A method and device for dynamically coordinating and optimizing carbon pollution control parameters in steel production

Abstract

The application provides a steel production carbon pollution control parameter dynamic collaborative optimization method and device, collects full-process dynamic production data; constructs a digital twin model embedded with metallurgical mechanism constraints and adopting neural tangent kernel adaptive weight training, realizes high-fidelity prediction of carbon, pollution and energy; introduces external economic signals such as carbon price and electricity price, constructs a multi-objective optimization function, solves the Pareto solution set by a mechanism-guided particle swarm algorithm, selects the best control target by fuzzy satisfaction degree decision and issues it to the bottom layer hardware to form a closed-loop control. The application embeds mechanism constraints in the model, ensures prediction accuracy under complex working conditions, avoids algorithm divergence, and ensures control safety; through collaborative optimization of carbon, pollution and energy, the system is broken down and contradictions such as carbon reduction and pollution increase are resolved; time lag compensation is introduced to predict the peak of end pollutants in advance and reduce the risk of exceeding the standard; combined with feedforward and feedback control, accurate command tracking and anti-disturbance execution are ensured.

Description

Technical Field

[0001] This invention relates to the field of steel production process control technology, and in particular to a method and apparatus for dynamic collaborative optimization of carbon pollution control parameters in steel production. Background Technology

[0002] In existing technologies, there has been considerable research and application of energy optimization scheduling for steel enterprises. For example, optimization scheduling methods for energy media such as oxygen, coal gas, electricity, and steam in steel enterprises utilize demand forecasting models and optimization algorithms to balance the supply and demand of various energy media, achieving safe, stable, and efficient operation of the energy system. These solutions can improve the supply and demand balance of energy media and increase the operating efficiency of the energy system; however, their optimization targets are usually limited to a specific type of energy media or a specific energy subsystem, focusing on energy conservation or economical energy scheduling. Furthermore, for end-of-pipe treatment systems such as flue gas desulfurization, denitrification, and dust removal, there are already optimization control methods based on online monitoring, predictive models, and control strategies. For example, by establishing simulation models of desulfurization and denitrification systems and pollutant concentration prediction models, control parameters such as slurry spraying volume and ammonia injection volume are optimized. These solutions can improve the operational stability of environmental protection facilities and the ability to control pollutants to meet standards; however, they mainly focus on end-of-pipe treatment facilities and typically do not consider the overall optimization from the perspective of the coupling relationship between steel production source process parameters, energy consumption, and carbon emissions. This may result in problems such as increased energy consumption, increased reagent consumption, or increased carbon emissions despite achieving pollutant standards.

[0003] However, the aforementioned existing technologies have significant shortcomings. First, in traditional steel enterprise information systems, the Energy Management System (EMS), Environmental Monitoring System (CEMS), and Production Scheduling System (MES) are isolated from each other, making data sharing difficult. Second, most existing carbon emission data is based on manual calculations from month-end material balance sheets, resulting in significant time lags. The underlying industrial controllers cannot adjust equipment operating parameters in real time based on dynamic carbon footprints, real-time pollutant generation rates, and fluctuating carbon trading market prices and electricity spot prices. Furthermore, existing so-called intelligent scheduling methods often simply apply general neural network models to historical steel plant data, lacking a deep integration of the complex metallurgical physicochemical reaction mechanisms in steelmaking, leading to a significant decrease in prediction accuracy when production boundaries fluctuate. More importantly, most existing methods target only a single optimization objective—either reducing energy consumption, controlling carbon emissions, or reducing pollutant emissions—rarely incorporating carbon emissions, pollutant emissions, and energy consumption into a unified multi-objective optimization framework for collaborative solution. This single-objective or fragmented optimization approach can easily lead to a dilemma where reducing carbon emissions may increase pollution, reducing pollution may increase energy consumption, and saving energy may affect quality, making it difficult to achieve coordinated pollution reduction and carbon reduction throughout the entire steelmaking process.

[0004] Therefore, there is an urgent need for a solution that enables collaborative management and control of the entire process and achieves multi-objective optimization. Summary of the Invention

[0005] In view of this, embodiments of the present invention provide a method and apparatus for dynamic collaborative optimization of carbon emission and pollution control parameters in steel production, in order to solve the problems in existing steel production control where carbon emission, pollution and energy consumption data cannot be perceived in real time and coordinated across processes, and where pure data-driven models lack metallurgical mechanism constraints, resulting in control commands not being able to be safely issued to underlying equipment.

