Method for Hybrid Oxygen Injection Control for EAF Post-Combustion Using Measured and Predicted Values and Electric Arc Furnace Control System

Abstract

A hybrid oxygen injection control method based on actual and predicted values ​​for electric furnace afterburning according to an embodiment of the present invention is performed in an electric furnace control system and comprises: receiving process data including at least one of actual values ​​of exhaust gas components and process operating variables from an electric furnace process; confirming a predicted value for one or more future points in time calculated based on the process data; calculating a feedback control component for oxygen injection control based on a target value for a control target and the actual values; calculating a feedforward control component for estimating the required amount of oxygen based on the predicted values; and determining an oxygen injection command based on the feedback control component and the feedforward control component.

Description

Technology Field

[0001] The present invention relates to the field of electric furnace afterburning control, and more specifically, to a hybrid control technology that determines oxygen injection using actual and predicted values ​​of exhaust gas components. More specifically, the invention relates to an electric furnace control technology for improving afterburning efficiency, operational stability, and responsiveness by generating a mixed oxygen injection command through weighted combination of a feedback control component based on a target value and actual value for a control objective and a feedforward control component based on a predicted value, and by calculating a final oxygen injection command by applying upper and lower limits and rate of change limits according to equipment constraints. Background Technology

[0002] Electric furnace afterburner control belongs to a technical field that aims to achieve energy efficiency, refractory protection, and compliance with emission regulations by optimizing oxygen input while considering the dynamics of carbon monoxide and carbon dioxide in the exhaust gas.

[0003] Conventionally, feedback control based on actual measured values ​​of exhaust gas components was the mainstream, and it was common practice to combine a certain level of feedforward correction using empirical rules. While this configuration has the advantage of being simple to implement, it has limitations such as being prone to lagging control due to sensor delay, noise, and rapid state changes caused by process step transitions, as well as causing overcorrection and vibration.

[0004] Furthermore, conventional feedforward applications generally relied on fixed constants or simple estimates of a single horizon, and procedures for calculating oxygen levels by reflecting prediction reliability or stoichiometrically interpreting alignment with control objectives were not sufficiently systematized. Consequently, there was a lack of principles regarding how to combine feedback and feedforward control components and how to adjust the proportion of this combination based on operating conditions or prediction uncertainty; as a result, there were issues with reduced consistency and reproducibility in oxygen input decisions.

[0005] In conventional technology, wrapping considering equipment constraints—specifically, the saturation processing of upper and lower limits and rate of change restrictions—often failed to organically link with the output of control commands. Significant trade-offs arise: if saturation and rate of change restrictions are applied loosely retrospectively, tracking of control objectives deteriorates; conversely, if applied excessively conservatively, opportunities for afterburning are missed. Furthermore, despite variations in response across equipment and differing characteristics at each operating stage, there is a limitation in the lack of a system for adjusting constraints in a state-dependent manner. Prior art literature

[0006] Republic of Korea Published Patent Application No. 10-2024-0058140 (Date of publication: May 3, 2024) The problem to be solved

[0007] The present invention aims to overcome the limitations of the background technology described above. An embodiment according to the present disclosure aims to solve the problem of lagging control, overcorrection, and vibration caused by sudden changes in exhaust gas components, sensor delays, and time axis misalignment between signals in an electric furnace process. To this end, an embodiment according to the present disclosure aims to simultaneously improve afterburning efficiency, operational stability, and responsiveness by calculating predicted values ​​for multiple different future points in time using process time series data and determining oxygen input by utilizing the predicted values ​​and actual values ​​together.

[0008] An embodiment according to the present disclosure aims to ensure the reliability and consistency of prediction signals by aligning process data in time and correcting for delays, and by generating derived features suitable for process steps, and to provide operating guidelines that simultaneously consider the near future and the slightly more distant future by calculating exhaust gas components and post-combustion values ​​(PCD) for multiple prediction intervals. In addition, it aims to stably maintain prediction quality during long-term field operation by providing an operation-integrated prediction pipeline that detects degradation of prediction performance or drift in data distribution to trigger retraining, and falls back to a baseline model when necessary.

[0009] Another embodiment according to the present disclosure aims to enable preemptive response while suppressing vibrations even during rapid state changes by generating a mixed oxygen injection command through weighted combination of a feedback control component based on actual values ​​and a feedforward control component based on predicted values ​​in terms of control, and by consistently applying upper and lower saturation and rate of change limits reflecting equipment constraints to calculate a final oxygen injection command. In particular, the objective is to achieve a balance between stability and responsiveness by adaptively adjusting weights according to a prediction reliability index, increasing the weight of preemptive compensation when the prediction is reliable, and prioritizing feedback following when it is uncertain.

[0010] Another embodiment according to the present disclosure aims to ensure that feedforward control aligns with process physics by specifying the control target as an afterburning index or a corresponding control indicator, and by providing a procedure for calculating the stoichiometrically required amount of oxygen based on predicted exhaust gas components and exhaust gas flow rates or their predicted values. Through this, the objective is to overcome the limitations of conventional feedforward control that relies on simple constant corrections and to improve quantitative tracking of the target indicator.

[0011] Furthermore, the embodiments according to the present disclosure aim to enable customized control that reflects the response deviation of each facility and the characteristics of each stage by allowing weights, upper and lower limits, and rate of change limits to be set and applied in a state-dependent manner according to operating state variables, such as whether the loop is closed, the charging stage, the stability of the exhaust gas flow rate, and the power input state. By doing so, the objective is to reduce the risk of damage to refractories and electrodes caused by excessive oxygen input without missing the opportunity for afterburning, and to simultaneously improve energy efficiency and facility lifespan.

[0012] However, the problems to be solved in this disclosure are not limited to those mentioned above, and can be expanded in various ways without departing from the spirit and scope of this disclosure. means of solving the problem

[0014] One technical aspect of the present invention proposes a hybrid oxygen injection control method based on actual and predicted values ​​for electric furnace afterburning. The method is performed in an electric furnace control system and comprises: receiving process data including at least one of actual values ​​of exhaust gas components and process operating variables from an electric furnace process; confirming a predicted value for one or more future points in time calculated based on the process data; calculating a feedback control component for oxygen injection control based on a target value for a control target and the actual values; calculating a feedforward control component for estimating the required amount of oxygen based on the predicted values; and determining an oxygen injection command based on the feedback control component and the feedforward control component.

[0015] In one embodiment, the step of calculating the feedback control component may include: defining an error signal based on the difference between the target value and the actual value; and applying a gain corresponding to at least one of a proportional term, an integral term, and a derivative term to the error signal, and calculating the feedback control component by summing the terms to which the gain is applied.

[0016] In one embodiment, the step of calculating the feedforward control component may include: a step of calculating the amount of oxygen required to achieve the target value according to stoichiometry based on the future exhaust gas component and afterburning index calculated from the predicted value and the predicted value of the exhaust gas flow rate included in the process data; and a step of calculating the feedforward control component using the calculation results of the amount of oxygen for each of one or more future time points.

[0017] In one embodiment, the step of determining the oxygen injection command includes the step of calculating a mixed oxygen injection command by weighted combining the feedback control component and the feedforward control component; wherein the weighted combination is defined according to the following mathematical formula according to weight α, and Here, O2 Mix is a mixed oxygen injection command, O2 FF is a feedforward control component, O2 PID is a feedback control component, and α is a weight between 0 and 1.

[0018] In one embodiment, the weight α is calculated based on a prediction reliability index, and the prediction reliability index may be calculated based on at least one of the prediction variance, the width between quantiles, or the residual in the recent window.

[0019] In one embodiment, the step of determining the oxygen injection command may include the step of saturating the mixed oxygen injection command by applying upper and lower limits according to equipment constraints.

[0020] In one embodiment, the step of determining the oxygen injection command may further include the step of calculating the final oxygen injection command by applying a rate of change limit to the difference between the oxygen injection command at the previous time and the mixed oxygen injection command.

[0021] In one embodiment, the weight α, the upper and lower limits and the rate of change limit are set and applied in a state-dependent manner according to an operating state variable, and the operating state variable may include at least one of whether the loop is closed, the charging stage, the exhaust gas flow rate stability, or the power input state.

[0022] Another technical aspect of the present invention proposes an electric furnace control system. The electric furnace control system comprises at least one processor; and a memory for storing instructions. When the instructions are executed individually or collectively by the at least one processor, the processor performs the following operations: receiving process data including at least one of a measured value of an exhaust gas component and a process operating variable from an electric furnace process; verifying a predicted value for one or more future points in time calculated based on the process data; calculating a feedback control component for oxygen input control based on a target value for a control goal and the measured value; calculating a feedforward control component for estimating the required amount of oxygen based on the predicted value; and determining an oxygen input command based on the feedback control component and the feedforward control component. Effects of the invention

[0024] According to various embodiments of the present disclosure, by using process time series data of an electric furnace process to predict exhaust gas components and post-combustion index (PCD) for multiple different future points in time, and by using both the predicted values ​​and the actual values ​​to determine oxygen injection, the opportunity for post-combustion is preemptively captured and over-correction and vibration caused by subsequent control are suppressed, thereby improving post-combustion efficiency, operational stability, and responsiveness.

[0025] Furthermore, according to various embodiments of the present disclosure, the reliability and consistency of the prediction signal are enhanced by time alignment, delay correction, and the generation of derived features suitable for process steps. Accordingly, control delays caused by sensor delays and time-axis misalignment between signals are mitigated, and prediction-based preemptive compensation becomes possible even in rapid change sections, thereby improving the performance of tracking target indicators.

[0026] In addition, according to various embodiments of the present disclosure, by including an operation procedure that monitors prediction performance degradation and data distribution drift to trigger retraining and falls back to a baseline model upon performance degradation, there is an effect of ensuring stable maintenance of prediction quality and service continuity during long-term operation.

