A method and system for fine desulfurization of blast furnace gas based on predictive control

Abstract

This invention relates to the field of desulfurization control technology and discloses a method and system for fine desulfurization of blast furnace gas based on predictive control. The method includes: collecting process data, performing redundancy consistency checks and weighted fusion to obtain fused data; performing state estimation within a sliding time-domain window based on the fused data to obtain the process state; establishing a mechanism-data fusion prediction model based on the process state and outputting the predicted state; constructing a constrained model predictive control problem within the prediction time-domain window and solving it to obtain control commands; combining actuator dynamics and time delays to perform feedforward compensation on the control commands and executing the control commands; and setting a safety barrier, switching to conservative or degraded operation when the redundancy consistency check fails or the predicted state violates the upper limit of total sulfur at the outlet or the process boundary. This application improves the precision and economic efficiency of desulfurization control and enhances safety under complex operating conditions.

Description

Technical Field

[0001] This invention relates to the field of desulfurization control technology, and more specifically, to a method and system for fine desulfurization of blast furnace gas based on predictive control. Background Technology

[0002] In the blast furnace gas purification process, the desulfurization stage directly affects the quality of downstream gas consumption and the stability of steel smelting. With the expansion of process scale and the tightening of emission standards, relying solely on empirical formulas or traditional time series models for sulfur concentration prediction suffers from insufficient accuracy and poor generalization ability. This can easily lead to insufficient desulfurizing agent dosage causing excessive emissions, or excessive dosage causing increased energy consumption and operating costs.

[0003] For example, invention patent CN120373092A – a desulfurization control method based on blast furnace gas sulfur concentration prediction – improves prediction accuracy and control adaptability to some extent by modeling the convection-diffusion equation and reaction rate equation of sulfur concentration, combining it with a long short-term memory network to output the predicted sulfur concentration, and then using an optimization model and reinforcement learning mechanism to adjust the dosing strategy. However, this scheme still has some problems: it does not fully consider the synergistic reaction and mass transfer effects of polysulfide species (such as hydrogen sulfide and carbon monoxide sulfides), and the prediction model does not adequately characterize the decay of absorbent activity; the optimization control is mainly based on the dosing target, and the energy consumption and dynamic constraints of the regeneration process are not adequately considered; it lacks accurate compensation for the dynamic characteristics and time delay of the actuator, which may lead to a disconnect between prediction and execution; in the event of sudden changes in operating conditions or sensor anomalies, the anomaly detection and safety switching mechanism still mainly relies on alarms and manual intervention, and the automatic degradation operation and conservative operation mechanisms are not yet perfect.

[0004] Therefore, it is necessary to design a blast furnace gas fine desulfurization method and system based on predictive control to solve the problems existing in the current technology. Summary of the Invention

[0005] In view of this, the present invention proposes a method and system for fine desulfurization of blast furnace gas based on predictive control, which aims to solve the problem that the current sulfur concentration prediction accuracy is insufficient, making it difficult to achieve stable desulfurization process under complex and fluctuating operating conditions.

[0006] In one aspect, the present invention proposes a method for fine desulfurization of blast furnace gas based on predictive control, comprising: Collect process data, including sulfur species sensor data and temperature, pressure, flow rate, pH, pressure drop, liquid level, regeneration parameters and actuator opening data. Perform redundancy consistency checks and weighted fusion on the process data to obtain fused data. Based on the fused data, state estimation is performed within a sliding time window to obtain the process state, which includes polysulfide species concentration, absorbent activity, and absorbent capacity. Based on the process status, a mechanism-data fusion prediction model is established to jointly characterize the mass transfer-reaction process of polysulfide species and the attenuation process of absorbent, and the predicted status is output. Within the prediction time window, a constrained model is constructed to predict the control problem. With the goal of achieving the target total sulfur at the outlet and minimizing the cost of addition and regeneration, hard constraints are applied to the process boundary and actuator boundary. The control commands are obtained by solving the problem. The control commands include addition amount command, regeneration intensity command and switching command. The control commands are fed forward and compensated by combining actuator dynamics and time delay, and the control commands of the current sampling period are executed in the rolling mode of the prediction time domain window; a safety barrier is set, and when the redundancy consistency check fails or the prediction state will violate the upper limit of total sulfur at the outlet or the process boundary, conservative operation or degraded operation is switched.

[0007] Furthermore, when collecting process data and performing redundancy consistency checks and weighted fusion on the process data to obtain fused data, the process data includes: The sulfur species sensor includes at least two types of sensors with different measurement principles to acquire sulfur species sensor data, and simultaneously collects temperature data, pressure data, flow data, pH data, pressure drop data, liquid level data, regeneration parameter data and actuator opening data; The process data is uniformly time-scaled and resampled, and a preliminary screening based on physical boundaries and rate of change is performed to remove out-of-bounds and abrupt data. The consistency error is calculated for the outputs of different sensors for the same measurand and compared with the consistency threshold. At the same time, a residual test is established based on process constraints to determine that data that does not conform to material balance, energy balance or known process coupling relationship is abnormal data. The sensor health score is calculated based on the consistency error, residual size, noise intensity and historical stability, and the fusion weight of each data channel is updated with the health score and noise covariance. Weighted least squares fusion is used to perform weighted fusion on the data that passed the test to generate the fused data. The fused data provides a fused estimate of polysulfide species concentration, temperature, pressure, flow rate, pH, pressure drop, liquid level, regeneration parameters and actuator opening under a unified time scale. When the consistency error exceeds the first threshold or the sensor health score is lower than the second threshold, the corresponding data channel is downweighted.

[0008] Furthermore, when performing state estimation in the sliding time domain based on the fused data to obtain the process state, it includes: The sliding time-domain window is set to several sampling periods, and the state vector is constructed as polysulfide species concentration, absorbent activity and absorbent capacity. The discrete state equation and discrete measurement equation of the mechanism-data fusion prediction model are used as constraints, where the measurement equation takes the fused data as the observation input and the state equation takes the actuator opening data in the fused data as the control input. A moving horizon estimation objective function is established to minimize the weighted sum of observation error and state evolution error within the sliding time window. Non-negative and upper bound constraints are imposed on polysulfide species concentration, and range constraints of [0, 1] and smoothing constraints on the rate of state change are imposed on absorbent activity and absorbent capacity. The observation noise covariance and process noise covariance are adaptively updated based on sensor health scores and historical stability. When the consistency error exceeds the first threshold or the corresponding data channel is downweighted, the observation noise covariance of the corresponding observation in the measurement equation is increased, thereby solving for the process state within the sliding time-domain window.

[0009] Furthermore, a mechanism-data fusion prediction model is established based on the aforementioned process state to jointly characterize the mass transfer-reaction and absorbent decay processes of polysulfide species. When outputting the predicted state, it includes: Using the concentration of polysulfide species, absorbent activity, and absorbent capacity in the process state as state variables, a mechanism-data fusion prediction model is constructed. The mechanism-data fusion prediction model consists of a mechanism sub-model and a data correction component. The mechanism sub-model takes temperature, pressure, flow rate and pH as inputs to characterize the mass transfer-reaction coupling of hydrogen sulfide and carbon monoxide sulfides in the gas-liquid phase and the hydrolysis conversion of carbon monoxide sulfides to hydrogen sulfide, and describes the absorbent decay process with absorbent activity and absorbent capacity. The data correction component takes the fused data as input and dynamically corrects the systemic bias of the mechanism sub-model. The mechanism-data fusion prediction model is discretized and iterated at a sampling period consistent with the sliding time-domain window to output the prediction state. The prediction state includes the predicted concentration of polysulfide species, the predicted activity of absorbent, and the predicted capacity of absorbent, and the predicted value of total sulfur at the outlet is calculated accordingly.

[0010] Furthermore, the parameter setting and updating of the mechanism-data fusion prediction model includes: During the historical phase, constrained parameter identification is performed based on the fused data with the goal of minimizing the weighted sum of prediction errors. Non-negative and upper bound constraints are imposed on the concentration of polysulfide species, and range and monotonicity constraints of [0, 1] are imposed on the absorbent activity and absorbent capacity. During the online phase, recursive least squares or Kalman-type updates are used to update the mass transfer coefficient and reaction rate coefficient of the mechanism sub-model, and the temperature sensitivity coefficient and acid-base sensitivity coefficient are limited to a preset range of variation. When the sensor health score drops or the consistency error exceeds the threshold, the affected parameters are frozen and the weight of the data correction component is increased.

[0011] Furthermore, when constructing a constrained model predictive control problem within the prediction time window, the following are included: Using the predicted state as the initial value, the objective function is set to satisfy the upper limit of total sulfur output and minimize the weighted sum of injection cost and regeneration cost. The decision variables are injection amount trajectory, regeneration intensity trajectory and switching time. The hard constraints include: the total sulfur at the outlet does not exceed the upper limit of the total sulfur at the outlet, the pressure drop of the absorber tower is within the pressure drop range, the liquid level of the absorber tower is within the liquid level range, the temperature of the regeneration unit is within the temperature range, the pressure of the regeneration unit is within the pressure range, the actuator opening is within the opening range, the rate of change of the actuator opening does not exceed the preset rate of change upper limit, and the switching time meets the minimum dwell time. The fusion prediction model explicitly includes actuator dynamics and actuator delay for feedforward prediction, thereby obtaining a control solution that satisfies the constraints within the prediction time window.

[0012] Furthermore, when solving for and obtaining control commands, the process includes: In each sampling period, the mechanism-data fusion prediction model is linearized to the current operating condition, and the constrained model predictive control problem is transformed into a bounded quadratic programming problem for online solution. A hierarchical objective is adopted to prioritize the feasibility of the total sulfur export ceiling constraint over the optimization of injection and regeneration costs; Only the control solution's dosage command, regeneration intensity command, and switching command for the current sampling period are issued, and the process rolls into the next sampling period; When a problem becomes temporarily infeasible due to a sudden change in operating conditions, the actuator opening rate is automatically tightened and the switching time is delayed until the minimum dwell time is met, without relaxing the upper limit of total sulfur at the outlet and the process boundary.

[0013] Furthermore, when performing feedforward compensation on the control commands by combining actuator dynamics and time delay, and executing the control commands for the current sampling period in a predictive time-domain window rolling manner, it includes: Dynamic models of actuators are established for the dosing pump and the regeneration valve respectively. The dynamic models of actuators adopt the form of first-order inertia plus pure time delay and the parameters are updated online. The feedforward compensation for the dosing command, regeneration intensity command and switching command is performed according to the actuator dynamic model, including: performing characteristic back calculation on the actuator opening degree-flow nonlinearity to obtain a linearized reference signal, constructing a Smith predictor for the actuator time delay to compensate for the time delay effect, and setting a ramp limiter on the reference signal so that the actuator opening degree change rate does not exceed the preset change rate upper limit. Dead zone compensation is performed when the reference signal falls into the actuator dead zone, and saturation projection is performed to satisfy the actuator boundary when the reference signal exceeds the opening range. The compensated and limited signal remains unchanged in the current sampling period through a zero-order hold. Only the dosage command, regeneration intensity command and switching command for the current sampling period are issued, and the switching command satisfies the minimum dwell time constraint.