[0006] On the one hand, this invention provides a method for dynamic collaborative optimization of carbon pollution control parameters in steel production, which includes the following steps: Acquire dynamic production data for the ironmaking process, steelmaking process, rolling process and end-of-pipe treatment process in the steel production process. The dynamic production data includes process control parameters and operating status parameters. A digital twin model of the entire steel production process, pre-trained based on Transformer or graph neural network and constrained by the physicochemical mechanism of the metallurgical process, is obtained as the mechanistic model. The digital twin model of the entire steel production process takes graph-structured dynamic production data as input and outputs predictions of instantaneous carbon emission intensity, instantaneous pollutant generation rate, and comprehensive energy consumption efficiency. The dynamic production data is constructed for the steel production process, with nodes representing monitoring points in the steel production process and recording their operating status, and edges representing material and energy flow directions and parameters. The system acquires dynamic prices from the carbon trading market, pollution control penalty prices, and real-time spot prices from external electricity markets. It then constructs multiple optimization objectives targeting carbon emission costs, pollution control penalty costs, and energy consumption costs. These multiple optimization objectives incorporate a time-series lag factor for spatiotemporal lag compensation. A multi-objective particle swarm optimization algorithm, guided by the aforementioned mechanism model, is used to find Pareto solutions that satisfy the multiple optimization objectives for each of the process control parameters. Multi-attribute decision-making is then introduced, based on the principle of maximizing fuzzy satisfaction, to select the optimal solution from the Pareto solution set, thereby determining the current optimal control target and its corresponding target parameters. Based on the target parameters of the current optimal control target position and the current parameters recorded in the dynamic production data, a control signal executable by the underlying hardware is generated and sent out.

[0007] In some embodiments, the method further includes: using a sliding window mid-range filtering algorithm to remove abnormal noise points in the dynamic production data; and calculating the coefficient of variation of multiple specified parameters in the dynamic production data according to a set time window to construct a supplementary feature vector reflecting the flow field fluctuations in the blast furnace or the combustion disturbances in the heating furnace.

[0008] In some embodiments, the pre-training step of the digital twin model of the entire steel production process includes: A training sample set is obtained, which consists of historical production data of the ironmaking process, steelmaking process, rolling process and end-of-pipe treatment process in the steel production process. The historical production data includes historical process control parameters and their corresponding historical supplementary feature vectors. Based on historical material and energy consumption parameters and historical flue gas parameters before and after end-of-pipe treatment, the instantaneous carbon emission intensity, instantaneous pollutant generation rate and comprehensive energy consumption efficiency at each moment are calculated as labels. The historical production data in the training sample set is structured into a graph. The historical production data is constructed for the steel production process. The nodes are the monitoring points in the steel production process and record the operating status. The edges are the material and energy flow directions and parameters. The initial neural network model is trained using the training sample set. The initial neural network model includes an input layer, a spatiotemporal feature extraction layer, and an output layer. The input layer is used to import the structured historical production data and extract initial process features. The spatiotemporal feature extraction layer extracts spatiotemporal attention features from the initial process features based on the self-attention mechanism of the Industrial Transformer. The output layer uses a multilayer perceptron to fuse the spatiotemporal attention features and outputs predicted values ​​for instantaneous carbon emission intensity, instantaneous pollutant generation rate, and comprehensive energy consumption efficiency. A data supervision loss is constructed based on the deviation between the predicted value and the label, and a total loss function is obtained by weighting and introducing carbon material balance constraint term, heat balance constraint term and pollutant combustion reaction kinetic upper limit constraint term; the total loss function is dynamically reconstructed in real time using a neural tangent kernel adaptive dynamic weight algorithm; the parameters of the initial neural network model are updated by minimizing the total loss function to obtain the digital twin model of the entire steel production process.

[0009] In some embodiments, the carbon material balance constraint term is used to calculate the carbon input-output residual between the total carbon amount entering the steel production process and the sum of the carbon amount of the produced material and the carbon amount emitted in the gas phase, and to construct an asymmetric residual loss term as the carbon element conservation loss based on the carbon input-output residual. The thermal balance constraint term is used in the constraint to calculate the instantaneous thermal balance residual, and an asymmetric residual loss term is constructed based on the instantaneous thermal balance residual as the thermal balance loss; In the upper limit constraint term of the pollutant combustion reaction kinetics, max(0, The function constructs a penalty term as pollutant loss, which is activated when pollutant production exceeds the physical limit. The formula for calculating the total loss function is: ( = ( + ( + ( ( ; in, ( Represents the total loss function, ( This represents the data supervision loss. ( This represents the carbon element conservation loss. ( This represents the heat balance loss. ( This indicates the loss of the pollutants; , and These are the weighting coefficients for each loss term.