[0027] In addition, according to various embodiments of the present disclosure, a feedforward control component is configured by calculating the required amount of oxygen according to stoichiometry from a predicted value, and a mixed oxygen injection command is generated by weighted combination with a feedback control component based on actual values, thereby ensuring that the feedforward matches process physics and complementarity with the feedback is secured, so that the reduction of overshoot / undershoot of the transient response and the improvement of steady-state precision are achieved simultaneously.

[0028] In addition, according to various embodiments of the present disclosure, by consistently applying upper and lower saturation limits and rate of change limits that reflect equipment constraints to mixed oxygen injection commands, the risk of damage to refractories and electrodes due to abrupt command changes is reduced, and the energy loss and emission burden resulting from excessive oxygen injection are reduced.

[0029] In addition, according to various embodiments of the present disclosure, weight α is adaptively adjusted according to prediction reliability indicators (prediction variance, interval width, recent residual, etc.) so that when the prediction is reliable, the weight of preemptive compensation is increased, and when it is uncertain, feedback following is prioritized, thereby having the effect of automatically maintaining a balance between vibration suppression and preemptiveness. This ensures stable control quality even under various operating regimes and equipment deviations.

[0030] In addition, according to various embodiments of the present disclosure, since prediction intervals (quantiles) and reliability are provided along with point prediction of multiple prediction intervals, it is possible to quantitatively support risk perception and decision-making in the operator interface and high-level optimization logic, and there is an effect that it can also be utilized for state-dependent adjustment of control parameters (weights, upper and lower limits, rate of change limits).

[0031] In addition, according to various embodiments of the present disclosure, process data quality management (handling missing values ​​and outliers, signal synchronization) and model operation (monitoring, fallback, retraining) are integrated into a single pipeline, thereby reducing the burden of operational management and maintaining data consistency between prediction and control, which has the effect of improving decision consistency.

[0032] The effects obtainable from the present disclosure are not limited to those mentioned above, and other unmentioned effects will be obvious to those skilled in the art to which the present disclosure pertains in light of the following detailed description and drawings. Brief explanation of the drawing

[0034] FIG. 1 is a block diagram illustrating an electric furnace control system according to one embodiment of the present invention. FIG. 2 is a flowchart illustrating a method for predicting electric furnace exhaust gas components based on multiple prediction intervals according to an embodiment of the present invention. FIG. 3 is a block diagram illustrating the detailed configuration of a prediction server according to one embodiment of the present invention. FIG. 4 is a flowchart illustrating a time alignment and delay correction procedure according to an embodiment of the present invention. FIG. 5 is a flowchart illustrating the procedure for outputting multiple prediction intervals and calculating and providing reliability according to an embodiment of the present invention. FIG. 6 is a flowchart illustrating a performance monitoring, fallback transition, and relearning management procedure according to an embodiment of the present invention. FIG. 7 is a flowchart illustrating a state-aware regime switching ensemble according to one embodiment of the present invention. FIG. 8 is a flowchart illustrating a hybrid oxygen injection control method based on actual and predicted values ​​for electric furnace afterburning according to one embodiment of the present invention. FIG. 9 is a block diagram illustrating the detailed configuration of a control server according to one embodiment of the present invention. FIG. 10 is a flowchart illustrating a feedforward oxygen amount calculation procedure according to one embodiment of the present invention. FIG. 11 is a flowchart illustrating a state-dependent parameter scheduling procedure according to one embodiment of the present invention. FIG. 12 is a flowchart illustrating a mixing command generation and constraint processing procedure in one embodiment of the present invention. FIG. 13 is a flowchart illustrating the burner and lance unit distribution optimization and sensitivity feedback procedure according to one embodiment of the present invention. Specific details for implementing the invention

[0035] Preferred embodiments of the present invention will be described below with reference to the attached drawings.

[0036] However, embodiments of the present invention may be modified in various different forms, and the scope of the present invention is not limited to the embodiments described below. Furthermore, embodiments of the present invention are provided to more fully explain the present invention to those with average knowledge in the relevant technical field.

[0037] That is, the aforementioned objectives, features, and advantages will be described in detail below with reference to the attached drawings, and accordingly, a person skilled in the art to which the present invention pertains will be able to easily implement the technical concept of the present invention. In describing the present invention, detailed descriptions of known technologies related to the present invention are omitted if it is determined that such descriptions may unnecessarily obscure the essence of the present invention. Hereinafter, preferred embodiments according to the present invention will be described in detail with reference to the attached drawings. In the drawings, the same reference numerals are used to indicate the same or similar components.

[0038] Additionally, singular expressions used in this specification include plural expressions unless the context clearly indicates otherwise. In this application, terms such as "composed of" or "comprising" should not be interpreted as necessarily including all of the various components or steps described in the specification, and should be interpreted as meaning that some of the components or steps may not be included, or that additional components or steps may be included.

[0039] In addition, various components and their sub-components are described below to explain the system according to the present invention. These components and their sub-components may be implemented in various forms, such as hardware, software, or a combination thereof. For example, each element may be implemented as an electronic configuration to perform a corresponding function, or as software itself that is executable in an electronic system, or as a functional element of such software. Alternatively, it may be implemented as an electronic configuration and corresponding executable software.

[0040] The various techniques described herein may be implemented with hardware or software, or, where appropriate, with a combination of all of these. Terms such as "Unit," "Server," and "System" as used herein may likewise be treated as equivalent to computer-related entities, namely hardware, a combination of hardware and software, software, or software at runtime. Furthermore, each function executed in the system of the present invention may be configured as a module and may be recorded in a single physical memory or distributed among two or more memories and recording media.

[0041] Various flowcharts are disclosed to explain embodiments of the present invention, but these are for the convenience of explaining each step and do not necessarily mean that each step must be performed in the order of the flowchart. That is, each step in the flowchart may be performed simultaneously, in the order according to the flowchart, or in the reverse order of the flowchart.

[0043] The key terms used in this specification have the following meanings. The technical concept of the present invention is not limited to the following definitions of terms, and unless otherwise defined in the context, the meanings defined in this specification shall prevail.

[0044] "Process data" and "process time series data" are collective terms for data acquired in chronological order in an electric furnace process, and include at least one of process operating variables such as actual values ​​of exhaust gas components (e.g., CO, CO₂), exhaust gas flow rate, power input, burner / lance status, and charging stage markings.

[0045] "Multi-Horizon" refers to a set of multiple different future time points (e.g., +15s, +30s, +60s, etc.).

[0046] "Exhaust gas components" include at least carbon monoxide (CO) and carbon dioxide (CO₂), and, if necessary, O₂ and NO X It may include additional ingredients such as...

[0047] The "Post-burning Discharge (PCD)" is a dimensionless indicator representing the degree to which CO is converted into CO₂.

[0048] "Prediction interval / quantile" is a quantitative value representing the uncertainty of the prediction distribution, and refers to a quantile (q10, q50, q90, etc.) or a prediction interval derived therefrom.

[0049] The "prediction reliability index" is a value that quantifies the reliability of a prediction and can be calculated as a function of prediction variance, interquantile range (IQR), and recent residuals.

[0050] "Performance indicators" are error-based indicators calculated based on predicted values ​​and corresponding actual values, and include window-based indicators such as MAE, RMSE, and MAPE.

[0051] "Drift indicators" is a general term for indicators used to detect changes in data distribution (e.g., PSI, KS-statistics, mean / variance shift, etc.).

[0052] A "baseline model" refers to an alternative model (linear, rule-based, simple time series, etc.) that differs in structure or parameters from the forecasting model.

[0053] "Fallback transition" refers to a procedure that switches to providing predicted values ​​as the output of the baseline model in the event of conditions such as performance degradation.

[0054] "Feedback control component" refers to a control quantity calculated by applying at least one of proportional, integral, or derivative gain to an error signal and summing the terms to which the gain is applied.

[0055] "Feedforward control component" refers to a control amount obtained by calculating and accumulating (including numerical integration) the amount of oxygen required to achieve a target value according to stoichiometry (e.g., 2CO+O₂→2CO₂) at future points in time based on the predicted exhaust gas component / PCD and exhaust gas flow rate or the predicted value thereof.

[0056] The "mixed oxygen injection command" is a control quantity formed by the weighted combination of feedforward and feedback control components.

[0057] The "final oxygen injection command" is a control quantity obtained by sequentially applying upper and lower saturation limits and rate of change limits to the mixed oxygen injection command.

[0059] FIG. 1 is a block diagram illustrating an electric furnace control system according to one embodiment of the present invention.

[0060] Figure 1 shows a simplified example for the sake of understanding, and configurations and signals not illustrated are also described below.

[0061] Referring to FIG. 1, exhaust gas generated from an electric furnace (10) is introduced into a gas detector (12) to measure components such as CO and CO₂ in real time, and the measurement result is transmitted to a controller (16). An oxygen supply unit (14) is a device including a burner and / or an oxygen lance, and supplies oxygen to the electric furnace (10) according to an oxygen injection command from the controller (16). The controller (16) may be a field PLC or an equivalent industrial controller.

[0062] The controller (16) communicates bidirectionally with the electric control system (100).

[0063] For example, the electric furnace control system (100) can perform the function of collecting actual values ​​of exhaust gas components (CO, CO₂) and process operation variables (exhaust gas flow rate, power, charging stage, equipment status, etc.) from the electric furnace (10) through a controller (16) and storing them as standardized process time series data.

[0064] For example, an electric furnace control system (100) can perform timestamp normalization, time alignment between signals, and delay correction on collected process time series data (including cross-correlation-based delay estimation), and generate derived features such as lag, moving average, difference, cumulative amount, and process step identifier.

[0065] For example, the electric furnace control system (100) can perform the function of calculating predicted values ​​of exhaust gas components (CO, CO₂) for multiple prediction intervals (multiple different future points in time) through a prediction server (110), receiving from an external source to verify the predicted values, and calculating the afterburn index (PCD) from the predicted values.