[0014] Furthermore, safety barriers are set up to switch to conservative or degraded operation when redundancy consistency checks fail or the predicted state will violate the upper limit of total sulfur at the outlet or the process boundary, including: Based on the consistency error, sensor health score, and predicted state, a consistency state flag, a predicted violation flag, and a feasibility flag are generated. When the consistency error exceeds the first threshold or the sensor health score is lower than the second threshold, or when the predicted state indicates that the total sulfur at the outlet exceeds the upper limit of the total sulfur at the outlet at any time within the prediction time window, or when the process state will cross the process boundary, or when the constrained model predictive control problem has no feasible solution, the operation mode switch is triggered. In conservative operation, the upper limit of total sulfur at the outlet is prioritized. The conservative baseline setting of the control command is adopted to ensure that the dosage command is not lower than the preset conservative dosage baseline and the regeneration intensity command is not lower than the preset conservative regeneration baseline. The switching command is delayed until the minimum residence time is met and a feasible solution is restored. At the same time, the actuator opening range and the actuator opening change rate constraint are maintained and the control command is issued according to the feedforward compensation and limiting strategy. When the consistency status flag continuously reaches the preset number of sampling cycles or an actual out-of-bounds alarm occurs, it enters degenerate operation. During degenerate operation, the solution of the constraint model predictive control problem is stopped and the online update of the mechanism-data fusion predictive model is frozen. The control command is set according to the rule-based conservative strategy, so that the dosage command takes the preset conservative coefficient of the upper limit of the actuator opening range, the regeneration intensity command takes the conservative upper limit of the temperature range and pressure range, and the switching command is set to priority regeneration and meets the minimum residence time. When the redundancy consistency test returns to normal and the predicted state meets the upper limit of total sulfur at the outlet and the process boundary within a continuous preset sampling period, the conservative operation or degradation operation is smoothly exited by the slope limit and the rolling execution of the constraint model predictive control is resumed.

[0015] Compared with existing technologies, the advantages of this invention are as follows: By verifying the redundancy and consistency of multi-source process data and weighting and fusing it, the reliability of the input data is ensured. Combined with state estimation using a sliding time-domain window, core process state information such as polysulfide species concentration, absorbent activity, and capacity is obtained. This allows for the simultaneous characterization of the mass transfer reaction process of polysulfide species and the attenuation characteristics of the absorbent within the mechanism-data fusion prediction model, improving the accuracy and adaptability of the prediction. Within the prediction time-domain window, a constrained model predictive control is employed, achieving stable compliance with total sulfur emissions at the outlet. This optimizes both the addition and regeneration costs, overcoming the shortcomings of existing methods in considering energy consumption and dynamic constraints. Through actuator dynamic and time-delay compensation mechanisms, the consistency between predicted commands and actual execution is ensured. Combined with a safety barrier design, it can automatically switch to conservative or degraded operating modes in case of sensor malfunctions or sudden changes in operating conditions, avoiding the risk of relying solely on manual intervention. This improves the precision and economic efficiency of desulfurization control and enhances safety under complex operating conditions.

[0016] On the other hand, this application also provides a blast furnace gas fine desulfurization system based on predictive control, for applying the above-mentioned blast furnace gas fine desulfurization method based on predictive control, including: The data acquisition module is configured to acquire process data, which includes sulfur species sensor data and temperature, pressure, flow rate, pH, pressure drop, liquid level, regeneration parameters and actuator opening data. The process data is subjected to redundancy consistency check and weighted fusion to obtain fused data. The estimation module is configured to perform state estimation within a sliding time-domain window based on the fused data to obtain the process state, which includes polysulfide species concentration, absorbent activity, and absorbent capacity. The prediction module is configured to establish a mechanism-data fusion prediction model based on the process state, jointly characterize the mass transfer-reaction and absorbent decay process of polysulfide species, and output the prediction state. The control module is configured to construct a constrained model predictive control problem within a prediction time window, with the goal of achieving the target total sulfur at the outlet and minimizing the cost of addition and regeneration. Hard constraints are applied to the process boundary and actuator boundary, and control commands are obtained by solving the problem. The control commands include addition amount command, regeneration intensity command and switching command. The execution module is configured to perform feedforward compensation on the control commands by combining actuator dynamics and time delay, and execute the control commands of the current sampling period in a predictive time domain window rolling manner; a safety barrier is set up to switch to conservative operation or degraded operation when the redundancy consistency check fails or the predicted state will violate the upper limit of total sulfur at the outlet or the process boundary.

[0017] It is understandable that the above-mentioned blast furnace gas desulfurization method and system based on predictive control have the same beneficial effects, and will not be elaborated further here. Attached Figure Description

[0018] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart of a blast furnace gas fine desulfurization method based on predictive control provided in an embodiment of the present invention; Figure 2 The structural block diagram of the blast furnace gas fine desulfurization system based on predictive control provided in the embodiments of the present invention. Detailed Implementation

[0019] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0020] In traditional blast furnace gas desulfurization control, the mass transfer-reaction coupling effect of polysulfide species is not fully modeled, leading to prediction biases in the synergistic reaction process of hydrogen sulfide and carbon monoxide sulfides. The dynamic characteristics of absorbent activity decay lack precise characterization, making it difficult to quantify the correlation between absorbent capacity and regeneration intensity. Energy consumption constraints in the regeneration process are not included in the optimization objective, and the dynamic characteristics and time delay effects of actuators are not compensated for in the control commands, resulting in discrepancies between actual dosage and theoretical solutions. Redundant sensor data fusion mechanisms do not consider physical constraints and process coupling relationships, and safety switching strategies under abnormal operating conditions rely on manual intervention. For example, in the operation of a blast furnace gas purification system, the gas flow rate suddenly increased from 120,000 cubic meters per hour under standard conditions to 135,000 cubic meters per hour within 30 seconds, with temperature fluctuations exceeding ±15℃. The sulfur species sensor detected the instantaneous increase in hydrogen sulfide concentration but failed to simultaneously capture the hydrolysis and transformation of carbon monoxide sulfides. The absorbent in the absorption tower experiences a decrease in effective capacity due to activity decay, but the existing solution fails to correct the mass transfer coefficient in real time, resulting in a deviation between the predicted total sulfur concentration at the outlet and the actual value. The regeneration unit switching command does not consider valve delay characteristics, leading to a missynchronization between the regeneration cycle and the dosage command, causing the absorber level to exceed the safety threshold. Redundant sensors generate consistency errors due to sudden changes in operating conditions, but this does not trigger a data weighting mechanism, and the fused data contains outliers, causing state estimation divergence.

[0021] For this, please refer to Figure 1 As shown, this application proposes a method for fine desulfurization of blast furnace gas based on predictive control, comprising: S100: Collects process data, including sulfur species sensor data, as well as temperature, pressure, flow rate, pH, pressure drop, liquid level, regeneration parameters, and actuator opening data. Redundancy consistency checks and weighted fusion are performed on the process data to obtain fused data.

[0022] S200: Based on the fused data, state estimation is performed within a sliding time-domain window to obtain the process state, which includes the concentration of polysulfide species, absorbent activity, and absorbent capacity.

[0023] S300: Establish a mechanism-data fusion prediction model based on process status, jointly characterize the mass transfer-reaction and absorbent decay process of polysulfide species, and output the predicted status.

[0024] S400: Construct a constrained model predictive control problem within the prediction time window. With the goal of achieving the target total sulfur at the outlet and minimizing the cost of addition and regeneration, apply hard constraints to the process boundary and actuator boundary, and solve to obtain control commands, including addition amount command, regeneration intensity command and switching command.

[0025] S500: Combines actuator dynamics and time delay to perform feedforward compensation on control commands, and executes control commands for the current sampling period in a predictive time-domain window rolling manner. A safety barrier is set; when redundancy consistency checks fail or the predicted state will violate the upper limit of total sulfur at the outlet or the process boundary, it switches to conservative or degraded operation.

[0026] Specifically, redundancy consistency verification refers to verifying the consistency of process data from different sensors. By calculating the consistency error between the outputs of different sensors for the same measurand and comparing it with a preset threshold, abnormal data is eliminated. Specifically, a preliminary screening method based on physical boundaries and rate of change can be adopted, combined with residual verification to identify data that does not conform to material balance, energy balance or process coupling relationship, thereby ensuring data reliability and solving the problem of interference from abnormal or sudden sensor data.

[0027] Weighted fusion refers to assigning different fusion weights to each data channel based on sensor health scores, noise covariance, and historical stability. Specifically, a weighted least squares algorithm can be used to fuse data that has passed the test, generating fused estimates of parameters such as polysulfide species concentration, temperature, and pressure, thereby improving data accuracy and solving the prediction bias problem caused by single sensor measurement errors. State estimation within a sliding time window refers to inferring process state based on a mechanism-data fusion model within a dynamic time domain. Specifically, a moving horizon estimation method is used, aiming to minimize the weighted sum of observation errors and state evolution errors. Combined with non-negativity constraints and state change rate smoothing constraints, it outputs real-time estimates of polysulfide species concentration, absorbent activity, and absorbent capacity, solving the problem that traditional static models cannot adapt to dynamic changes in operating conditions.

[0028] Among them, the mechanism-data fusion prediction model refers to the combination of mass transfer-reaction mechanism equation and data-driven deviation correction model. Specifically, by discretizing the state equation and measurement equation, it dynamically corrects the systematic deviation of the mechanism model and outputs the predicted values ​​of polysulfide species concentration and absorbent activity decay, thus solving the problem of insufficient prediction accuracy caused by parameter mismatch in single mechanism models.

[0029] Among them, the constrained model predictive control problem refers to applying hard constraints on the process boundary and actuator boundary within the prediction time domain with the objectives of achieving the target total sulfur at the outlet and minimizing the cost of adding regeneration. Specifically, the control command is obtained by solving the quadratic programming problem online, which solves the risk of actuator overshoot or out-of-bounds due to neglecting dynamic constraints in traditional control methods.

[0030] Feedforward compensation refers to modifying control commands by combining the actuator's dynamic model and time delay characteristics. Specifically, it uses a first-order inertial model and a Smith predictor to compensate for time delay. By back-calculating the actuator's nonlinear characteristics and using ramp limiting, it ensures that the commands are executable and solves the control deviation problem caused by actuator response lag.

[0031] Among them, the safety barrier refers to switching the operating mode based on the redundancy consistency test results and the predicted state violation risk. Specifically, it triggers conservative or degraded operating strategies through consistency status flags and predicted violation flags to solve the risk of system runaway caused by sensor failure or sudden changes in operating conditions.

[0032] This application combines redundancy consistency verification, weighted fusion data preprocessing, and sliding time-domain state estimation to construct a mechanism-data fusion prediction model. Based on constrained model predictive control, it achieves multi-objective dynamic optimization. At the same time, it ensures the reliable execution of control commands and system safety through feedforward compensation and safety barrier mechanisms, forming a closed-loop predictive control architecture. This effectively addresses the nonlinear complex control problems of mass transfer-reaction coupling of polysulfide species, absorbent decay dynamics, and actuator delay during blast furnace gas desulfurization.

[0033] The working process and principle of this application are as follows: Process data is collected, including sulfur species sensor data and data on temperature, pressure, flow rate, pH, pressure drop, liquid level, regeneration parameters, and actuator opening. Redundancy consistency checks and weighted fusion are performed on these process data to obtain fused data. Based on the fused data, state estimation is performed within a sliding time-domain window to obtain the process state, including polysulfide species concentration, absorbent activity, and absorbent capacity. A mechanism-data fusion prediction model is established based on the process state, jointly characterizing the mass transfer-reaction process of polysulfides and the absorbent decay process, and outputting the predicted state. Within the prediction time-domain window, a constrained model predictive control problem is constructed, aiming to achieve the target total sulfur at the outlet and minimize the cost of addition and regeneration. Hard constraints are applied to the process boundary and actuator boundary, and control commands, including addition amount commands, regeneration intensity commands, and switching commands, are obtained. Feedforward compensation is performed on the control commands based on actuator dynamics and time delay, and the control commands for the current sampling period are executed in a rolling manner according to the prediction time-domain window. A safety barrier is set; when the redundancy consistency check fails or the predicted state violates the upper limit of total sulfur at the outlet or the process boundary, conservative operation or degraded operation is switched.

