows: Individual coding: Encode the adjustment amount of smelting process parameters as a real number vector; Fitness function: in, For process consumption costs, To control stability indicators, These are the weighting coefficients; Genetic operations: Real crossover and Gaussian mutation were used; Elite retention: Retaining the best individuals from each generation.

8. The method for maintaining constant temperature during the smelting stage of an electric arc furnace flat-walled molten pool according to claim 1, characterized in that, It also includes online learning steps: Collect new smelting data in real time; When the accumulated data reaches a threshold, the error correction model is retrained. With gated model ; Use sliding windows to maintain data timeliness; If cross-validation performance improves, then replace the original error correction model and gating model.

9. The method for maintaining constant temperature during the smelting stage of an electric arc furnace flat-walled molten pool according to claim 1, characterized in that, The temperature deviation processing also includes feedforward control: based on the scrap steel composition and process requirements, predicting future temperature change trends, generating a feedforward compensation amount, and comparing it with the temperature deviation... The calculated feedback control quantities are superimposed to achieve feedforward-feedback composite control.

10. A constant temperature control system for the smelting stage of an electric arc furnace flat-walled molten pool, characterized in that, The method for maintaining constant temperature during the smelting stage of an electric arc furnace flat-pool as described in any one of claims 1-9 includes: Data acquisition module: Real-time acquisition of smelting process parameters and temperature data; Mechanism Model Calculation Module: Temperature Prediction Based on Energy Balance Computation ; Partitioned Machine Learning Module: Based on Error Correction Model and Gating Model Output ; Temperature prediction module: calculation ; Hierarchical control module: includes PID control submodule, MPC submodule and genetic algorithm optimization submodule; Execution control module: Adjusts the corresponding smelting process parameters; Termination control module: When Enter The closed loop is paused until the next startup.

Images