[0066] For example, an electric furnace control system (100) can calculate a prediction reliability index based on a prediction interval (quantile) or a prediction variance.

[0067] For example, the electric furnace control system (100) can perform the function of monitoring prediction performance by calculating a window-based error indicator in the interval where predicted values ​​and actual values ​​are secured, and determining whether performance deterioration has occurred by calculating a data distribution drift indicator in parallel.

[0068] For example, the electric furnace control system (100) can perform the function of triggering retraining of the prediction model when performance degradation persists, drift is detected, or a cycle arrives, and continuing to provide prediction values ​​by falling back to the baseline model when necessary.

[0069] For example, the electric furnace control system (100) can perform the function of defining an error signal based on the difference between the target value and the actual value for a control target at the control server (120), and calculating a feedback control component by summing terms to which at least one of proportional, integral, or derivative gain is applied.

[0070] For example, the electric furnace control system (100) can perform the function of calculating a feedforward control component by calculating and accumulating the amount of oxygen required for each future section according to stoichiometry (e.g., 2CO+O₂→2CO₂) based on the predicted exhaust gas component / PCD and the exhaust gas flow rate or the predicted value.

[0071] For example, the electric furnace control system (100) can generate a mixed oxygen injection command by combining a feedforward control component and a feedback control component with a weight α, and perform the function of adaptively adjusting α according to a prediction reliability index or an operating state variable.

[0072] For example, the electric furnace control system (100) can perform the function of sequentially applying upper and lower saturation and change rate limits reflecting equipment constraints to a mixed oxygen injection command to calculate a final oxygen injection command, and transmitting it to a controller (16) to execute it on an oxygen supply unit (14).

[0073] For example, the electric furnace control system (100) can perform the function of maintaining the stability and responsiveness of the control in a balanced manner by setting and applying weights α, upper and lower limits and a rate of change limit in a state-dependent manner according to operating state variables such as whether the loop is closed, the charging stage, the stability of the exhaust gas flow rate, and the power input state.

[0074] For example, the electric furnace control system (100) can perform the function of visualizing the predicted value, PCD and reliability, mixed / final oxygen injection command and judgment basis (pharmaceutical application history, α value change, etc.) on the HMI and recording it in a storage along with logs and events to ensure traceability and auditability.

[0075] The electric furnace control system (100) may include a prediction server (110) and a control server (120).

[0076] The prediction server (110) performs multiple prediction interval predictions, and the control server (120) performs hybrid oxygen injection control. The prediction server (110) and the control server (120) may be implemented as logic modules within a single physical device or may be separated into different physical servers.

[0077] The prediction server (110) receives process time series data (CO and CO₂ component values ​​of the gas detector (12), equipment status, power information, etc.) collected through the controller (16) as input and calculates prediction values ​​for multiple prediction intervals for multiple different future points in time.

[0078] According to one embodiment, the prediction server (110) outputs exhaust gas components (CO, CO₂) at future time points and a post-combustion index (PCD) calculated from them. The post-combustion index can be defined, for example, as in [Equation 1] below.

[0079] [Mathematical Formula 1]

[0080]

[0081] The prediction server (110) may, if necessary, calculate and provide to the control server (120) prediction intervals (quantiles) or prediction reliability indicators (e.g., width between quantiles, prediction variance, based on recent residuals). The prediction reliability indicators may be used as a basis for setting weights when combining weights as described later.

[0082] The control server (120) uses the predicted value provided by the prediction server (110) and the actual value provided by the controller (16) together to calculate the actual-based feedback control component (e.g., selected from P / PI / PD / PID) and the prediction-based feedforward control component, and generates a mixed oxygen injection command (O2_Mix) by weighted combining them. The mixed oxygen injection command can be defined, for example, as shown in [Equation 2] below.

[0083] [Mathematical Formula 2]

[0084]

[0085] Here, is a feedforward control component, is a feedback control component, is a weight between 0 and 1.

[0086] The control server (120) sequentially applies upper and lower saturation (clip) and rate limit (rate_limit) reflecting facility constraints to the mixed oxygen injection command to calculate the final oxygen injection command (O2_Final) and transmits it to the controller (16). The formula related to this can be defined as [Equation 3] below.

[0087] [Mathematical Formula 3]

[0088]

[0089] The controller (16) applies the final oxygen injection command to the oxygen supply unit (14) to inject oxygen into the electric furnace (10).

[0090] Explains signal paths and data flow.

[0091] The measurement path is formed in the order of the electric furnace (10), gas detector (12), controller (16), and electric furnace control system (100), and the prediction path transmits the predicted values ​​(future CO, CO₂, PCD) of multiple prediction intervals and optional reliability indicators from the prediction server (110) to the control server (120). The command path transmits the final oxygen injection command from the control server (120) to the oxygen supplier (14) through the controller (16), and the result is applied to the electric furnace (10).

[0092] To explain an example of the operation based on FIG. 1, when the measurement value of the gas detector (12) is transmitted to the prediction server (110) through the controller (16), the prediction server (110) calculates the predicted values ​​of CO, CO₂, and PCD for multiple prediction intervals. The control server (120) uses the predicted values ​​from the prediction server (110) and the actual measured values ​​provided by the controller (16). and Calculate and combine them by weight. It generates. Afterwards, the control server (120) By applying upper and lower saturation and rate of change limits to Obtain the and transmit it to the controller (16) to execute the oxygen supply device (14).

[0093] Each component illustrated in FIG. 1 may be implemented as a single device or as a distributed unit. The prediction server (110) and the control server (120) may be software modules coexisting on a single computing hardware. Hereinafter, the prediction server (110) and the control server (120) are collectively referred to as the electric furnace control system (100), but according to the description, the electric furnace control system (100) may include the functions of either the prediction server (110) or the control server (120), or both. The gas detector (12) may be a single sensor or a group of multiple sensors distributed in the upper and lower ducts. The oxygen supply unit (14) may be a combination of multiple burners and lances, and the controller (16) may be configured to perform individual / group control of these actuators.

[0095] Hereinafter, with reference to FIGS. 2 to 13, various embodiments of an electric furnace control system (100) according to the present invention and an electric furnace control method performed therein will be described.

[0097] FIG. 2 is a flowchart illustrating a method for predicting electric furnace exhaust gas components based on multiple prediction intervals according to an embodiment of the present invention.

[0098] Each step illustrated in FIG. 2 can be performed in parallel or repeatedly depending on the situation, and variations are possible within the scope of the same technical concept unless the order is specifically limited.

[0099] In step S210, the prediction server (110) receives process time series data including actual values ​​of exhaust gas components and at least one of process operating variables from the electric furnace (10) through the electric furnace control system and loads it into a buffer.

[0100] In step S220, the prediction server (110) can perform time alignment and delay correction between signals based on the timestamps of the process time series data.

[0101] Specifically, the prediction server (110) normalizes the timestamps of the process time series data and aligns the sampling intervals. Subsequently, the prediction server (110) estimates the relative delay based on cross-correlation analysis between signals and records the estimated delay value. The prediction server (110) applies a variable buffer according to the delay value obtained in this way to perform time alignment between signals and delay correction.

[0102] In step S230, the prediction server (110) generates derived features for improving prediction performance from time-aligned and lag-corrected data. As an example, the derived features include at least one of a lag, a moving average, a difference, a cumulative amount, and a process step identifier.

[0103] In step S240, the prediction server (110) provides the derived feature and the process time series data to the prediction model to calculate the predicted values ​​of the exhaust gas components for each of the multiple future time points corresponding to multiple prediction intervals.

[0104] In one embodiment, the prediction server (110) calculates one or more quantiles or prediction intervals for the prediction value calculated in step S240. Subsequently, the prediction server (110) calculates a prediction reliability index based on the results of the calculated data or recent residuals.

[0105] The prediction server (110) calculates the post-combustion index (PCD) using the CO and CO₂ predicted in step S240. The post-combustion index can be defined by the aforementioned [Equation 1].

[0106] Subsequently, the prediction server (110) performs post-processing by applying physical consistency constraints to the predicted values ​​of exhaust gas components for each of the afterburning index and multiple different future time points. The physical consistency constraints include non-negative values ​​of CO and CO₂ and a range of 0 to 1 for the afterburning index.

[0107] In step S250, the prediction server (110) stores or provides the aforementioned prediction value and the derived indicator calculated therefrom. The derived indicator may include a post-burn index and a prediction reliability index. For example, the prediction server (110) may provide the prediction value, the post-burn index, and the prediction reliability index to the control server (120) and the operator interface.

[0108] In step S260, the prediction server (110) forms a predetermined time window after the actual value corresponding to the prediction value is obtained to match the prediction value with the actual value, calculates an error-based performance indicator based on this, and monitors the trend of the error-based performance indicator. The prediction server (110) determines whether performance deterioration has occurred by comparing the error-based performance indicator with a reference value.

[0109] In step S270, the prediction server (110) detects whether there is a change in the distribution of process data based on the drift indicator. Subsequently, if performance degradation is determined in step S260 or drift is detected in step S270, the prediction server (110) performs a fallback transition to replace the predicted value with the output of a baseline model having a different structure or parameter configuration from the prediction model.

[0110] In step S280, the prediction server (110) triggers retraining of the prediction model when at least one of the conditions of the duration of a performance degradation state, detection of drift, or arrival of a predefined retraining cycle is satisfied, i.e., when the retraining condition of the prediction model is satisfied.

[0111] Subsequently, the prediction server (110) performs data splitting considering leakage prevention for retraining, proceeds with training and verification, and registers a new model as a candidate when the verification performance satisfies a threshold. Subsequently, the prediction server (110) deploys the new model, intensively monitors performance indicators during the initial period after deployment, and performs a rollback if necessary.

[0112] The prediction server (110) continues to perform such procedures and periodically repeats at least a part of the entire procedure to continuously provide the predicted values ​​of multiple prediction intervals, the post-burn index, and the prediction reliability index to the control server (120).