[0034] The accuracy of state estimation was improved through multi-source data fusion and redundancy consistency checks. The mechanism-data fusion prediction model jointly characterized the mass transfer-reaction and absorbent decay processes of polysulfide species, enhancing prediction accuracy. The constrained model predictive control problem considered multi-objective optimization of achieving outlet total sulfur standards and minimizing regeneration costs, while imposing process and actuator boundary constraints to ensure the feasibility of the control scheme. Feedforward compensation and rolling optimization mechanisms addressed the control mismatch caused by actuator dynamic characteristics and time delays. The safety barrier design enabled automatic switching under abnormal operating conditions, improving operational safety and reliability.

[0035] As a preferred embodiment, the specific implementation of this application is as follows: When collecting process data, two sulfur species sensors based on different principles are used to collect hydrogen sulfide and carbon monoxide sulfides concentration data, with a sampling frequency of 1Hz. Simultaneously, temperature, pressure, flow rate, pH, pressure drop, liquid level, regeneration temperature, regeneration pressure, and the opening of the dosing pump and regeneration valve are collected. The process data undergoes unified time-scale alignment and 1-minute resampling. The physical boundaries for hydrogen sulfide concentration are set to 0-2000ppm, and for carbon monoxide sulfides concentration to 0-500ppm; data exceeding these boundaries are discarded. The consistency error between the two sulfur species sensors is calculated; data exceeding 50ppm is considered abnormal. It should be noted that the time parameters in this application can be set using a layered time-scale configuration: the original sensor acquisition frequency is used for initial anomaly screening and consistency error calculation, while the unified time-scale resampling period is used for multi-source fusion, state estimation, and predictive control solutions. The aforementioned 1Hz sampling and 1-minute resampling correspond to the original acquisition layer and the control calculation layer, respectively. In different devices or embodiments, a unified timescale such as 1 second, 5 seconds, 10 seconds, or 1 minute can also be used. Unless otherwise specified, the subsequent "sampling period," "sliding time domain window length," and "prediction time domain window length" are based on the unified timescale of the control calculation layer. The zero-order hold period can be the same as the sampling period of the control calculation layer, or set to its integer division or harmonic period, but should remain consistent within the same embodiment. The consistency threshold of 50ppm is an example of an absolute threshold in one embodiment.

[0036] In other embodiments, the consistency threshold can be set jointly by an absolute threshold and a relative threshold. The relative threshold can be 0.5%–3% of the corresponding sensor range, and the larger of the converted values ​​of the absolute threshold and the relative threshold is taken as the judgment threshold. The threshold tuning can be determined based on the repeatability error during the sensor calibration stage, the noise standard deviation of the steady-state operating data, and the allowable false alarm rate and false alarm rate targets. In this application, "first threshold" uniformly represents the data anomaly threshold used for consistency error judgment, and "second threshold" uniformly represents the threshold used for sensor health score judgment. The sensor health score can be expressed using a normalized caliber or a percentage caliber: when using a normalized caliber, the health score ranges from 0 to 1. When using a percentage caliber, the health score ranges from 0 to 100 points. The two can be converted using "percentage score = normalized score × 100", for example, 0.6 corresponds to 60 points. 0.7 in the embodiments can be used as an example threshold for triggering enhanced observation noise covariance, and is not necessarily equivalent to the second threshold. The consistency error threshold used when the safety barrier is triggered can be an amplified threshold of the first threshold. This amplified threshold is determined by multiplying the first threshold by a preset amplification factor to distinguish between the two trigger levels: "data channel weight reduction" and "operation mode switching." A residual check is established based on material balance; when the difference between the inlet total sulfur and the outlet total sulfur exceeds 10% of the absorption capacity, it is considered abnormal data. Sensor health scores are calculated based on consistency error, residual size, noise intensity, and historical stability, and the fusion weights are updated. Weighted least squares is used to fuse the data that pass the check. A sliding time window of 10 minutes is set, and a state vector containing polysulfide species concentration, absorbent activity, and absorbent capacity is constructed. Using the discrete state equation and measurement equation of the mechanism-data fusion prediction model as constraints, a moving horizon estimation objective function is established, imposing a constraint of 0-2000 ppm on polysulfide species concentration and a range constraint of 0-1 on absorbent activity and capacity. The mechanism-data fusion prediction model includes a mechanism sub-model and a data correction component. The mechanistic sub-model characterizes the mass transfer-reaction coupling between hydrogen sulfide and carbon monoxide sulfides in the gas-liquid phase, as well as the hydrolytic conversion of carbon monoxide sulfides to hydrogen sulfide, and describes the decay process of absorbent activity and capacity. Data correction components dynamically correct for systematic biases in the mechanistic sub-model.

[0037] For ease of implementation, the mechanistic sub-model can be expressed using a combination of total sulfur equivalent conservation constraints and species kinetics: the hydrolysis of carbon monoxide sulfides can be modeled as a 1:1 stoichiometric relationship of COS + H2O → H2S + CO2. The gas-liquid mass transfer fluxes of hydrogen sulfide and carbon monoxide sulfides are calculated as the product of the overall mass transfer coefficient and the mass transfer driving force, respectively. The hydrolysis reaction rate is determined by the rate constant, species concentration, and pH correction term, and the rate constant can be described by the Arrhenius relation as a function of temperature. The absorbent activity and absorbent capacity are characterized by discrete decay-regeneration equations and limited to the range [0,1]. Data correction components can be constructed based on the mechanistic prediction residuals to build linear or nonlinear correction terms to compensate for unmodeled disturbances and slow drift. A 30-minute prediction time window is set to construct a constrained model predictive control problem. The objective function is to ensure that the total sulfur at the outlet does not exceed 30 ppm and to minimize the weighted sum of the injection cost and the regeneration cost. The decision variables are the injection trajectory, regeneration intensity trajectory, and switching time within 30 minutes.

[0038] The lengths of the sliding time-domain window and the prediction time-domain window can be tuned based on the device's dominant time constant, total actuator delay, sensor response time, and regeneration switching cycle. A 10-minute sliding time-domain window and a 30-minute prediction time-domain window are examples of an implementation. Generally, the sliding time-domain window can cover 1-3 times the state estimation dominant response time, and the prediction time-domain window can cover at least the total actuator delay and the lead time for a single switching decision. Under a 1-minute unified timescale, the sliding time-domain window can be selected from 5 to 20 minutes, and the prediction time-domain window can be selected from 15 to 60 minutes. These can be comprehensively determined through historical playback simulation using indicators such as the prediction error of total sulfur at the outlet, the number of constraint violations, and the fluctuation of control quantities. The following hard constraint and safety barrier trigger values ​​are a set of example parameters applicable to high-load, second-level actuator adjustment scenarios; in other embodiments, equivalent conversions or readjustments can be performed based on the device scale, regeneration conditions, and the unified timescale definition. The following hard constraints are applied: total sulfur at the outlet does not exceed 30 ppm; pressure drop in the absorber tower is within the range of 0.5-5 kPa; liquid level is within the range of 50%-90%; regeneration temperature is within the range of 60-80℃; regeneration pressure is within the range of 0.1-0.3 MPa; actuator opening is within the range of 0-100%; opening change rate does not exceed 2% / s; and switching times meet a minimum dwell time of 30 minutes. A first-order inertial plus pure time-delay actuator dynamic model is established for the dosing pump and regeneration valve, and parameters are updated online. Feedforward compensation is applied to control commands based on the actuator dynamic model, including characteristic back-calculation, Smith predictor compensation, and ramp limiting. A safety barrier is set: when the consistency error exceeds 100 ppm, the sensor health score is below 0.6, or the predicted status indicates that the total sulfur at the outlet will exceed 30 ppm, a conservative operation mode is triggered. In conservative operation, the dosing command is not less than 120% of the current value, and the regeneration intensity command is not less than 110% of the current value. When the consistency status is abnormal for 5 minutes or an actual out-of-bounds alarm occurs, it enters degenerate operation. Control commands are set according to the rule-based strategy, with the dosage set to 80% of the actuator opening limit and the regeneration intensity set to the upper limit of the temperature and pressure range. When the redundancy consistency check returns to normal and the predicted state meets the constraints for 10 consecutive minutes, it smoothly exits conservative or degenerate operation.

[0039] Through the above scheme, this application achieves accurate modeling of the mass transfer-reaction process of polysulfide species, improving the prediction accuracy of the synergistic reaction of hydrogen sulfide and carbon monoxide sulfides. The fusion of multi-source heterogeneous data and the introduction of a redundancy consistency verification mechanism eliminates the interference of sensor anomalies on state estimation. The predictive model, which integrates mechanism and data, accurately characterizes the dynamic characteristics of absorbent activity decay, achieving coupled modeling of absorbent capacity and regeneration process parameters. The feedforward compensation strategy combined with a rolling optimization mechanism solves the control mismatch problem caused by actuator dynamic characteristics and time delay, ensuring synchronization between control commands and process response. The multi-layer safety barrier design enables automatic switching between conservative and degraded operating modes in the event of sudden changes in operating conditions or sensor failure, improving the safety and reliability of system operation.

[0040] This application further proposes a method for collecting process data, performing redundancy consistency checks and weighted fusion on the process data, and obtaining fused data. This includes: acquiring sulfur species sensor data using at least two types of sensors with different measurement principles, and simultaneously acquiring temperature, pressure, flow, pH, pressure drop, liquid level, regeneration parameter data, and actuator opening data. The process data undergoes unified time-scale alignment and time resampling processing, and initial screening based on physical boundaries and rates of change is performed to remove out-of-bounds and abrupt data. Consistency errors are calculated for the outputs of different sensors for the same measurand and compared with a consistency threshold. Simultaneously, a residual check is established based on process constraints to determine data inconsistent with material balance, energy balance, or known process coupling relationships as abnormal data. Sensor health scores are calculated based on consistency errors, residual magnitude, noise intensity, and historical stability, and the fusion weights of each data channel are updated using the health scores and noise covariance. Weighted least squares fusion is used to weight and fuse the data that has passed the test, generating fused data. The fused data provides fused estimates of polysulfide species concentration, temperature, pressure, flow rate, pH, pressure drop, liquid level, regeneration parameters, and actuator opening under a unified time scale. When the consistency error exceeds the first threshold or the sensor health score is lower than the second threshold, the corresponding data channel is downweighted.

[0041] The sulfur species sensor employs both electrochemical and optical measurement principles, using cross-validation to improve the reliability of sulfur concentration data. Time-based resampling uses linear interpolation to align sensor data from different sampling frequencies. Physical boundaries are set according to process design parameters, such as a temperature range of 50℃ to 150℃, and the rate of change threshold is determined based on historical data statistics. Residual verification compares actual measured values ​​with predicted values ​​based on material balance; for example, the difference in sulfur mass flow rate between the absorber inlet and outlet should be within a preset error range. Sensor health scores are calculated using a weighted average method, integrating consistency error, residuals, noise covariance, and historical stability indicators. Noise covariance is calculated statistically using a sliding window. During fusion weight updates, sensors with higher health scores are assigned higher weights, and channels with lower noise covariance receive increased weights. Weighted least squares fusion aims to minimize the sum of squared weighted residuals to generate a fusion estimate. Weight reduction processing lowers the weights of abnormal channels to 10% to 30% of their original values ​​to prevent abnormal data from affecting the overall fusion result.

[0042] Specifically, during the data acquisition phase, sulfur concentration data is simultaneously acquired using two types of sulfur species sensors, combined with simultaneous acquisition of multiple parameters such as temperature and pressure, ensuring data temporal consistency. Time-stamp alignment uses a unified clock signal to label data from each sensor, and time resampling unifies data from different frequencies to a fixed sampling period. In the initial screening phase, data exceeding reasonable ranges are removed based on process physical boundaries; for example, data with temperatures below 50℃ or above 150℃ are marked as abnormal. The rate of change threshold is set based on the standard deviation of historical data; for example, a temperature change rate exceeding 5℃ / second is considered abrupt change data. Consistency error calculation uses the root mean square error of measurements from different sensors; when the error exceeds a preset threshold, residual verification is triggered. Residual verification verifies data rationality using material balance equations; for example, the difference between the inlet sulfur mass flow rate and the outlet sulfur mass flow rate should equal the absorbent absorption amount; exceeding this error range is considered abnormal. When calculating the health score, noise intensity is evaluated using the data variance within a sliding window, and historical stability is determined by statistically analyzing the number of data anomalies over a past period. The fusion weight is dynamically adjusted based on the score; for example, for every 10% decrease in the health score, the corresponding weight decreases by 20%. In the weighted least squares fusion process, data from each sensor are linearly combined according to their weights to generate fused multi-parameter estimates. When the consistency error of a sensor exceeds a threshold or its health score falls below a critical value, its weight is dynamically reduced to ensure that the fused data is not affected by faulty sensors. For example, when the health score of an electrochemical sensor is below 60%, its weight is adjusted from 0.5 to 0.1, while the weight of an optical sensor is increased to 0.9.

[0043] Through the above steps, the fused data is generated based on time-scaled unification, anomaly removal, and weight optimization, providing reliable input for state estimation.

[0044] As a preferred embodiment, the specific implementation of this application is as follows: The sulfur species sensor includes two types of sensors with different measurement principles to acquire sulfur species sensor data. The first type is an electrochemical sensor with a measurement range of 0-2000 ppm and a response time of less than 10 seconds. The second type is a laser absorption spectroscopy sensor with a measurement range of 0-5000 ppm and a response time of less than 5 seconds. Temperature, pressure, flow rate, pH, pressure drop, liquid level, regeneration parameters, and actuator opening data are collected simultaneously. The process data undergoes unified timescale alignment and time resampling processing. The sampling period used for initial screening of the sensor layer and consistency error calculation is set to 1 second. The unified timescale used for multi-source fusion, state estimation, and predictive control can be further resampled to 5 seconds to 1 minute depending on the dynamic nature of the operating conditions. Initial screening based on physical boundaries and rates of change is performed to remove out-of-bounds and abrupt data. The physical boundaries of temperature, pressure, and flow rate data are set according to the corresponding measurement point type and process design range. For example, temperature data can be set to 0-100℃ or 0-150℃ according to the corresponding measuring point, pressure data can be set to 0-1MPa, and flow data can be set to the corresponding range according to the measuring point of the main pipeline or branch (for example, the gas flow rate of the main pipeline can be set to 0-200000Nm3 / h, and the flow rate of the branch can be set according to the equipment range).

[0045] Consistency errors are calculated for the outputs of different sensors for the same measurand and compared with a consistency threshold. The consistency threshold is set to 1% of the measurement range. A residual check is established based on process constraints to identify data that does not conform to material balance, energy balance, or known process coupling relationships as anomalous data. Sensor health scores are calculated based on consistency error, residual magnitude, noise intensity, and historical stability. The health score uses a 0-100 scale. The fusion weights of each data channel are updated using the health score and noise covariance. Weighted least squares fusion is used to weight and fuse the data that pass the check, generating fused data. The fused data provides fused estimates of polysulfide species concentration, temperature, pressure, flow rate, pH, pressure drop, liquid level, regeneration parameters, and actuator opening under a unified time scale. When the consistency error exceeds the first threshold or the sensor health score falls below the second threshold, the corresponding data channel is downweighted. The first threshold is set to 2% of the measurement range, and the second threshold is set to 60 points.

[0046] Through the above technical solution, this application achieves redundancy consistency verification and weighted fusion of multi-source heterogeneous sensor data. This improves the reliability and accuracy of process data, laying the foundation for subsequent state estimation and predictive control.

[0047] Furthermore, by introducing a sensor health scoring mechanism, the system's adaptability to sensor failures is enhanced.

[0048] Specifically, when a sensor malfunctions, the system can automatically reduce its weight, thereby ensuring the quality of the fused data. For example, when the hydrogen sulfide sensor drifts, its impact is identified and mitigated, preventing overall performance degradation caused by a single sensor failure. The application further proposes a method for state estimation within a sliding time domain based on fused data to obtain the process state. This includes: setting the sliding time domain window as several sampling periods, constructing a state vector consisting of polysulfide species concentration, absorbent activity, and absorbent capacity, and using the discrete state equation and discrete measurement equation of the mechanism-data fusion prediction model as constraints. The measurement equation uses fused data as observation input, and the state equation uses actuator opening data from the fused data as control input. A moving horizon estimation objective function is established to minimize the weighted sum of observation error and state evolution error within the sliding time domain window. Non-negativity and upper bound constraints are applied to the polysulfide species concentration, and [0,1] range constraints and state change rate smoothing constraints are applied to the absorbent activity and absorbent capacity. The observation noise covariance and process noise covariance are adaptively updated based on sensor health scores and historical stability. When the consistency error exceeds a first threshold or the corresponding data channel is downweighted, the observation noise covariance of the corresponding observation in the measurement equation is increased, thereby solving for the process state within the sliding time domain window.

[0049] The length of the sliding time-domain window is set to five to twenty minutes based on the dynamic characteristics of the process. When the unified time-scale sampling period is 1 minute, it corresponds to five to twenty sampling periods; when the unified time-scale sampling period is 5 to 10 seconds, it corresponds to thirty to two hundred and forty sampling periods. The state vector includes hydrogen sulfide concentration, carbon monoxide sulfide concentration, absorbent activity coefficient, and the proportion of remaining absorbent capacity. The discrete state equation is obtained by discretizing the absorption reaction kinetic equation, and the discrete measurement equation is a linear mapping between the observed polysulfide species concentrations in the fused data and the state vector. In the moving horizon estimation objective function, the observation error term is weighted by Mahalanobis distance, the state evolution error term is weighted by quadratic form, and the weight matrix is ​​dynamically adjusted according to the noise covariance. The non-negativity constraint of polysulfide species concentration is achieved through inequality constraints, and the range constraints of absorbent activity and capacity are applied through a projection algorithm. The smoothing constraint of the state change rate is achieved by a quadratic penalty term for the state difference between adjacent sampling periods. When updating the observation noise covariance, if the sensor health score decreases, the diagonal elements of the noise covariance matrix of the corresponding observation channel increase by two to three times. The process noise covariance is calculated online recursively from the covariance of the residuals estimated based on historical conditions.

[0050] Specifically, within a sliding time-domain window, the observed concentrations of polysulfide species, temperature, and pressure data from the fused data are first input into the measurement equation, while the actuator opening data is input into the state equation. The objective function is estimated by moving the horizon, simultaneously optimizing the state sequence of the current moment and the previous four sampling periods within the window. This ensures that the state estimation conforms to both the current observation data and the state evolution law. For data channels with reduced weight, their influence on the objective function is reduced by increasing their noise covariance. During the solution process, the interior-point method is used to handle state constraints, ensuring that the polysulfide species concentration does not become negative and that the absorbent activity coefficient remains between zero and one. When a sudden change in absorbent activity is detected, abnormal jumps are suppressed by smoothing the rate of change of state, maintaining the physical rationality of the state trajectory. By adaptively adjusting the noise covariance, the robustness of the state estimation is maintained when sensor health fluctuates. For example, when the pH sensor health score is below 0.7, its corresponding noise covariance coefficient is increased to three times the baseline value, thereby reducing the interference of abnormal data on the state estimation.

[0051] As a preferred embodiment, the specific implementation of this application is as follows: A sliding time-domain window of 60 sampling periods is set, and a state vector is constructed including H2S concentration, COS concentration, absorbent activity, and absorbent capacity. The discrete state equation and discrete measurement equation of the mechanism-data fusion prediction model are used as constraints, where the measurement equation uses fused data as observation input, and the state equation uses actuator opening data from the fused data as control input. A moving horizon estimation objective function is established to minimize the weighted sum of observation error and state evolution error within 60 sampling periods. A non-negative upper bound constraint of 0-1000 ppm is applied to H2S and COS concentrations, and a [0,1] range constraint and a smoothing constraint of state change rate not exceeding 0.01 / min are applied to absorbent activity and absorbent capacity. The observation noise covariance and process noise covariance are adaptively updated based on sensor health scores and historical stability. When the consistency error exceeds a first threshold (or its relative threshold expression) or the corresponding data channel is downweighted, the observation noise covariance of the corresponding observation in the measurement equation is increased by a preset factor, for example, to 2 to 3 times the baseline value. Under extreme and abnormal operating conditions, the factor can be further increased to a higher level, thereby solving for the process state within a sliding time-domain window of 60 sampling periods.

[0052] Through the above technical solutions, this application achieves accurate estimation of polysulfide species concentration, absorbent activity, and capacity. The robustness and smoothness of the state estimation are improved by employing a sliding time-domain window and moving horizon estimation. Adaptive adjustment of the noise covariance enhances the adaptability to sensor anomalies and operating condition fluctuations. Applied constraints ensure the physical rationality of the estimation results. This provides reliable process state input for predictive control, contributing to improved accuracy and stability of desulfurization control.

[0053] This application further proposes a mechanism-data fusion prediction model using the concentration of polysulfide species, absorbent activity, and absorbent capacity in the process state as state variables. The mechanism-data fusion prediction model consists of a mechanism sub-model and a data correction component. The mechanism sub-model takes temperature, pressure, flow rate, and pH as inputs to characterize the mass transfer-reaction coupling between hydrogen sulfide and carbon monoxide sulfides in the gas-liquid phase, as well as the hydrolysis conversion of carbon monoxide sulfides to hydrogen sulfide. It also describes the absorbent decay process using absorbent activity and absorbent capacity. The data correction component takes the fused data as input and dynamically corrects the systematic bias of the mechanism sub-model. After discretization, the mechanism-data fusion prediction model iterates at a sampling period consistent with the sliding time-domain window, outputting the predicted state, which includes the predicted concentration of polysulfide species, the predicted absorbent activity, and the predicted absorbent capacity. Based on these, the predicted value of the total sulfur at the outlet is calculated.

[0054] The mechanistic sub-model uses gas-liquid phase equilibrium equations and reaction kinetic equations to describe the mass transfer-reaction coupling process between hydrogen sulfide and carbon monoxide sulfides, specifically including gas film mass transfer coefficient, liquid film mass transfer coefficient, and hydrolysis reaction rate parameters. The data correction component employs a residual feedback mechanism based on fused data, compensating for unmodeled dynamics in the mechanistic model by updating the correction gain matrix online. The discretization process uses an implicit Euler method to maintain numerical stability, and the sampling period is set to a fixed time interval synchronized with the sliding time-domain window, such as 5 or 10 seconds. The predicted total sulfur at the outlet can be obtained by converting the predicted polysulfide species concentrations to sulfur equivalents or by calibrated weighted summation. The conversion or weighting coefficients are pre-set according to the detection calibrator, calibration data, or process calculation calibrator.