[0114] FIG. 3 is a block diagram illustrating the detailed configuration of a prediction server according to one embodiment of the present invention.

[0115] The prediction server (110) illustrated in FIG. 3 includes the functions of collecting and structuring process time series data, time alignment and delay correction, generating derived features, predicting components in multiple prediction intervals, calculating post-burning index and physical consistency post-processing, monitoring prediction performance and fallback switching, managing relearning, and storing or providing data.

[0116] Referring to FIG. 3, the prediction server (110) may be composed of a collection module (111), an alignment correction module (112), a derived feature generation module (113, hereinafter referred to as the derived feature generation module), a component prediction module (114), a post-burning index calculation module (115), a performance monitoring module (116), a retraining management module (117), an interface module (118), a data repository, and a model repository (not shown).

[0117] The collection module (111) receives process time series data including actual values ​​of exhaust gas components and at least one of process operating variables through a controller and a field gateway, loads the data into a buffer while preserving timestamps, and performs lightweight preprocessing including missing value correction and outlier marking.

[0118] The alignment correction module (112) performs timestamp normalization and sampling interval alignment, estimates relative delay based on cross-correlation analysis between signals, etc., and performs time alignment and delay correction by applying a variable buffer according to the estimated delay.

[0119] The derived feature generation module (113, derived feature generation module) generates a derived feature including at least one of a lag, a moving average, a difference, a cumulative amount, and a process step identifier, and combines the generated feature with the raw input to provide a predicted input.

[0120] The component prediction module (114) receives time-aligned and delay-corrected data and derived features as input and calculates predicted values ​​of exhaust gas components for multiple different future points in time. If necessary, one or more quantiles or prediction intervals may be calculated, and a prediction reliability index may be calculated based on the width between quantiles or the prediction variance and recent residuals.

[0121] The post-combustion index calculation module (115) calculates the post-combustion index (PCD) using the predicted CO and predicted CO₂ calculated by the component prediction module (114). Additionally, it performs post-processing to apply physical consistency constraints to the predicted values ​​and the post-combustion index (PCD) to satisfy the non-negative nature of CO and CO₂ and the range of PCD from 0 to 1.

[0122] The performance monitoring module (116) configures a predetermined time window after the actual value corresponding to the predicted value is obtained, calculates an error-based performance indicator to determine whether performance has deteriorated, and detects whether there is a change in data distribution according to the drift indicator.

[0123] The retraining management module (117) triggers the retraining of the prediction model under at least one of the conditions of performance degradation persisting for a certain period, drift detection, or the arrival of a predefined retraining cycle, and manages training, validation, deployment, and rollback if necessary. Additionally, if performance degradation is determined or drift is detected, it performs a fallback transition to replace the predicted value with the output of a baseline model having a different structure or parameter configuration from the prediction model.

[0124] The interface module (118) stores or provides a predicted value and a derived indicator calculated from the predicted value, and transmits the result to the control server and the operator interface.

[0125] The data store and model store (not shown) store input data, predicted values, PCD, prediction reliability indicators, performance indicators, and models, and manage the conversion history of deployed models and candidate models. The store may be embedded within the prediction server (110) or implemented as a separate upper-level storage system.

[0126] The prediction server (110) configured in this manner produces control-friendly prediction results through a series of procedures including collection, alignment and correction, generation of derived features, multiple prediction interval prediction, PCD calculation and physical consistency post-processing, performance monitoring and drift detection, fallback transition and retraining management, storage or provision, and reliably provides them to the control server.

[0128] FIG. 4 is a flowchart illustrating a time alignment and delay correction procedure according to an embodiment of the present invention.

[0129] One embodiment of FIG. 4 is performed by an alignment correction module (112) within a prediction server (110) and includes steps for aligning process time series data having different sample intervals and transmission delays to a common time axis and correcting the delay.

[0130] Referring to FIG. 4, in step S410, the alignment correction module (112) receives raw time series data from the collection module (111) including actual values ​​of exhaust gas components and at least one of process operation variables.

[0131] In step S420, the alignment correction module (112) performs time alignment between signals based on timestamps. Specifically, the alignment correction module (112) normalizes the timestamps of the input data into a standard format, unifies the time zone and reference time, and marks unreliable timestamps. Subsequently, the alignment correction module (112) identifies the sample interval per channel, performs resampling or interpolation to match the target sample interval, and converts the irregular sample interval into a uniform interval.

[0132] In step S430, the alignment correction module (112) detects missing intervals and outlier intervals, generates a mask, and corrects according to defined rules. Timestamp consistency is maintained even after processing outliers and missing values. Specifically, the alignment correction module (112) detects missing intervals and outlier intervals and generates a mask. The alignment correction module (112) can perform corrections such as linear interpolation or maintaining adjacent values ​​according to missing value processing rules.

[0133] In step S440, the alignment correction module (112) calculates the cross-correlation between the reference signal and the target signal to calculate an indicator for estimating the relative delay within the candidate delay range.

[0134] For example, a reference signal in continuous time and target signal cross-correlation regarding It can be defined as [Equation 4] below, and in discrete time as [Equation 5] below.

[0135] [Mathematical Formula 4]

[0136]

[0137] [Mathematical Formula 5]

[0138]

[0139] Here, represents the continuous-time waveform of the reference signal, and represents the continuous-time waveform of the target signal, represents the relative delay in continuous time. refers to cross-correlation in the analysis window, and class represents the discrete time series sample values ​​of the reference signal and the target signal, respectively, and means discrete shift.

[0140] In step S450, the alignment correction module (112) determines the delay value at which the above-described indicator becomes maximum as the delay estimate value and records the said delay value as metadata.

[0141] For example, the delay estimate can be determined as shown in [Equation 6] below.

[0142] [Mathematical Formula 6]

[0143]

[0144] In one embodiment, the alignment correction module (112) ensures the stability of the delay estimation by adjusting the delay estimation range and the calculation window length according to the operating state, such as a process step identifier. Additionally, the fixed response delay of the exhaust gas analyzer can be reflected as a correction constant, and the fixed response delay When set to, the total correction delay It can be defined as.

[0145] Here, represents the relative lag where the cross-correlation has a maximum value, and represents the physical response delay of the sensor, and represents the total delay correction amount applied to the actual alignment.

[0146] In one embodiment, the alignment correction module (112) ensures the stability of the delay estimation by adjusting the delay estimation range and the calculation window length according to the operating state, such as a process step identifier.

[0147] In step S460, the alignment correction module (112) performs alignment by applying a variable buffer according to the delay estimate to advance or delay the channel-specific signal on the time axis.

[0148] In this case, the sample interval in the discrete-time implementation is When set to, integer or fractional offset By applying It can be sorted as, If is not an integer, sample values ​​can be estimated using linear interpolation or spline interpolation. If necessary, the window length to reduce residual high-frequency noise A moving average filter for can be applied as shown in [Equation 7] below.

[0149] [Mathematical Formula 7]

[0150]

[0151] Here, represents the time interval between discrete samples, and represents the shift amount converted from the total delay into the number of samples, and represents the raw sample before alignment, represents an aligned sample with delayed correction applied. represents the window length of the moving average, and represents a relative index within the window. The length is It refers to the moving average result of.

[0152] In one embodiment, the alignment correction module (112) can reconstruct the frame to align the alignment results with a common time axis and perform synchronization so that all channels have the same timestamp sequence.

[0153] In step S470, the alignment correction module (112) calculates an alignment quality index, which consists of a correlation coefficient after alignment, a remaining time error, and a ratio of successful alignment intervals, in order to evaluate the alignment quality. Subsequently, the alignment correction module (112) determines whether the alignment is successful by comparing the alignment quality index with a preset threshold value. If the alignment quality index is below the threshold value, as a post-processing step for alignment failure, a conservative rule is applied, such as maintaining the previous delay value, relaxing the delay correction, or excluding the corresponding interval from the prediction input.

[0154] Subsequently, the alignment correction module (112) provides the data set with completed alignment and delay correction to the derived feature generation module (113) and the component prediction module (114).

[0155] Additionally, the alignment correction module (112) logs the delay estimate, alignment quality indicator, and applied correction rule, and stores them in a data repository and a model repository (not shown) as needed.

[0156] According to this procedure, the alignment correction module (112) aligns process time series data having heterogeneous time axes based on a common standard and corrects delays between signals, thereby providing an input that improves prediction performance and consistency in the derived characteristic generation module (113) and the component prediction module (114).

[0158] FIG. 5 is a flowchart illustrating the procedure for outputting multiple prediction intervals and calculating and providing reliability according to an embodiment of the present invention.

[0159] One embodiment illustrated in FIG. 5 is performed by a component prediction module (114), a post-burning index calculation module (115), and an interface module (118) within a prediction server (110).

[0160] Referring to FIG. 5, in step S510, the component prediction module (114) receives an input frame from the alignment correction module (112) and the derived feature generation module (113) in which time alignment and delay correction are completed and derived features are combined.

[0161] In step S520, the component prediction module (114) calculates point prediction values ​​of CO and CO₂ for each of a plurality of different future time points corresponding to multiple prediction intervals based on the input frame.

[0162] Specifically, the component prediction module (114) calculates one or more quantiles or prediction intervals for each future time point prediction value calculated in step S520. Subsequently, the component prediction module (114) calculates a prediction reliability index based on the calculated quantile interval width or prediction variance and, if necessary, recent residuals.

[0163] In step S530, the post-burning index calculation module (115) calculates the post-burning index for each future point in time based on the predicted CO and predicted CO₂ of step S520. It calculates ), which can be set as shown in [Equation 7] below. [Equation 7] is substantially the same formula as [Equation 1], but at a future point in time It is a mathematical formula that specifies and represents it.

[0164] [Mathematical Formula 7]

[0165]

[0166] Here, is a future point in time It refers to the predicted carbon dioxide value at, is a future point in time It refers to the predicted carbon monoxide value at, represents the predicted value of the afterburning index at that point in time. Here, epsilon (å) is a constant added to prevent the denominator from becoming zero.