[0055] Specifically, the mechanistic sub-model first calculates the driving force for mass transfer between the gas and liquid phases based on temperature, pressure, flow rate, and pH, and then obtains the interphase mass transfer flux between hydrogen sulfide and carbon monoxide sulfides by combining the absorption tower structural parameters. Subsequently, the conversion rate of carbon monoxide sulfides to hydrogen sulfide through hydrolysis is calculated based on the reaction kinetic equation; this rate is affected by pH and temperature. The absorbent activity and capacity are correlated with the current dosage and regeneration intensity through a decay function, which uses an exponential form to characterize the nonlinear decrease in absorbent performance over time. The data correction component receives real-time observations of polysulfide species concentrations from the fused data. The residual between the computer mechanistic sub-model's predicted values ​​and the actual observed values ​​is updated using a recursive least squares algorithm to dynamically adjust the output bias of the mechanistic sub-model. During discretization, the implicit Euler method transforms the continuous differential equations into difference equations, ensuring numerical convergence over a large sampling period. When calculating the predicted total sulfur export value, the hydrogen sulfide concentration and carbon monoxide sulfide concentration can be weighted and summed using sulfur equivalent conversion factors or calibrated weighting factors, and then output to the control module. For example, in one embodiment, an exemplary combination of weighting factors obtained through historical calibration (such as 0.8 and 0.2) can be used for engineering approximation calculations, but it is not limited to this set of values.

[0056] As a preferred embodiment, the specific implementation of this application is as follows: A mechanism-data fusion prediction model is constructed based on the concentration of polysulfide species, absorbent activity, and absorbent capacity under process conditions. This model consists of a mechanism sub-model and a data correction component. The mechanism sub-model, using temperature, pressure, flow rate, and pH as inputs, characterizes the mass transfer-reaction coupling of hydrogen sulfide and carbon monoxide sulfides in the gas-liquid phase, as well as the hydrolysis conversion of carbon monoxide sulfides to hydrogen sulfide. Simultaneously, the mechanism sub-model describes the absorbent decay process through absorbent activity and absorbent capacity. The data correction component, using the fused data as input, dynamically corrects the systematic bias of the mechanism sub-model.

[0057] Specifically, the mechanistic sub-model employs the two-film theory to describe the gas-liquid mass transfer process, considering the parallel absorption of hydrogen sulfide and carbon monoxide sulfides. The gas-phase mass transfer coefficient is calculated using Onda correlations, while the liquid-phase mass transfer coefficient is estimated using Higbie permeation theory. Reaction kinetics utilizes a second-order reaction rate expression, where the reaction rate constant changes with temperature according to the Arrhenius equation. The absorbent activity decay is modeled using an exponential decay model, with the decay coefficient related to the cumulative treated gas volume. The absorbent capacity is dynamically updated using the material balance equation. Data correction employs a recursive least squares algorithm, based on fused data within a sliding time window, to online correct key parameters of the mechanistic sub-model. Corrected parameters include the gas-liquid mass transfer coefficient, the reaction rate constant pre-factor, and the absorbent activity decay coefficient. During parameter initialization, the initial values ​​of the gas-phase mass transfer coefficient, liquid-phase mass transfer coefficient, and equivalent overall mass transfer coefficient can be calculated using Onda correlations, Higbie permeation theory, or equivalent empirical correlations, based on the absorber tower structure dimensions, packing parameters, gas-liquid load, and temperature and pressure conditions. The initial value of the hydrolysis reaction rate constant and the initial value of the activation energy can be obtained by fitting empirical values ​​from published literature, bench test data, or historical steady-state data of the device. The initial values ​​of the absorbent activity decay coefficient and capacity recovery coefficient can be obtained by performing constrained identification using operating data from the past 30-90 days. To improve cross-device applicability, the above parameters can be normalized to their design nominal values, and the initial historical identification values ​​can be set within the range of 0.8-1.2 times the nominal values, and limited to 0.5-2.0 times the initial values ​​during online updates.

[0058] Furthermore, the mechanism-data fusion prediction model, after discretization, iterates at a sampling period consistent with the sliding time-domain window. This outputs predicted states, including predicted polysulfide species concentration, predicted absorbent activity, and predicted absorbent capacity. Based on these predicted states, the predicted value of total sulfur at the outlet is calculated. Through the above technical solution, this application achieves a joint characterization of the mass transfer-reaction process of polysulfide species and the absorbent decay process. The mechanism sub-model provides a deep understanding of the essence of the process, while the data correction component compensates for the shortcomings of the mechanism model and improves prediction accuracy. This fusion method retains the interpretability and extrapolation capability of the mechanism model while utilizing the adaptability of the data-driven method. Therefore, when facing complex and variable blast furnace gas desulfurization conditions, it can more accurately predict changes in polysulfide species concentration and absorbent performance decay, laying the foundation for subsequent precise control. In some of the above solutions in this application, the parameter setting and updating process of the mechanism-data fusion prediction model suffers from insufficient adaptability of model parameters under dynamic operating conditions. The identification of parameters in historical stages did not fully consider physical constraints, leading to deviations of the initial model parameters from reality. The lack of adaptive adjustments to sensor anomalies during online parameter updates may introduce errors. The failure to impose reasonable limits on temperature and pH sensitivity coefficients can easily lead to model instability.

[0059] This application further proposes a parameter setting and updating method for a mechanism-data fusion prediction model, including: performing constrained parameter identification based on fused data during the historical phase, aiming to minimize the weighted sum of prediction errors, and imposing non-negativity and upper bound constraints on polysulfide species concentration, and range and monotonicity constraints of [0,1] on absorbent activity and absorbent capacity. During the online phase, recursive least squares or Kalman-type updates are used to update the mass transfer coefficient and reaction rate coefficient of the mechanism sub-model, and the temperature sensitivity coefficient and pH sensitivity coefficient are limited to preset variation ranges. When the sensor health score decreases or the consistency error exceeds a threshold, the affected parameters are frozen and the weight of the data correction component is increased. The constrained parameter identification ensures that the model parameters conform to physical laws by introducing non-negativity and upper bound constraints on polysulfide species concentration, and range and monotonicity constraints on absorbent activity and capacity. The recursive least squares or Kalman-type updates use sliding window data to adjust the mass transfer coefficient and reaction rate coefficient online, improving the model's ability to track dynamic operating conditions. Temperature and pH sensitivity coefficients are limited to preset ranges to prevent abrupt changes in model parameters due to environmental fluctuations. Parameter updates are frozen when sensor health is abnormal to avoid abnormal data contaminating the model, while the weight of the data correction component is increased to compensate for biases in the mechanistic sub-model.

[0060] Specifically, during the historical phase, fused data is input into the parameter identification module. An optimization algorithm solves for the parameter set that satisfies the constraints of concentration non-negativity and monotonically decreasing absorbent activity and capacity, ensuring the initial model conforms to the process mechanism. During the online phase, new fused data is collected each sampling period. The mass transfer coefficient and reaction rate coefficient are updated using recursive least squares, with the temperature sensitivity coefficient limited to ±5% and the pH sensitivity coefficient limited to ±3%. When the sulfur species sensor health score falls below a threshold, the mass transfer coefficient update is frozen, and the weighting coefficient of the data correction component is increased from 0.3 to 0.7, using the data-driven component to compensate for deviations in the mechanism model. For example, when the temperature sensitivity coefficient is detected to exceed a preset range, it is forcibly constrained to the boundary value to avoid model prediction distortion. Through constrained parameter updates and anomaly handling mechanisms, the model maintains the rationality of the mechanism while enhancing its adaptability to operating conditions and its anti-interference capabilities. In one embodiment, the forgetting factor for the recursive least squares method can be set to 0.95-0.995, and the initial covariance matrix can be set to a diagonal matrix according to the parameter normalization scale. When using Kalman averaging, the process noise covariance and observation noise covariance can be adaptively adjusted based on the variance of the innovation sequence. The parameter freezing trigger condition can be a combination of the criteria of "sensor health score below the threshold" and "consistency error exceeding the limit for multiple consecutive sampling periods". Parameter unfreezing can be performed after the consistency error recovers to below the threshold and continues for a preset time, so as to reduce the contamination of model parameters by abnormal data.

[0061] As a preferred embodiment, the specific implementation of the scheme in this application is as follows: The parameter setting and updating of the mechanism-data fusion prediction model includes the following steps: First, in the historical stage, constrained parameter identification is performed based on the fused data. Using minimizing the weighted sum of prediction errors as the objective function, non-negativity and upper bound constraints are applied to the polysulfide species concentration, and [0,1] range and monotonicity constraints are applied to the absorbent activity and absorbent capacity. Initial estimates of the model parameters are obtained by solving this constrained optimization problem. Second, in the online stage, recursive least squares or Kalman averaging updates are used to update the mass transfer coefficient and reaction rate coefficient of the mechanism sub-model in real time.

[0062] Specifically, recursive least squares can be used to recursively update parameters based on new observation data in each sampling period. Alternatively, extended Kalman filtering or similar methods can be employed to jointly estimate parameters as extended state variables. Simultaneously, the temperature sensitivity coefficient and pH sensitivity coefficient are limited to preset ranges, such as ±20% for the temperature sensitivity coefficient and ±10% for the pH sensitivity coefficient. This avoids excessive fluctuations in parameter estimation. Finally, when the sensor health score decreases or the consistency error exceeds a threshold, the affected parameters are frozen and the weight of the data correction component is increased. For example, when the health score of a sensor falls below 0.8 or the consistency error exceeds 5%, updates to model parameters related to that sensor are paused, and the weight of the data correction component is increased by 50% to enhance the model's robustness to abnormal data. Through these technical solutions, this application achieves an organic combination of mechanistic modeling and data-driven methods, improving the model's prediction accuracy and adaptability. Constrained parameter identification ensures the physical rationality of the model parameters. The use of an online parameter update method enables the model to adaptively track changes in operating conditions. Setting limits on parameter variation ranges avoids excessive fluctuations in parameter estimation. Introducing sensor health scores and consistency error assessments enhances the model's robustness to abnormal data. Therefore, this approach can fully utilize online data to improve prediction accuracy while preserving the model's physical meaning, providing a reliable foundation for subsequent model predictive control.

[0063] This application further proposes constructing a constrained model predictive control problem within a prediction time window. The objective is to achieve the target total sulfur concentration at the outlet while minimizing the costs of injection and regeneration. Hard constraints are applied to the process and actuator boundaries to obtain the control command. During the control problem construction, the fusion prediction model explicitly includes actuator dynamics and time delays for feedforward prediction, ensuring that the control solution satisfies the process and actuator constraints. The objective function is set to satisfy the upper limit of total sulfur concentration at the outlet and minimize the weighted sum of injection and regeneration costs. Decision variables include the injection trajectory, regeneration intensity trajectory, and switching time. Hard constraints include the upper limit of total sulfur concentration at the outlet, the absorber pressure drop range, the liquid level range, the temperature and pressure range of the regeneration unit, the actuator opening range, the upper limit of the actuator opening change rate, and the minimum residence time at the switching time. The fusion prediction model characterizes the mass transfer-reaction and absorbent decay processes of polysulfide species through mechanism-data fusion, and explicitly embeds the actuator dynamic model and time delay compensation mechanism within the prediction time window.

[0064] Specifically, the objective function balances the cost of desulfurizing agent addition and regeneration energy consumption through weighted coefficients, and continuously optimizes the combination strategy of addition amount, regeneration intensity, and switching time within the prediction time window. Process boundary constraints are calculated in real time by the prediction model to control the pressure drop, liquid level, and temperature and pressure of the absorption tower and regeneration unit, ensuring that the equipment operates within a safe range. Actuator boundary constraints limit the opening range and rate of change to avoid overshoot or oscillation of the actuators. The fusion prediction model uses actuator dynamics and time delay as feedforward inputs to predict the impact of control commands on the process state, improving the executability of the control solution. When the solver detects a sudden change in operating conditions that leads to infeasibility, it automatically tightens the actuator rate of change constraints and postpones the switching time, restoring a feasible solution while ensuring that the total sulfur content at the outlet meets the standard. Through hierarchical objective priority settings, the feasibility of the upper limit constraint for total sulfur content at the outlet always takes precedence over economic optimization, ensuring emission compliance.

[0065] As a preferred embodiment, the solution of this application is implemented as follows: When constructing a constrained model predictive control problem within the prediction time domain window, the predicted state is used as the initial value, and the objective function is set to satisfy the upper limit of total sulfur at the outlet and minimize the weighted sum of the injection cost and the regeneration cost. The decision variables are the injection trajectory, the regeneration intensity trajectory, and the switching time. The hard constraints include: the total sulfur at the outlet does not exceed the upper limit of total sulfur at the outlet; the pressure drop of the absorber tower is within the pressure drop range; the liquid level of the absorber tower is within the liquid level range; the temperature of the regeneration unit is within the temperature range; the pressure of the regeneration unit is within the pressure range; the actuator opening is within the opening range; the rate of change of the actuator opening does not exceed the preset upper limit of the rate of change; and the switching time satisfies the minimum residence time.

[0066] Furthermore, the mechanism-data fusion prediction model explicitly includes actuator dynamics and actuator delay for feedforward prediction, thereby obtaining a control solution that satisfies the constraints within the prediction time domain window. Specifically, the following is a parameter example of another embodiment, which defines the control calculation layer timescale at the minute level and is applicable to low-temperature regeneration conditions. The prediction time domain window can be set to 30 minutes, and the sampling period is 1 minute. The weight ratio of the addition cost to the regeneration cost in the objective function can be set to 3:1. The upper limit of total sulfur at the outlet is set to 35 mg / m3, the pressure drop of the absorption tower is 2-5 kPa, the liquid level is 60%-80%, the temperature range of the regeneration unit is 40-60℃, and the pressure range is 0.1-0.3 MPa. The actuator opening range is 0-100%, the upper limit of the opening change rate is 2% / min, and the minimum residence time for switching is 60 min. It should be noted that the upper limit of total sulfur at the outlet can be ppm or mg / m3 according to the enterprise's implementation standards. 3 It is noted that the temperature range, upper limit of actuator opening rate of change, and minimum dwell time in different embodiments may vary due to differences in process routes, equipment specifications, and unified time scale definitions. Comparisons should be made using a unified time scale and the same operating condition. This is intended to illustrate parameter tuning methods and is not limited to a single value. Actuator dynamics can be described using a first-order inertia plus pure time delay model, with a time constant of 30s and a pure time delay of 15s. The prediction step size of the mechanism-data fusion prediction model is consistent with the sampling period, which is 1 minute. Accurate compensation for actuator dynamics and time delay is achieved by online identification and updating actuator model parameters. Thus, a constrained optimization problem is solved within the prediction time domain window, yielding a sequence of control solutions that satisfy all hard constraints. Only the control command for the current sampling period is executed, and the solution is updated in the next period, achieving closed-loop feedback control. Through the above technical solution, this application achieves joint prediction and optimization control of polysulfide species concentration, absorbent activity, and capacity. By explicitly considering actuator dynamics and time delay, prediction accuracy and control effect are improved. Simultaneously, comprehensive process and actuator constraints were applied to ensure the feasibility and safety of the control scheme. A rolling optimization strategy was adopted to enhance adaptability to fluctuations in operating conditions. Thus, while ensuring that the total sulfur content at the outlet meets the standard, the overall cost of addition and regeneration was minimized.

[0067] This application further proposes linearizing the mechanism-data fusion prediction model under the current operating condition in each sampling period, transforming the constrained model predictive control problem into a bounded quadratic programming problem for online solution. A hierarchical objective is adopted, prioritizing the feasibility of the outlet total sulfur upper limit constraint over the optimization of injection and regeneration costs. Only injection, regeneration intensity, and switching commands for the control solution in the current sampling period are issued, and the process rolls into the next sampling period. When a sudden change in operating conditions renders the problem temporarily infeasible, the actuator opening rate is automatically tightened and the switching time is delayed until the minimum dwell time is met, without relaxing the outlet total sulfur upper limit and process boundary. Specifically, the current operating condition linearization preserves the first-order dynamic characteristics of the mechanism-data fusion prediction model through Taylor expansion, transforming the nonlinear prediction equation into a linear time-varying model, ensuring the quadratic programming problem satisfies convex optimization conditions. The bounded quadratic programming uses the Lagrange multiplier method to handle inequality constraints, setting the upper limit of the actuator opening rate to 0.5%–2% of full scale per second. The hierarchical objective adopts a two-layer optimization structure. The priority layer constructs the feasible region using the upper limit of total sulfur at the outlet as a hard constraint. The optimization layer minimizes the weighted sum of the costs of addition and regeneration within the feasible region. The minimum residence time is set as an integer multiple of the regeneration cycle to ensure sufficient absorbent regeneration.

[0068] Specifically, within each sampling period, the mechanism-data fusion prediction model is first locally linearized based on the current process state, generating a linear time-varying state-space equation. Then, the optimization objective within the prediction time-domain window is transformed into a quadratic cost function, and the process boundary and actuator boundary are transformed into linear inequality constraints, forming a bounded quadratic programming problem. During the online solution process, priority is given to ensuring that the predicted total sulfur value at the outlet does not exceed a preset upper limit, followed by optimization of the dosing pump frequency and regeneration valve position settings. When a constraint conflict is detected, the upper limit of the actuator opening change rate is automatically reduced from the preset value to 50%–80% of the original value, and the switching time is postponed by at least one regeneration cycle. Finally, only the optimized solution at the current moment is output, achieving dynamic adjustment through a rolling time-domain mechanism. For example, when a sudden increase in gas flow causes the absorber level to approach its upper limit, the system automatically limits the regeneration valve opening change rate and postpones the switch to regeneration mode to ensure that the level constraint is not breached.

[0069] As a preferred embodiment, the solution of this application is implemented as follows: In each sampling period, the mechanism-data fusion prediction model is linearized according to the current operating condition, and the constrained model predictive control problem is transformed into a bounded quadratic programming problem for online solution. Specifically, firstly, a first-order Taylor expansion is performed on the nonlinear prediction model near the current operating point to obtain a linearized state-space model. Then, the linearized model is substituted into the objective function and constraints to construct a standard form quadratic programming problem. The decision variables are the control sequence in the future prediction time domain, the objective function is the weighted sum of the squares of the deviations between the total sulfur output and the setpoint and the changes in the control quantity, and the constraints include the state equation, the upper limit of the total sulfur output, the pressure drop range, the liquid level range, the temperature range, the pressure range, the actuator opening range, and the rate of change limit, etc.

[0070] Furthermore, a hierarchical objective is adopted to prioritize the feasibility of the total sulfur emission ceiling constraint at the outlet over the optimization of injection and regeneration costs. Specifically, the degree of constraint violation is used as the first-level objective, and injection and regeneration costs as the second-level objectives, with hierarchical optimization achieved by assigning different weight coefficients. Thus, only injection quantity commands, regeneration intensity commands, and switching commands for the control solution in the current sampling period are issued, and the process rolls into the next sampling period. For example, if the prediction time domain is 1 hour and the sampling period is 1 minute, then only the control commands for the next minute are executed each time, and a new control sequence is obtained by resolving the problem after 1 minute. When a sudden change in operating conditions makes the problem temporarily infeasible, the actuator opening rate is automatically tightened and the switching time is delayed until the minimum dwell time is met, without relaxing the total sulfur emission ceiling at the outlet and the process boundary.

[0071] Specifically, the first step is to reduce the upper limit of the actuator opening change rate by 50%, and if this is still infeasible, further reduce it to 25%. Simultaneously, the switching time is delayed until the minimum dwell time constraint is met. If no feasible solution is found after these adjustments, a conservative operation mode is triggered. Through this technical solution, this application achieves efficient online solving of constrained model predictive control problems. The hierarchical objective ensures the priority of achieving the total sulfur emission standard at the outlet, while also considering the optimization of the costs of injection and regeneration. The rolling optimization method improves the real-time performance and robustness of the control. Furthermore, by automatically adjusting the actuator change rate and switching time, the system's ability to cope with sudden changes in operating conditions is enhanced, avoiding control interruptions due to temporary infeasibility.

[0072] This application further proposes a scheme to perform feedforward compensation for control commands by combining actuator dynamics and time delay, and to execute control commands for the current sampling period in a predictive time-domain window rolling manner. Specifically, this includes: establishing actuator dynamic models for the dosing pump and regeneration valve respectively, with the actuator dynamic models adopting a first-order inertial plus pure time delay form and updating parameters online. Feedforward compensation is performed on the dosing quantity command, regeneration intensity command, and switching command based on the actuator dynamic models. This includes inverse calculation of the actuator opening-flow nonlinearity to obtain a linearized reference signal, constructing a Smith predictor for the actuator time delay to compensate for the time delay effect, and setting a ramp limiter on the reference signal to ensure that the actuator opening change rate does not exceed a preset upper limit. Dead-zone compensation is performed when the reference signal falls into the actuator dead zone, and saturation projection is performed to satisfy the actuator boundary when the reference signal exceeds the opening range. The compensated and limited signal remains unchanged within the current sampling period through a zero-order hold, and only the dosing quantity command, regeneration intensity command, and switching command for the current sampling period are issued, with the switching command satisfying the minimum dwell time constraint. The actuator dynamic model employs a first-order inertial model with added pure time delay. By online identification of the response time constant of the injection pump and the hysteresis parameters of the regeneration valve, the model parameters are dynamically adjusted to adapt to actuator aging or changes in operating conditions. Characteristic back-calculation establishes a nonlinear mapping relationship between actuator opening and flow rate, generating a linearized reference signal to eliminate the influence of actuator nonlinearity on control commands. The Smith predictor constructs a feedforward compensation loop based on actuator time delay parameters, incorporating the time delay effect into the forward calculation of the predictive model to avoid control deviations caused by time delay. The ramp limiter sets a slope threshold based on the upper limit of the actuator opening change rate, ensuring a smooth transition of control commands. Dead zone compensation uses superimposed offsets to ensure the reference signal avoids the actuator's insensitive area, and saturation projection forcibly constrains the reference signal exceeding the actuator's physical boundaries to within allowable limits.