[0167] In one embodiment, the post-burning index calculation module (115) may perform post-processing by applying physical consistency constraints to the calculated values ​​of steps S520 and S530. The physical consistency constraints include non-negative properties of CO and CO₂ and a range of 0 to 1 for the post-burning index.

[0168] In step S540, the component prediction module (114) and the post-burning index calculation module (115) construct an output package in which point predictions of multiple prediction intervals, quantiles or prediction intervals, prediction reliability indicators, and PCDs are aligned on the time axis.

[0169] For example, the interface module (118) converts such an output package into visualization data for display on the operator interface. The visualization data may include predicted trend lines and quantile bands for CO and CO₂, and predicted trend lines and ranges for PCD.

[0170] Afterwards, the interface module (118) can provide the output package and visualization data to the control server (120) and display them on the operator interface.

[0171] The interface module (118) can store predicted values ​​and derived indicators in a data store and a model store. The derived indicators may include PCD and prediction reliability indicators.

[0172] According to the above procedure, the prediction server (110) calculates point predictions and quantiles or prediction intervals and prediction reliability indicators for multiple prediction intervals, and a post-burning index for each future time point ( The result of applying physical consistency constraints by calculating ) can be provided to the control server (120) and the operator interface.

[0174] FIG. 6 is a flowchart illustrating a performance monitoring, fallback transition, and relearning management procedure according to an embodiment of the present invention.

[0175] One embodiment of FIG. 6 is performed by a performance monitoring module, a drift detector, a fallback switcher, and a relearning management module within a prediction server (110), and includes steps to maintain the prediction quality in multiple prediction intervals stably for a long period.

[0176] Referring to FIG. 6, in step S610, the prediction server (110) stores the predicted values ​​of the multiple prediction intervals provided in FIG. 5. For example, the prediction server (110) registers the predicted values ​​of the multiple prediction intervals in a monitoring buffer and stores together the identifier and timestamp for each future point in time, the prediction interval or quantile, and the prediction reliability index.

[0177] In step S620, the prediction server (110) aligns the predicted value with the actual value at the time when the actual value is obtained. At this time, the prediction horizon Residuals regarding cast Calculate as. Here, is the point in time Predicted horizon at It refers to the actual measured value corresponding to, represents the predicted value of the same horizon.

[0178] In step S630, the prediction server (110) has a predetermined time window It calculates an error-based performance metric between the predicted and measured values. For example, as an error-based performance metric, the horizon With respect to , the prediction horizon as shown in [Equation 8] below Mean absolute error in , prediction horizon Square root mean square error in and forecast horizon Average absolute percentage error in It can produce.

[0179] [Mathematical Formula 8]

[0180] ,

[0181] ,

[0182]

[0183] Here, means the number of samples included in the window, is a small positive constant to prevent a zero denominator. Also, is the prediction horizon As the mean absolute error in, the unit is the same as the target variable, and is the prediction horizon As the square root mean squared error, it is more sensitive to large errors, and its unit is the same as the target variable, and is the prediction horizon It is the average absolute percentage error, and the unit is percentage.

[0184] In step S640, the prediction server (110) compares the performance indicator of each prediction horizon with a reference value and determines deterioration based on whether the exceedance of the reference value continues continuously. Minimum duration for determining continuity or minimum number of times It is possible to make it insensitive to temporary spikes by defining it.

[0185] In step S650, the prediction server (110) detects whether there is a change in the data distribution based on the drift indicator. That is, the baseline distribution with respect to the distribution of the input features and current distribution Using the proportions of each interval, the Population Stability Index (PSI) is calculated as shown in [Equation 9] below, or KS statistics for the cumulative distribution of specific characteristics are calculated. Here, represents the range index, and and represents the proportion of the corresponding interval in the reference period and the current period, respectively.

[0186] [Mathematical Formula 9]

[0187]

[0188] In step S660, the prediction server (110) determines that the fallback switching condition is satisfied if at least one of the deterioration determination of step S640 or the drift detection of step S650 is satisfied. The fallback switcher determines the switching method to the baseline model output, selects one of variable-unit fallback, horizon-unit fallback, and total fallback, and the fallback holding time Set the return condition together.

[0189] In step S670, the prediction server (110) evaluates the retraining trigger conditions. If at least one of the persistence of performance degradation, exceeding the drift metric, or the arrival of a cycle is satisfied, it transitions to a retraining ready state.

[0190] In one embodiment, the prediction server (110) constructs a retraining dataset. To prevent leakage, the training section and the verification section are separated by preserving the chronological order, and the alignment and delay correction rules of FIG. 4 and the physical consistency post-processing rules of FIG. 5 are applied identically. To ensure representativeness of the feature distribution, the ratio of operating state variables is checked.

[0191] Afterward, the prediction server (110) trains candidate models and evaluates performance indicators and robustness indicators in the validation phase. Candidates satisfying the horizon-specific criteria, overall average criteria, and variance criteria are selected.

[0192] In one embodiment, the prediction server (110) performs shadow mode verification on the candidate model. It calculates predictions in parallel by synchronizing with online input and records the performance comparison without reflecting them in the actual service output.

[0193] If the distribution approval criteria are satisfied, the prediction server (110) performs a gradual distribution. It sequentially expands by limiting it to some sections or horizons, and intensively observes performance indicators in the initial distribution monitoring window.

[0194] If normal performance is observed during the initial deployment period, the prediction server (110) releases or reduces the fallback, and immediately rolls back upon detection of an abnormality. Rollback means switching to the last normal version or baseline model.

[0195] Afterward, the prediction server (110) records fallback transition events, retraining triggers, training and validation settings, deployment and rollback history in an audit log and stores them in the data store and model store as needed.

[0196] According to this procedure, the prediction server (110) can continuously monitor performance in multiple prediction intervals, respond quickly to changes in distribution, and ensure prediction stability and consistency in long-term operation through fallback and relearning.

[0198] FIG. 7 is a flowchart illustrating a state-aware regime switching ensemble according to one embodiment of the present invention.

[0199] One embodiment of FIG. 7 is applied to the multiple prediction interval calculation step of the prediction server (110) to improve prediction stability and preemptiveness in the data distribution transition interval.

[0200] Referring to FIG. 7, in step S710, the prediction server (110) receives an input frame with time alignment and delay correction completed from the alignment correction module (112) and the derived characteristic generation module (113), and combines operating state variables (loop closure status, charging stage, exhaust gas flow rate stability, power input status, etc.) and state indicators derived therefrom. The state indicators may include the recent power input change rate, gas flow rate stability, process stage identifier, and recent residual statistics.

[0201] In step S720, the prediction server (110) applies a regime determination model to the time point regime posterior probability vector in Calculates. Here is the number of defined regimes, and is "the current regime It is a posterior probability. Regimes can be defined, for example, as a high-variability regime at the beginning of charging, a stable power input regime, an unstable exhaust gas flow rate regime, and a regime of reduced afterburning efficiency.

[0202] The regime determination model calculates the posterior probability of which operating regime the current operating state belongs to, and uses this probability to determine which prediction submodel to prioritize and how to smoothly blend the two submodels if a transition is in progress. In other words, regime determination quantifies "how much to trust a particular model" and is directly used for model selection, calculation of blending weights, and transition constraints (hysteresis and dwell time).

[0203] In addition, the prediction server (110) Smoothed with an exponential moving average to mitigate chatter caused by instantaneous fluctuations Calculate. Here is a smoothing coefficient between 0 and 1.

[0204] In step S730, the prediction server (110) performs a regime transition decision by applying hysteresis and dwell time constraints.

[0205] For example, the immediately prior confirmed regime Called, and the candidate transition regime When saying that, as a transition entry condition The state is continuous If maintained above Approve the metastasis.

[0206] As a condition for holding the metastasis If so, the current regime is maintained. Here is a hysteresis condition, is the minimum residence time. At this stage, the regime posterior probability vector itself is used as the input to the transition decision formula.

[0207] In step S740, the prediction server (110) determines the current regime confirmed as the transition decision result Set of submodels by regime based on Main model Select and additionally select one or more auxiliary models based on transition candidates or adjacency criteria. Adjacency can be defined based on past transition frequencies or feature space similarity.

[0208] In step S750, the prediction server (110) selects each submodel Prediction of multiple prediction intervals regarding Calculate in parallel. Here, the prediction of multiple prediction intervals is the point in time regime It is the model's prediction vector and includes exhaust gas components such as CO and CO₂.

[0209] In step S760, the prediction server (110) performs a first blend based on the regime posterior probability if a regime transition is in progress. For example, transition candidate regimes and Transition score for Transition weighting from Calculate and use the following [Equation 10] to submodel prediction vector of Calculates.

[0210] [Mathematical Formula 10]

[0211]

[0212] In a normal state without metastasis As this converges to 0 or 1, the main model output remains unchanged. This becomes a submodel prediction vector of refers to the primary mixture result for the purpose of metastasis mitigation.

[0213] In step S770, the prediction server (110) determines the uncertainty indicator of each submodel. Calculates the uncertainty indicator. The uncertainty indicator can be defined as at least one of the width between predicted quantiles, the predicted variance, or the recent residual variance. This is used for confidence correction.

[0214] In step S780, the prediction server (110) calculates a confidence correction weight combining regime information and uncertainty, and performs a second blend using this. Uncertainty mapping After defining [Equation 11] below, normalized weights are calculated, and finally, a confidence-corrected mixed prediction is produced using [Equation 12] below.

[0215] [Mathematical Formula 11]

[0216]

[0217] [Mathematical Formula 12]

[0218]

[0219] Here, "reliability-adjusted weighting" is Referring to, regime posterior probability Uncertainty correction It is the weight obtained by multiplying. Even in the case of no transition. It is allocated mainly based on the main model It is identical to or close to it. Conversely, if a transition is in progress and the uncertainty of the auxiliary model is high, the contribution of that model is automatically reduced.