[0073] Specifically, in the dosing pump control process, a dynamic model of the actuator is first constructed based on the online identified first-order inertial parameters and pure time lag. For example, the transfer function of the dosing pump can be expressed as G(s)=K / (Ts+1)e-τs, where K is the gain, T is the inertial time constant, and τ is the time lag. By acquiring actuator opening and flow data in real time, the T and τ parameters are updated online using the recursive least squares method. Subsequently, the dosing quantity command output by the model predictive control is input into the Smith predictor, which generates a compensated reference signal based on the time lag τ. Simultaneously, based on the pre-calibrated opening-flow curve, a piecewise linearization method is used to back-calculate the nonlinear characteristics, generating a linearized reference signal. The reference signal is processed by a ramp limiter to ensure that the opening change rate does not exceed a preset upper limit of 0.5% / s. When the reference signal is within the actuator dead zone, a fixed offset of 0.5% is added to cross the dead zone. When the signal exceeds the opening upper limit of 95%, saturation projection is performed to limit it to 95%. Finally, the compensated and limited signal is maintained within the current sampling period by a zero-order hold, ensuring that only the optimized command at the current moment is executed. During this process, the minimum dwell time constraint for the switching command is implemented through a time counter, ensuring that the regeneration unit operates stably for at least the aforementioned minimum dwell time (e.g., 30 minutes) before switching, consistent with the constraint setting. Through these steps, the impact of actuator dynamic characteristics and time delay on control accuracy is eliminated, synchronization between predicted commands and executed actions is achieved, and the stable compliance of the total sulfur concentration at the outlet is guaranteed.

[0074] As a preferred embodiment, the specific implementation of this application is as follows: A dynamic model of the actuator is established for the injection pump with a first-order inertial element time constant of 5 seconds and a pure time delay of 3 seconds. The dynamic model of the regeneration valve adopts a structure with a first-order inertial element time constant of 2 seconds and a pure time delay of 1 second. The gain coefficient of the inertial element is updated online using the recursive least squares method, and parameter updates are triggered when the valve opening change exceeds 10%. The flow-opening curve of the injection pump is piecewise linearized and inversely calculated to generate a feedforward compensation value for the flow reference signal. A Smith predictor is constructed to generate compensation for the time delay of the regeneration valve. The upper limit of the injection pump opening change rate is set to 5% per minute, and the upper limit of the regeneration valve rate is set to 15% per minute. The step command is converted into a ramp signal using a ramp limiter. When the flow reference signal falls into the dead zone interval of 5%–8% opening, a fixed offset of 3% is superimposed for dead zone compensation. Commands exceeding the 0–100% opening range are subject to boundary truncation processing, and a saturation projection operation is performed. The compensated control commands are maintained for a 20-second sampling period via a zero-order hold, outputting only the current cycle's dosing pump opening command (62%), regeneration valve opening command (85%), and switching delay command. The switching command triggers the regeneration unit switching action after the absorber has operated for at least 30 minutes. Through this technical solution, this application overcomes the control deviation caused by actuator dynamic lag and time delay effects, achieving precise matching between control commands and actuator responses through a feedforward compensation mechanism. The combined effect of nonlinear characteristic back-calculation and the Smith predictor eliminates the influence of actuator nonlinearity and transmission delay on the prediction model, while the ramp limiting mechanism prevents process parameter oscillations caused by actuator overshoot. Dead zone compensation and saturated projection processing ensure the effectiveness and safety of actuator actions, and the zero-order hold mechanism ensures the stability of control commands within the sampling period. Minimum residence time constraints avoid incomplete absorbent regeneration caused by frequent switching, improving system operational reliability.

[0075] This application further proposes a method for setting up safety barriers to switch between conservative and degraded operation when redundancy consistency checks fail or the predicted state will violate the upper limit of total sulfur at the outlet or the process boundary. Specifically, this includes generating consistency status flags, prediction violation flags, and feasibility flags based on consistency error, sensor health scores, and predicted state. When the consistency error exceeds a first threshold (or its safety trigger amplification threshold), the sensor health score is lower than a second threshold, the predicted state at any moment within the prediction time window indicates that the total sulfur at the outlet exceeds the upper limit, the process state will cross the process boundary, or the constraint-based model predictive control problem has no feasible solution, the operation mode switch is triggered. In conservative operation, meeting the upper limit of total sulfur at the outlet is prioritized, and a conservative baseline is used to set control commands while maintaining actuator boundary and rate of change constraints. Specifically, the conservative baseline setting refers to limiting the dosage command and regeneration intensity command to no less than the preset conservative dosage baseline and preset conservative regeneration baseline, respectively, without relaxing the aforementioned hard constraints, and performing delay or freezing processing on the switching command to prioritize ensuring the feasibility of achieving the total sulfur at the outlet standard. When the consistency status flag remains abnormal for a preset number of sampling periods or an actual out-of-bounds alarm occurs, the system enters a degraded operation phase. During this phase, the solution to the constrained model predictive control problem is stopped, and the online update of the mechanism-data fusion predictive model is frozen. Control commands are then set according to a rule-based conservative strategy. This strategy involves directly generating control commands based on pre-configured threshold-action mapping rules, without relying on online optimization in the current period. Prioritized regeneration prioritizes switching commands to or maintains regeneration conditions, while meeting minimum residence time constraints, to improve absorbent recovery capabilities and reduce the risk of total sulfur exceeding the outlet limit. When the redundancy consistency check returns to normal and the predicted state meets the upper limit of total sulfur at the outlet and the process boundary within a continuous preset sampling period, the system smoothly exits conservative or degraded operation with a ramp-limited approach and resumes rolling execution of the constrained model predictive control. Through these settings, the system's safety boundary and basic desulfurization capacity can be maintained in the event of sensor anomalies, model mismatches, or sudden changes in operating conditions, reducing the need for manual intervention.

[0076] As a preferred embodiment, the solution of this application is implemented as follows: In the total sulfur concentration monitoring system at the outlet of the absorption tower, an electrochemical sensor and an ultraviolet spectral sensor are redundantly configured as sulfur species detection units. When the output value of the electrochemical sensor exceeds the physical range of 0-200ppm, an out-of-bounds data screening and rejection mechanism is triggered. If the difference between the measurements of the two sensors exceeds 50ppm and continues for three sampling cycles, the consistency error is determined to be out of limit and a consistency status anomaly flag is generated. At this time, the moving horizon estimation module automatically increases the observation noise covariance weight, the prediction model freezes the mass transfer coefficient update and starts data correction component compensation. If the prediction status shows that the outlet total sulfur concentration will exceed the aforementioned outlet total sulfur upper limit (e.g., 30ppm) within the next five sampling cycles, a prediction violation flag is triggered. When the above-mentioned abnormal flag continues to exist, the control module switches to a conservative operation mode. The dosing pump command is set to be no less than the preset baseline value of 120L / min, the regeneration valve opening command is maintained above 50% and switching operations are prohibited. This stage corresponds to the switching delay / freeze processing in the aforementioned conservative baseline setting. The actuator dynamic compensation module limits the start-up ramp of the injection pump to ensure that the flow rate change does not exceed 10 L / min. 2 If the abnormal flag is not eliminated within ten sampling cycles and the actual outlet total sulfur concentration exceeds the aforementioned outlet total sulfur upper limit for a preset duration, a degraded operation mode will be entered: the dosing pump opening will be forcibly locked to 80%, and the regeneration intensity will be set to the conservative upper limit of the aforementioned temperature and pressure ranges. For example, under the aforementioned example constraints, the regeneration temperature can be set to 80℃ and the regeneration pressure to 0.3MPa. Simultaneously, after meeting the minimum residence time constraint, the switching command will prioritize the regeneration condition to correspond to the aforementioned priority regeneration strategy. When the sensor difference recovers to within 20ppm and the predicted state is compliant for five consecutive cycles, the actuator opening will smoothly recover to the normal control mode at a rate of 5% / min.

[0077] Through the above technical solutions, this application achieves a multi-layered safety protection mechanism under abnormal operating conditions, avoiding the risk of excessive emissions due to sensor failure or model mismatch. By automatically switching between conservative and degrading operating modes, it ensures that the system maintains basic desulfurization efficiency under extreme conditions, while preventing equipment damage caused by frequent actuator operations. The state switching mechanism based on dynamic flags significantly reduces the frequency of manual intervention, ensuring the continuous and stable operation of the blast furnace gas purification process.

[0078] In the above embodiments, the reliability of input data is ensured through redundancy consistency checks and weighted fusion of multi-source process data. Combined with state estimation via a sliding time-domain window, core process state information such as polysulfide species concentration, absorbent activity, and capacity is obtained. This allows the mechanism-data fusion prediction model to simultaneously characterize the mass transfer reaction process of polysulfides and the absorbent decay characteristics, improving prediction accuracy and adaptability. Constrained model predictive control within the prediction time-domain window achieves stable compliance with total sulfur emissions at the outlet, optimizing both addition and regeneration costs and addressing the shortcomings of existing methods in considering energy consumption and dynamic constraints. Dynamic and time-delay compensation mechanisms for actuators ensure consistency between predicted commands and actual execution. Combined with a safety barrier design, the system automatically switches to conservative or degraded operating modes in case of sensor malfunctions or sudden changes in operating conditions, avoiding the risks of relying solely on manual intervention. This improves the precision and economic efficiency of desulfurization control and enhances safety under complex operating conditions.

[0079] In another preferred embodiment based on the above embodiments, see [reference] Figure 2 As shown, this embodiment provides a blast furnace gas desulfurization system based on predictive control, used to apply the above-mentioned blast furnace gas desulfurization method based on predictive control, including: The data acquisition module is configured to acquire process data, including sulfur species sensor data, as well as temperature, pressure, flow rate, pH, pressure drop, liquid level, regeneration parameters, and actuator opening data. Redundancy consistency checks and weighted fusion are performed on the process data to obtain fused data.

[0080] The estimation module is configured to perform state estimation within a sliding time-domain window based on fused data to obtain the process state, which includes the concentration of polysulfide species, absorbent activity, and absorbent capacity.

[0081] The prediction module is configured to establish a mechanism-data fusion prediction model based on the process status, jointly characterize the mass transfer-reaction and absorbent decay process of polysulfide species, and output the predicted status.

[0082] The control module is configured to construct a constrained model predictive control problem within the prediction time window. With the goal of achieving the target total sulfur at the outlet and minimizing the cost of addition and regeneration, hard constraints are applied to the process boundary and actuator boundary to obtain control commands, which include addition amount command, regeneration intensity command and switching command.

[0083] The execution module is configured to perform feedforward compensation of control commands by combining actuator dynamics and time delay, and execute the control commands of the current sampling period in a rolling prediction time-domain window manner. A safety barrier is set to switch to conservative operation or degraded operation when redundancy consistency check fails or the predicted state will violate the upper limit of total sulfur at the outlet or the process boundary.

[0084] Through the above technical solutions, this application effectively solves the prediction bias problem caused by inaccurate modeling of the synergistic reaction kinetics of polysulfide species, and achieves dynamic compensation for the absorbent activity decay process through a mechanism-data fusion prediction model. The system explicitly incorporates the dynamic characteristics of the actuators during the control command generation stage, avoiding the control lag phenomenon caused by time delays in traditional methods. The combination of a safety barrier mechanism and a redundant data fusion strategy ensures that the system can still maintain the outlet sulfur concentration meeting the standard when sensors malfunction or operating conditions change abruptly, while rolling optimization reduces energy consumption fluctuations during the regeneration process.

[0085] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for fine desulfurization of blast furnace gas based on predictive control, characterized in that, include: Collect process data, including sulfur species sensor data and temperature, pressure, flow rate, pH, pressure drop, liquid level, regeneration parameters and actuator opening data. Perform redundancy consistency checks and weighted fusion on the process data to obtain fused data. Based on the fused data, state estimation is performed within a sliding time window to obtain the process state, which includes polysulfide species concentration, absorbent activity, and absorbent capacity. Based on the process status, a mechanism-data fusion prediction model is established to jointly characterize the mass transfer-reaction process of polysulfide species and the attenuation process of absorbent, and the predicted status is output. Within the prediction time window, a constrained model is constructed to predict the control problem. With the goal of achieving the target total sulfur at the outlet and minimizing the cost of addition and regeneration, hard constraints are applied to the process boundary and actuator boundary. The control commands are obtained by solving the problem. The control commands include addition amount command, regeneration intensity command and switching command. The control commands are fed forward and compensated by combining actuator dynamics and time delay, and the control commands of the current sampling period are executed in the rolling mode of the prediction time domain window; a safety barrier is set, and when the redundancy consistency check fails or the prediction state will violate the upper limit of total sulfur at the outlet or the process boundary, conservative operation or degraded operation is switched.

2. The blast furnace gas desulfurization method based on predictive control according to claim 1, characterized in that, When collecting process data and performing redundancy consistency checks and weighted fusion on the process data to obtain fused data, the process data includes: The sulfur species sensor includes at least two types of sensors with different measurement principles to acquire sulfur species sensor data, and simultaneously collects temperature data, pressure data, flow data, pH data, pressure drop data, liquid level data, regeneration parameter data and actuator opening data; The process data is uniformly time-scaled and resampled, and a preliminary screening based on physical boundaries and rate of change is performed to remove out-of-bounds and abrupt data. The consistency error is calculated for the outputs of different sensors for the same measurand and compared with the consistency threshold. At the same time, a residual test is established based on process constraints to determine that data that does not conform to material balance, energy balance or known process coupling relationship is abnormal data. The sensor health score is calculated based on the consistency error, residual size, noise intensity and historical stability, and the fusion weight of each data channel is updated with the health score and noise covariance. Weighted least squares fusion is used to perform weighted fusion on the data that passed the test to generate the fused data. The fused data provides a fused estimate of polysulfide species concentration, temperature, pressure, flow rate, pH, pressure drop, liquid level, regeneration parameters and actuator opening under a unified time scale. When the consistency error exceeds the first threshold or the sensor health score is lower than the second threshold, the corresponding data channel is downweighted.

3. The blast furnace gas fine desulfurization method based on predictive control according to claim 2, characterized in that, When performing state estimation in the sliding time domain based on the fused data to obtain the process state, the following are included: The sliding time-domain window is set to several sampling periods, and the state vector is constructed as polysulfide species concentration, absorbent activity and absorbent capacity. The discrete state equation and discrete measurement equation of the mechanism-data fusion prediction model are used as constraints, where the measurement equation takes the fused data as the observation input and the state equation takes the actuator opening data in the fused data as the control input. A moving horizon estimation objective function is established to minimize the weighted sum of observation error and state evolution error within the sliding time window. Non-negative and upper bound constraints are imposed on polysulfide species concentration, and range constraints of [0, 1] and smoothing constraints on the rate of state change are imposed on absorbent activity and absorbent capacity. The observation noise covariance and process noise covariance are adaptively updated based on sensor health scores and historical stability. When the consistency error exceeds the first threshold or the corresponding data channel is downweighted, the observation noise covariance of the corresponding observation in the measurement equation is increased, thereby solving for the process state within the sliding time-domain window.

4. The blast furnace gas desulfurization method based on predictive control according to claim 3, characterized in that, Based on the aforementioned process conditions, a mechanism-data fusion prediction model is established to jointly characterize the mass transfer-reaction and absorbent decay processes of polysulfide species. The output prediction status includes: Using the concentration of polysulfide species, absorbent activity, and absorbent capacity in the process state as state variables, a mechanism-data fusion prediction model is constructed. The mechanism-data fusion prediction model consists of a mechanism sub-model and a data correction component. The mechanism sub-model takes temperature, pressure, flow rate and pH as inputs to characterize the mass transfer-reaction coupling of hydrogen sulfide and carbon monoxide sulfides in the gas-liquid phase and the hydrolysis conversion of carbon monoxide sulfides to hydrogen sulfide, and describes the absorbent decay process with absorbent activity and absorbent capacity. The data correction component takes the fused data as input and dynamically corrects the systemic bias of the mechanism sub-model. The mechanism-data fusion prediction model is discretized and iterated at a sampling period consistent with the sliding time-domain window to output the prediction state. The prediction state includes the predicted concentration of polysulfide species, the predicted activity of absorbent, and the predicted capacity of absorbent, and the predicted value of total sulfur at the outlet is calculated accordingly.

5. The blast furnace gas desulfurization method based on predictive control according to claim 4, characterized in that, The parameter setting and updating of the mechanism-data fusion prediction model include: During the historical phase, constrained parameter identification is performed based on the fused data with the goal of minimizing the weighted sum of prediction errors. Non-negative and upper bound constraints are imposed on the concentration of polysulfide species, and range and monotonicity constraints of [0, 1] are imposed on the absorbent activity and absorbent capacity. During the online phase, recursive least squares or Kalman-type updates are used to update the mass transfer coefficient and reaction rate coefficient of the mechanism sub-model, and the temperature sensitivity coefficient and acid-base sensitivity coefficient are limited to a preset range of variation. When the sensor health score drops or the consistency error exceeds the threshold, the affected parameters are frozen and the weight of the data correction component is increased.

6. The blast furnace gas fine desulfurization method based on predictive control according to claim 1, characterized in that, When constructing a constrained model predictive control problem within a prediction time window, the following are included: Using the predicted state as the initial value, the objective function is set to satisfy the upper limit of total sulfur output and minimize the weighted sum of injection cost and regeneration cost. The decision variables are injection amount trajectory, regeneration intensity trajectory and switching time. The hard constraints include: the total sulfur at the outlet does not exceed the upper limit of the total sulfur at the outlet, the pressure drop of the absorber tower is within the pressure drop range, the liquid level of the absorber tower is within the liquid level range, the temperature of the regeneration unit is within the temperature range, the pressure of the regeneration unit is within the pressure range, the actuator opening is within the opening range, the rate of change of the actuator opening does not exceed the preset rate of change upper limit, and the switching time meets the minimum dwell time. The fusion prediction model explicitly includes actuator dynamics and actuator delay for feedforward prediction, thereby obtaining a control solution that satisfies the constraints within the prediction time window.

7. The blast furnace gas fine desulfurization method based on predictive control according to claim 6, characterized in that, When solving for control commands, the following steps are included: In each sampling period, the mechanism-data fusion prediction model is linearized to the current operating condition, and the constrained model predictive control problem is transformed into a bounded quadratic programming problem for online solution. A hierarchical objective is adopted to prioritize the feasibility of the total sulfur export ceiling constraint over the optimization of injection and regeneration costs; Only the control solution's dosage command, regeneration intensity command, and switching command for the current sampling period are issued, and the process rolls into the next sampling period; When a problem becomes temporarily infeasible due to a sudden change in operating conditions, the actuator opening rate is automatically tightened and the switching time is delayed until the minimum dwell time is met, without relaxing the upper limit of total sulfur at the outlet and the process boundary.

8. The blast furnace gas desulfurization method based on predictive control according to claim 7, characterized in that, When feedforward compensation is applied to the control commands based on actuator dynamics and time delay, and control commands for the current sampling period are executed using a predictive time-domain window rolling method, the following is included: Dynamic models of actuators are established for the dosing pump and the regeneration valve respectively. The dynamic models of actuators adopt the form of first-order inertia plus pure time delay and the parameters are updated online. The feedforward compensation for the dosing command, regeneration intensity command and switching command is performed according to the actuator dynamic model, including: performing characteristic back calculation on the actuator opening degree-flow nonlinearity to obtain a linearized reference signal, constructing a Smith predictor for the actuator time delay to compensate for the time delay effect, and setting a ramp limiter on the reference signal so that the actuator opening degree change rate does not exceed the preset change rate upper limit. Dead zone compensation is performed when the reference signal falls into the actuator dead zone, and saturation projection is performed to satisfy the actuator boundary when the reference signal exceeds the opening range. The compensated and limited signal remains unchanged in the current sampling period through a zero-order hold. Only the dosage command, regeneration intensity command and switching command for the current sampling period are issued, and the switching command satisfies the minimum dwell time constraint.

9. The method for fine desulfurization of blast furnace gas based on predictive control according to claim 8, characterized in that, Set up safety barriers to switch to conservative or degraded operation when redundancy consistency checks fail or the predicted state will violate the upper limit of total sulfur at the outlet or process boundary, including: Based on the consistency error, sensor health score, and predicted state, a consistency state flag, a predicted violation flag, and a feasibility flag are generated. When the consistency error exceeds the first threshold or the sensor health score is lower than the second threshold, or when the predicted state indicates that the total sulfur at the outlet exceeds the upper limit of the total sulfur at the outlet at any time within the prediction time window, or when the process state will cross the process boundary, or when the constrained model predictive control problem has no feasible solution, the operation mode switch is triggered. In conservative operation, the upper limit of total sulfur at the outlet is prioritized. The conservative baseline setting of the control command is adopted to ensure that the dosage command is not lower than the preset conservative dosage baseline and the regeneration intensity command is not lower than the preset conservative regeneration baseline. The switching command is delayed until the minimum residence time is met and a feasible solution is restored. At the same time, the actuator opening range and the actuator opening change rate constraint are maintained and the control command is issued according to the feedforward compensation and limiting strategy. When the consistency status flag continuously reaches the preset number of sampling cycles or an actual out-of-bounds alarm occurs, it enters degenerate operation. During degenerate operation, the solution of the constraint model predictive control problem is stopped and the online update of the mechanism-data fusion predictive model is frozen. The control command is set according to the rule-based conservative strategy, so that the dosage command takes the preset conservative coefficient of the upper limit of the actuator opening range, the regeneration intensity command takes the conservative upper limit of the temperature range and pressure range, and the switching command is set to priority regeneration and meets the minimum residence time. When the redundancy consistency test returns to normal and the predicted state meets the upper limit of total sulfur at the outlet and the process boundary within a continuous preset sampling period, the conservative operation or degradation operation is smoothly exited by the slope limit and the rolling execution of the constraint model predictive control is resumed.

10. A blast furnace gas desulfurization system based on predictive control, used for applying the blast furnace gas desulfurization method based on predictive control as described in any one of claims 1-9, characterized in that, include: The data acquisition module is configured to acquire process data, which includes sulfur species sensor data and temperature, pressure, flow rate, pH, pressure drop, liquid level, regeneration parameters and actuator opening data. The process data is subjected to redundancy consistency check and weighted fusion to obtain fused data. The estimation module is configured to perform state estimation within a sliding time-domain window based on the fused data to obtain the process state, which includes polysulfide species concentration, absorbent activity, and absorbent capacity. The prediction module is configured to establish a mechanism-data fusion prediction model based on the process state, jointly characterize the mass transfer-reaction and absorbent decay process of polysulfide species, and output the prediction state. The control module is configured to construct a constrained model predictive control problem within a prediction time window, with the goal of achieving the target total sulfur at the outlet and minimizing the cost of addition and regeneration. Hard constraints are applied to the process boundary and actuator boundary, and control commands are obtained by solving the problem. The control commands include addition amount command, regeneration intensity command and switching command. The execution module is configured to perform feedforward compensation on the control commands by combining actuator dynamics and time delay, and execute the control commands of the current sampling period in a predictive time domain window rolling manner; a safety barrier is set up to switch to conservative operation or degraded operation when the redundancy consistency check fails or the predicted state will violate the upper limit of total sulfur at the outlet or the process boundary.

Images