[0220] Afterwards, the prediction server (110) predicts a reliability correction mixed prediction. It receives the input and applies physical consistency constraints. Upon receiving a reliability correction mixed prediction, the prediction server (110) first removes physically impossible negative or abnormal values ​​from the CO and CO₂ values ​​to exclude values ​​less than 0, and then matches the consistency between the gas composition (predicted CO, predicted CO₂) and the post-combustion index (PCD). Specifically, if the gas prediction is more reliable, the PCD is recalculated and used according to that composition, and conversely, if the PCD prediction is prioritized, the ratio of CO and CO₂ is adjusted to match the PCD while maintaining the total amount of gas. Afterward, it finally checks the boundary conditions to ensure that CO and CO₂ are not negative and that the PCD is between 0 and 1, applies only the minimum necessary corrections, records the details of the application, and then transmits this physical consistency guarantee result to the providing step and control server of FIG. 5.

[0221] Finally, the prediction server (110) provides the mixed prediction, regime discrimination probability, mixed weight, and uncertainty indicator to the control server (120) via the prediction API. The control server (120) uses this information to provide the feedforward component and weight State dependency settings can be performed more precisely.

[0222] As described above, first-order blending refers to a procedure that temporally smoothly connects the submodel outputs of the current regime and the regime to be transitioned by utilizing regime identification probabilities, hysteresis, and residence time constraints. This prevents model switching chatter caused by instantaneous fluctuations in regime identification signals and reduces prediction discontinuities by continuously interpolating feature values ​​that change abruptly at regime boundaries. Second-order blending refers to a procedure that automatically reduces the contribution of submodels with high uncertainty during transition situations or abnormal data intervals by applying reliability-corrected weights based on the uncertainty index of each submodel. Since first-order blending focuses on transition stability and second-order blending focuses on model reliability, combining the two steps reduces the amplitude of prediction variability during distribution transitions and mitigates the phenomenon where prediction deviations are skewed toward specific submodels. As a result, multiple prediction intervals for CO and CO₂ are generated in a continuous and noise-robust form. Spikes in mixed oxygen injection commands during the control phase are reduced, decreasing losses associated with rate of change limitation and saturation processing, and improving the stability of tracking the afterburning index target. Additionally, even if performance degradation or data loss occurs in parts of the submodel, weights are redistributed to increase fault tolerance. Furthermore, there is the advantage of being able to realize adaptive behavior at the runtime level without modifying the existing model, while integrating dynamic characteristics that differ depending on the operating state.

[0224] FIG. 8 is a flowchart illustrating a hybrid oxygen injection control method based on actual and predicted values ​​for electric furnace afterburning according to one embodiment of the present invention.

[0225] An embodiment illustrated in FIG. 8 is performed by a control server (120), which combines a predicted value provided by a prediction server (110) with a measured value obtained through a controller (16) to generate a mixed oxygen injection command and calculates a final oxygen injection command by reflecting facility constraints.

[0226] Referring to FIG. 8, in step S810, the control server (120) receives process data from the electric furnace (10) via the controller (16), including actual values ​​of exhaust gas components and at least one of process operating variables. The control server (120) may also acquire operating state variables (such as whether the loop is closed, the charging stage, the stability of the exhaust gas flow rate, the power input status, etc.).

[0227] In step S820, the control server (120) checks the prediction values ​​for one or more future points in time calculated and provided by the prediction server (110). The prediction values ​​may include future CO, CO₂, PCD predictions, and prediction reliability indicators.

[0228] In step S830, the control server (120) calculates the feedback control component of the oxygen injection control based on the target value and the actual value for the control target.

[0229] First, the control server (120) defines an error signal based on the target value and actual value for the control goal as shown in [Equation 13] below. If necessary, flow rate correction is applied and [Equation 14] is used.

[0230] [Mathematical Formula 13]

[0231]

[0232] [Mathematical Formula 14]

[0233]

[0234] Here is the point in time exhaust gas flow rate, is the reference flow rate (median).

[0235] Subsequently, the control server (120) applies the gains of the proportional, integral, and derivative terms to the error signal to the feedback control component It calculates [it], and if expressed as a mathematical formula, it is as shown in [Mathematical Formula 15] below.

[0236] [Mathematical Formula 15]

[0237]

[0238] Here is the gain of each term.

[0239] In step S840, the control server (120) provides a feedforward control component that estimates the required amount of oxygen based on the predicted value. It calculates. For example, a trapezoidal integral can be applied, and first, at a future point in time The required oxygen flow rate in is defined as [Equation 16] below.

[0240] [Mathematical Formula 16]

[0241]

[0242] Here is the point in time The value of the required PCD increment converted into the number of CO oxidation reactions, is the predicted value of the exhaust gas flow rate at that time.

[0243] Next, When saying this For discrete prediction horizons such as the above, the following [Equation 17] is calculated.

[0244] [Mathematical Formula 17]

[0245]

[0246] Also, reaction equation Based on is calculated.

[0247] In step S850, the control server (120) combines the feedback component and the feedforward component to issue a mixed oxygen injection command. It calculates. This is calculated as shown in [Equation 18] below.

[0248] [Mathematical Formula 18]

[0249]

[0250] Here is a merge weight and can be adaptively set depending on model reliability or operating conditions.

[0251] In one embodiment, the control server (120) according to the operating state variable , upper limit , lower limit , limit on rate of change It can be set in a state-dependent manner.

[0252] Afterwards, the control server (120) saturates the mixed oxygen injection command by applying upper and lower limits of the equipment constraints, and this can be expressed as [Equation 19] below.

[0253] [Mathematical Formula 19]

[0254]

[0255] Here, It is an operation that cuts within the boundary.

[0256] Afterward, the control server (120) applies a change rate limit to produce a final oxygen injection command. This can be expressed as [Equation 20] below.

[0257] [Mathematical Formula 20]

[0258]

[0259] Here, The amount of change per unit time It is an operation that limits so as not to exceed.

[0260] Afterwards, the control server (120) calculates The controller (16) provides the control to the controller (16), and the controller (16) applies commands to the oxygen supply unit (burner and lance, etc.) to the electric furnace (10). The control result is displayed on the HMI, and the relevant log is stored.

[0261] The control server (120) is in operation It records the history of state-dependent settings and the trends of prediction reliability indicators, and can update adjustment rules as needed.

[0263] FIG. 9 is a block diagram illustrating the detailed configuration of a control server according to one embodiment of the present invention.

[0264] A control server (120) according to an embodiment illustrated in FIG. 9 calculates a feedback control component and a feedforward control component by combining a predicted value provided by a prediction server (110) and a measured value obtained through a controller (16), generates a mixed oxygen injection command, and then calculates and provides a final oxygen injection command by applying equipment constraints and a rate of change limit.

[0265] Referring to FIG. 9, the control server (120) includes a synchronization module (211), a prediction value management module (212), an error calculation module (213), a feedback control module (214), a feedforward control module (215), a reliability analysis module (216), a mixed command management module (217), and an interface module (218). Depending on the embodiment, it may further include a data storage and a model storage (not shown), an HMI and logging module (not shown), and a safety interlock and exception handling module (not shown).

[0266] The synchronization module (211) receives process data including actual values ​​of exhaust gas components and at least one of process operating variables from the controller (16), and performs timestamp matching and unit normalization.

[0267] The prediction value management module (212) receives CO prediction values, CO₂ prediction values, afterburning index prediction values, and prediction reliability indicators for multiple prediction intervals from the prediction server (110), performs time synchronization with the control cycle, and in the event of reception failure, maintains the last normal packet or applies a conservative replacement rule.

[0268] The error calculation module (213) calculates an error signal based on the target value and actual value for the control target, and if necessary, also calculates a flow rate correction error based on inventor data.

[0269] The feedback control module (214) applies at least one gain among a proportional term, an integral term, and a derivative term to the error signal, and calculates a feedback control component by summing the terms to which the gain is applied.

[0270] The feedforward control module (215) applies stoichiometry based on the predicted exhaust gas components, the predicted afterburn index value, and the predicted exhaust gas flow rate value, and calculates the feedforward control component according to the trapezoidal integration procedure described in the inventor's data.

[0271] The reliability analysis module (216) performs an analysis of the prediction reliability indicator, recent residuals, and operating state variables, and provides clues necessary for calculating the mixed weights and setting constraint parameters to the mixed command management module (217).

[0272] The mixing command management module (217) generates a mixed oxygen injection command by combining a feedforward component and a feedback component. Additionally, the mixing command management module (217) incorporates and performs constraint application and change rate limit functions. Specifically, it performs saturation processing by applying an upper limit O₂_max and a lower limit O₂_min according to equipment constraints, and calculates the final oxygen injection command O₂_Final by applying a change rate limit r_max to the difference from the previous command. The above α, O₂_max, O₂_min, and r_max can be set in a state-dependent manner according to the analysis results provided by the reliability analysis module (216) and the operating state variables.

[0273] The interface module (218) provides the final oxygen injection command to the controller (16) and maintains a closed-loop operation by receiving the command application result and field feedback. Additionally, the interface module (218) can transmit the predicted value, mixing weight α, constraint application history, change rate limit application history, error, and performance indicator to the HMI and logging module (not shown) to perform display and storage.

[0274] The data store and model store (not shown) preserve input data, predicted values, blends and final commands, weight and constraint parameter settings, audit logs and events, and manage versions and change history.

[0275] The safety interlock and exception handling module (not shown) performs safety stop or fallback procedures in the event of oxygen over-injection, communication anomalies, or data quality anomalies.

[0276] The control server (120) configured in this manner provides prediction-based control through input synchronization, prediction value management, error calculation, feedback component calculation, feedforward component calculation, reliability analysis, mixed command generation and integrated constraint processing, final command provision and recording.

[0278] FIG. 10 is a flowchart illustrating a feedforward oxygen amount calculation procedure according to one embodiment of the present invention.

[0279] One embodiment of FIG. 10 is performed by a feedforward control module (215) within a control server (120), and by applying stoichiometric and discrete integral rules based on future predicted values ​​provided by a prediction server (110), thereby providing a feedforward control component Produces.

[0280] In step S1010, the control server (120) receives input of a synchronized multiple prediction interval from the prediction server (110). The input includes CO prediction values, CO₂ prediction values, PCD prediction values, and exhaust gas flow rate prediction values ​​for each future time point, and confirms that these are values ​​to which the physical consistency post-processing of FIG. 5 has been applied.

[0281] In step S1020, the control server (120) calculates the required amount of CO oxidation at each future point in time by comparing the target value and the predicted value for the control goal. The calculation is based on the stoichiometry of the reaction equation and the definition of PCD, and calculates the amount of CO oxidation required to achieve or maintain the target PCD at each future point in time.

[0282] In step S1030, the control server (120) calculates a feedforward control component based on the amount of CO oxidation required at each future point in time.

[0283] Specifically, the control server (120) converts the CO oxidation requirement at each future point in time into the required oxygen amount. The conversion is performed based on the oxidation reaction equation of carbon monoxide, and obtains the instantaneous value of the oxygen flow rate required at each point in time.

[0284] Subsequently, the control server (120) combines the instantaneous value of the oxygen flow rate required at each time point with the predicted value of the exhaust gas flow rate to convert it into the oxygen consumption amount for each time interval. At this time, the amount of oxygen for each interval is prepared using the same time division as the prediction horizon array.

[0285] Subsequently, the control server (120) calculates the accumulated oxygen amount by applying a trapezoidal integration procedure over the entire prediction horizon, and this as a feedforward control component It is defined as follows. Integration is performed by averaging the oxygen flow rates of adjacent future times and multiplying them by the interval length, and is calculated cumulatively as the sum of the entire interval.

[0286] In one embodiment, the control server (120) may apply time alignment and interpolation rules with the control cycle to the calculated feedforward control component. Sample hold or linear interpolation is performed on the integral result to match the current control cycle, and if necessary, the decimal resolution is matched to the controller command resolution.

[0287] In step S1040, the control server (120) feedforward control component according to the operating state variable The confidence interval and usage range are set. For example, if the prediction reliability is low, the reliability summary value is provided to the reliability analysis module (216) so that weight adjustment in the mixing stage is possible.

[0288] Afterwards, the control server (120) calculates It is transmitted to the mixing command management module (217). The mixing command management module (217) performs weighted combination with the feedback component and determines the final oxygen injection command by applying equipment constraints and change rate limits.

[0289] Finally, the control server (120) logs the used prediction packet identifier, oxygen amount per time interval, integration settings, and alignment rules, and stores them in a data repository and a model repository (not shown) if necessary.

[0290] According to this procedure, the control server (120) provides a preemptive oxygen amount reflecting the future process state. It stably calculates and, combined with the mixing step of Fig. 8, implements prediction-based hybrid oxygen injection control.

[0292] FIG. 11 is a flowchart illustrating a state-dependent parameter scheduling procedure according to one embodiment of the present invention.

[0293] An embodiment illustrated in FIG. 11 is performed in a reliability analysis module (216) and a mixed command management module (217) within a control server (120), and the reliability indicator and operating state variable provided by the prediction server (110) are analyzed to stably determine the weight α and constraint parameter, and then provided for the generation of mixed commands.

[0294] In step S1110, the reliability analysis module (216) receives, in synchronization, the quantiles or prediction intervals for multiple prediction intervals, the prediction variance, the residual summary in the recent window, and the metadata for each horizon from the prediction value management module (212).

[0295] In step S1120, the reliability analysis module (216) calculates a reliability score based on the predicted reliability for each horizon.

[0296] Specifically, the reliability analysis module (216) calculates or updates the prediction reliability indicator for each horizon. The indicator consists of at least one of the interval between quantiles, the prediction variance, and the recent residual-based indicator, and displays the quality degradation interval by considering the channel quality flag and the data valid sample ratio together.

[0297] Subsequently, the reliability analysis module (216) normalizes and summarizes the reliability indicators for each horizon to calculate the reliability score to be applied to the current control cycle. Both the single score and the horizon score are prepared through a horizon-weighted average or conservative selection rule, and a conservative rating is assigned if there is a quality degradation flag.

[0298] In step S1130, the reliability analysis module (216) determines weight candidate α according to a predefined reliability and weight mapping table or interval rule. As reliability increases, α is assigned larger, and as reliability decreases, α is assigned smaller, and saturated so as not to exceed the set minimum and maximum value ranges.

[0299] In one embodiment, the reliability analysis module (216) applies time smoothing and hysteresis to suppress rapid fluctuations in α. If the difference from the previous cycle is below a threshold value, the existing value is maintained, and if a communication anomaly or data quality degradation is detected, α is reduced or temporarily frozen according to safety rules.

[0300] In step S1140, the reliability analysis module (216) identifies the current operating state and sets the state-dependent parameters.

[0301] Specifically, the reliability analysis module (216) obtains operating state variables such as whether the loop is closed, the charging stage, the exhaust gas flow rate stability, and the power input status from the synchronization module (211), and identifies the current operating state according to the state classification rule.

[0302] Afterward, the reliability analysis module (216) sets state-dependent parameters by referring to a state-specific schedule table. The parameters include an allowable range of weight α, upper and lower limits of oxygen input, and a limit on the rate of change, which are conservatively reduced in unstable states and relaxed in stable states.

[0303] In one embodiment, the reliability analysis module (216) applies a rule that reflects the characteristics of each horizon and, if necessary, places weight on the short horizon and reduces weight on the long horizon. At this time, an additional repair factor is applied to the state transition interval to mitigate the overreaction.

[0304] In step S1150, the reliability analysis module (216) transmits the final determined α and state-dependent parameter set to the mixing command management module (217). The mixing command management module (217) combines the feedforward component and the feedback component with the received α, and in the same step, applies upper and lower saturation and rate of change limits to produce the final oxygen injection command.

[0305] Afterwards, the reliability analysis module (216) and the mixed command management module (217) log the applied α, status, constraint parameter, quality flag, and result command, and if the indicator falls below the threshold for a long period, transmit a tuning request or relearning trigger signal to the relearning management module (not shown).

[0306] According to this procedure, the control server (120) can stably calculate weight α and constraint parameters by reflecting the prediction reliability and operating status, and can implement mixed command generation and batch processing.

[0308] FIG. 12 is a flowchart illustrating a mixing command generation and constraint processing procedure in one embodiment of the present invention.

[0309] An embodiment according to FIG. 12 is performed by a mixing command management module (217) within a control server (120) and describes a process of combining a feedforward control component and a feedback control component and applying equipment constraints and a rate of change limit to produce a final oxygen injection command.

[0310] Referring to FIG. 12, in step S1210, the mixing command management module (217) receives from the feedforward control module (215) a feedforward control component, i.e., preemptive oxygen amount, and from the feedback control module (214) a feedback control component, i.e., corrected oxygen amount, and from the reliability analysis module (216) a weight α and a lower limit O₂, which is a state-dependent parameter. min , upper limit O₂ max , limit on rate of change r max Receives by synchronizing.

[0311] In step S1220, the mixing command management module (217) generates a mixed oxygen injection command by weighting the two control components according to the aforementioned combining rule. At this time, the weight α can be updated periodically to reflect the results of the reliability analysis.

[0312] In step S1230, the mixing command management module (217) can perform saturation processing on the mixing command and apply a change rate limit to the saturated command to produce a final oxygen injection command.

[0313] Specifically, the mixed command management module (217) checks whether constraint parameters corresponding to the current operating state are being applied, and if necessary, updates the lower limit, upper limit, and rate of change limit with the latest values ​​transmitted from the reliability analysis module (216).

[0314] Afterwards, the mixing command management module (217) applies a lower limit and an upper limit to the mixing command to reflect equipment constraints and performs saturation processing to limit values ​​exceeding the boundary to the boundary.

[0315] The mixing command management module (217) calculates the final oxygen injection command by applying a change rate limit so that the amount of change per unit time does not exceed the set limit for the saturated command.

[0316] In one embodiment, the mixed command management module (217) checks whether saturation has occurred and whether a rate of change limit is applied, and if repeated saturation or excessive limit is detected, it notifies the reliability analysis module (216) of a request to reduce weight α or relax the rate of change limit in the next control cycle.

[0317] In step S1240, the interface module (218) transmits the final oxygen injection command to the controller (16) and receives an application acknowledgment. If a communication anomaly or application failure is detected, a safety interlock and exception handling module (not shown) is called to perform safety value maintenance or fallback.

[0318] In one embodiment, the mixing command management module (217) and the interface module (218) log the inputs used, mixing commands, saturation processing results, final commands, applied α and constraint parameters, saturation and rate of change limit events, and store them in a data repository and a model repository (not shown) as needed.

[0319] In one embodiment, the mixed command management module (217) monitors the actual response after the command is applied and, if an over- or under-response persists for a certain period, transmits a parameter adjustment request to the reliability analysis module (216).

[0320] According to this procedure, the control server (120) consistently performs the combination of feedforward and feedback, the application of upper and lower saturation limits, and the rate of change limit to stably determine the final oxygen injection command.

[0322] FIG. 13 is a flowchart illustrating the burner and lance unit distribution optimization and sensitivity feedback procedure according to one embodiment of the present invention.

[0323] One embodiment of FIG. 13 is a procedure for safely and efficiently distributing the total mixed oxygen injection command calculated in hybrid oxygen injection control to burners and lances, and is configured to ensure a balance of target PCD tracking, CO slip reduction, thermal uniformity maintenance, and oxygen usage efficiency even before distribution or when equipment conditions change.

[0324] Referring to FIG. 13, in step S1310, the control server (120) receives the total mixed oxygen injection command and state-dependent parameters of the current control cycle from the mixing command management module (217), and collects unit-specific constraint information such as operating state variables, a summary of prediction reliability, availability and maintenance status of each unit, upper and lower flow rate limits, and rate of change limits from the synchronization module (211) and the prediction value management module (212). Thus, the total amount, constraints, reliability, and facility status required for distribution optimization are organized into one frame.

[0325] In step S1320, the control server (120) estimates spatial hints of the heat demand distribution and CO oxidation demand based on operating state variables such as the charging stage, power input status, exhaust gas flow rate stability, and flame image quality (not shown). By referring to the output of multiple prediction intervals provided by the prediction server (110) and recent residuals, it expresses in the form of weights which zone has a large potential for oxygen reaction and which unit has high reaction stability. This output is used as an initial value for the target contribution per unit.

[0326] In step S1330, the control server (120) breaks down the entire mixed oxygen injection command into unit-specific initial distributions. The initial distributions are calculated in proportion to the unit rated capacity, current availability, the result of the previous distribution, and the space hint from step S1320, and are corrected by immediately applying unit-specific upper and lower limits to ensure that physically impossible initial proposals are not generated. These initial proposals are used as the starting point for the next optimization.

[0327] Subsequently, the control server (120) executes a distribution evaluation model for the unit-specific distribution plan. The distribution evaluation model is a lightweight response model that quickly approximates the effect of unit-specific oxygen changes on PCD, CO slip, temperature deviation, and combustion stability, and adjusts the model confidence weights by reflecting prediction reliability and recent operation log quality. The results of this step are summarized as expected performance indicators and constraint violation risk for each unit-specific distribution plan.

[0328] In step S1340, the control server (120) performs distribution optimization by configuring the distribution evaluation results into an objective function and constraints. The objective function includes reducing the target PCD tracking error, minimizing CO slip, improving temperature uniformity, and enhancing oxygen usage efficiency, and applies a remunerative weighting in low-reliability sections. The constraints include upper and lower limits on flow rates per unit, limits on the rate of change, minimum ignition maintenance conditions, and mutual interference avoidance rules. Optimization is calculated in a manner that guarantees a solution within a limited time, and if no valid solution exists, re-optimization is performed automatically by applying conservative relaxation rules. As a result, the target distribution value per unit is calculated.

[0329] Subsequently, the control server (120) converts the target distribution value per unit into an executable command. By comparing it with the actual command per unit from the previous cycle, a change rate limit is applied, and units with poor ignition status or flow sensor quality are conservatively adjusted or temporarily excluded. The converted command per unit is transmitted to the controller (16) through the interface module (218), and once the application result is confirmed, the process proceeds to the next step.

[0330] In step S1350, the control server (120) collects short-term responses after the application of unit-specific commands. The collected items include the latest actual values ​​of exhaust gas components, flame stability indices, unit flow rate tracking errors, and indices of variation in local temperature or flow velocity. This data is used to determine the validity of the distribution results and as input for the next sensitivity estimation.

[0331] Additionally, the control server (120) performs sensitivity estimation for each unit. Sensitivity refers to the local influence of oxygen distribution in each unit on PCD error, CO slip, and temperature deviation. It is calculated by sequentially fine-modulating unit commands with very small amplitudes within a safe range, or, if modulation is not possible, by regression estimation using the natural occurrence interval of recent changes. Modulation is automatically deactivated during the initial loading stage or in an unstable state, and safety is ensured in the section where modulation is applied by prioritizing the change rate limit and upper and lower limits.

[0332] Subsequently, the control server (120) stores the estimated sensitivity in the sensitivity register and accumulates it in a noise-insensitive manner by applying time weights. Units with uncertain sensitivity or frequently reversed signs are marked as having low reliability, and the objective function weights and constraint margins of the next distribution optimization are automatically consolidated.

[0333] In step S1360, the control server (120) updates the distribution weighting schedule by reflecting the latest sensitivity and reliability information. Units that contribute significantly to reducing PCD error have their weights slightly increased, while units that worsen CO slip or increase temperature deviation have their weights lowered. This schedule is directly used for the initial distribution in step S1330 and the optimization in step S1350 of the next control cycle, so that the learning from the previous step leads to performance improvement in the next step.

[0334] Subsequently, if a unit failure, valve stall, or sensor abnormality is detected, the control server (120) automatically isolates the unit and immediately re-performs the distribution optimization described above for the remaining units. The reason for isolation and the alternative distribution results are recorded in the HMI and logging module (not shown), and if repeated isolation occurs, a maintenance request signal is transmitted to an external system (not shown).

[0335] Based on this procedure, the present embodiment stably and adaptively distributes the total mixed oxygen injection command to burner and lance units, and continuously learns unit-specific sensitivity through small-scale, safe experimental modulation and response analysis to reflect it in the distribution optimization of the next cycle. As a result, even in situations where distribution transitions and equipment deviations exist, PCD target tracking, CO slip suppression, thermal uniformity maintenance, and oxygen efficiency improvement can be achieved simultaneously; furthermore, service continuity and safety can be ensured through automatic isolation and re-optimization even in the event of performance degradation or failure of some units.

[0337] The methods according to the various embodiments of the present application described above may be implemented in the form of program instructions that can be executed through various computer means and recorded on a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, etc., either alone or in combination. The program instructions recorded on the medium may be those specifically designed and configured for the embodiments, or they may be those known and available to those skilled in the art of computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes; optical recording media such as CD-ROMs and DVDs; magneto-optical media such as floptical disks; and hardware devices specifically configured to store and execute program instructions, such as ROM, RAM, and flash memory. Examples of program instructions include machine code, such as that generated by a compiler, as well as high-level language code that can be executed by a computer using an interpreter, etc. The hardware devices described above may be configured to operate as one or more software modules to perform the operation of the embodiments, and vice versa.

[0338] Although the embodiments have been described above with reference to limited examples and drawings, those skilled in the art can make various modifications and variations from the description above. For example, suitable results can be achieved even if the described techniques are performed in a different order than described, and / or the components of the described system, structure, device, circuit, etc. are combined or assembled in a form different from described, or replaced or substituted by other components or equivalents.

[0339] Therefore, other implementations, other embodiments, and equivalents to the claims also fall within the scope of the claims set forth below.

Claims

Claim 1 A hybrid oxygen injection control method based on actual values ​​and predicted values ​​for electric furnace afterburning, comprising: receiving process data including at least one of actual values ​​of exhaust gas components and process operating variables from an electric furnace process; confirming a predicted value for one or more future time points calculated based on the process data; calculating a feedback control component for oxygen injection control based on a target value for a control target and the actual values; calculating the amount of oxygen required to achieve the target value according to stoichiometry for each of one or more future time points based on the predicted values ​​of future exhaust gas components, afterburning index, and exhaust gas flow rate calculated from the predicted values, and calculating a feedforward control component using the calculation results for one or more future time points; and determining an oxygen injection command based on the feedback control component and the feedforward control component. Claim 2 A hybrid oxygen injection control method based on actual and predicted values ​​for electric furnace afterburning, comprising: a step of calculating the feedback control component in claim 1, a step of defining an error signal based on the difference between the target value and the actual value; and a step of applying a gain corresponding to at least one of a proportional term, an integral term, and a derivative term to the error signal, and calculating the feedback control component by summing the terms to which the gain is applied. Claim 3 delete Claim 4 In claim 1, the step of determining the oxygen injection command comprises the step of calculating a mixed oxygen injection command by weighted combining the feedback control component and the feedforward control component; wherein the weighted combination is defined according to the following mathematical formula according to the weight α, and Here, O2 Mix is a mixed oxygen injection command, O2 FF is a feedforward control component, O2 PID A hybrid oxygen injection control method based on actual and predicted values ​​for electric furnace afterburning, characterized in that is a feedback control component and α is a weight of 0 or more and 1 or less. Claim 5 A hybrid oxygen injection control method based on actual and predicted values ​​for electric furnace afterburning, wherein, in paragraph 4, the weight α is calculated based on a prediction reliability index, and the prediction reliability index is calculated based on at least one of the prediction variance, the width between quantiles, or the residual in the recent window. Claim 6 A hybrid oxygen injection control method based on actual and predicted values ​​for electric furnace afterburning, wherein, in claim 4, the step of determining the oxygen injection command comprises the step of saturating the mixed oxygen injection command by applying upper and lower limits according to equipment constraints. Claim 7 A hybrid oxygen injection control method based on actual and predicted values ​​for electric furnace afterburning, wherein, in claim 6, the step of determining the oxygen injection command further comprises the step of calculating the final oxygen injection command by applying a rate of change limit to the difference between the oxygen injection command at a previous time and the mixed oxygen injection command. Claim 8 A hybrid oxygen injection control method based on actual and predicted values ​​for electric furnace afterburning, wherein, in claim 7, the weighting α, the upper and lower limits and the rate of change limit are set and applied in a state-dependent manner according to an operating state variable, and the operating state variable includes at least one of whether a loop is closed, a charging stage, exhaust gas flow rate stability, or a power input state. Claim 9 A computer-readable recording medium having a computer program stored therein that enables the methods of claims 1, 2 and 4 through 8 to be performed in combination with a computer, which is hardware. Claim 10 An electric furnace control system comprising at least one processor; and a memory for storing instructions, wherein when the instructions are executed individually or collectively by the at least one processor, the processor performs the following operations: receiving process data including at least one of actual values ​​of exhaust gas components and process operating variables from an electric furnace process; confirming a predicted value for one or more future points in time calculated based on the process data; calculating a feedback control component for oxygen injection control based on a target value for a control target and the actual values; calculating the amount of oxygen required to achieve the target value according to stoichiometry for each of one or more future points in time based on the predicted values ​​of future exhaust gas components, afterburning index, and exhaust gas flow rate calculated from the predicted values, and calculating a feedforward control component using the calculation results for one or more future points in time; and determining an oxygen injection command based on the feedback control component and the feedforward control component.

Images