[0010] In some embodiments, the method further includes: For weighting coefficients , and An adaptive dynamic weighting algorithm for neural tangent kernels is introduced for updating, and the calculation formula is as follows: ; in, Indicates weighting coefficient exist The value at time, Indicates weighting coefficient exist The value at time; Indicate the impact of various losses on network parameters The gradient; Represents the absolute value of the gradient; β This is a smoothing factor.

[0011] In some embodiments, multiple optimization objectives are constructed for carbon emission costs, pollutant treatment penalty costs, and energy consumption costs, and the calculation formula is as follows: ； in, This refers to the carbon emission cost calculated using the dynamic price of the carbon trading market for each operation X in the process control parameters. This refers to the cost of pollutant treatment penalties calculated by introducing the pollutant treatment penalty price for each operation X in the process control parameters. This represents the energy cost calculated using the real-time spot price of electricity for each operation X in the process control parameters; a time lag factor is introduced for the variables therein. , is used to constrain the delayed response time of the operation adjustment of the j detection point by the i-th process.

[0012] In some embodiments, a multi-objective particle swarm optimization algorithm guided by the aforementioned mechanism model is employed to find the Pareto solution set satisfying the multiple optimization objectives for each item in the process control parameters, including: Let the position of each particle represent a set of process control parameters for steel production, and its velocity update formula is: ; The position update formula is: ; in, This represents the velocity of particle i in the d-th dimension of the control variable during the k-th iteration. Let w represent the position of particle i in the d-th dimension of the control variable during the k-th iteration, and w represent the inertia weight. and For individual and social learning factors, and A random number between 0 and 1. Let i be the optimal value of the control variable in the d-th dimension for the k-th generation particle i. It is the globally optimal solution selected from the external Pareto non-dominated solution set by the crowding distance up to generation k; Mechanism-guided correction terms; ; in, Indicating mechanism guiding factor, A random number between 0 and 1. This represents the theoretically optimal parameter.

[0013] In some embodiments, multi-attribute decision-making is introduced to select the optimal solution from the Pareto solution set based on the principle of maximizing fuzzy satisfaction, including: For the m-th objective function of the j-th Pareto solution Establish membership function The calculation formula is: ; The current optimal control target position satisfies: ; in, This represents the lower bound of the m-th objective function. This represents the upper bound of the m-th objective function. Let be the weighting coefficient, satisfying .

[0014] In some embodiments, based on the target parameters of the current optimal control target position and the current parameters recorded in the dynamic production data, a control signal executable by the underlying hardware is generated and sent out, including: The feedforward control benchmark is constructed using the following formula: ; in, The target parameters are those for the current optimal control target position. Represents the industrial control mapping function. This represents the feedforward control reference; The formula for calculating the feedback deviation correction value is as follows: ; e(t) = X opt (t) -Y real (t) ; in, This represents the feedback deviation correction value. Y real (t) This represents the current actual parameters of the current optimal control target position. e(t) Indicates parameter deviation. This indicates the cumulative amount of parameter deviation. This represents the rate of change of parameter deviation. , and These are the weighting coefficients; The final control signal is obtained based on the feedforward control reference and the feedback deviation correction value, calculated as follows: ; in, This indicates the final control signal issued.

[0015] On the other hand, the present invention also provides a dynamic collaborative optimization device for control parameters of the entire steel production process, including a processor, a memory, and a computer program or instructions stored in the memory. The processor is used to execute the computer program or instructions, and when the computer program / instructions are executed, the device implements the steps of the above method.

[0016] On the other hand, the present invention also provides a computer-readable storage medium having a computer program or instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.

[0017] On the other hand, the present invention also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the steps of the above-described method.

[0018] The present invention discloses a method and apparatus for dynamic collaborative optimization of carbon pollution control parameters in steel production. First, it utilizes a distributed edge computing gateway to collect dynamic production data in real time, including process control parameters, energy consumption parameters, and flue gas emission parameters from the ironmaking, steelmaking, and rolling processes. Then, it constructs a digital twin model based on an industrial Transformer or graph neural network. This model embeds metallurgical mechanism constraint equations such as carbon material balance, thermodynamic balance, and reaction kinetics into a graph structure. An adaptive dynamic weighting algorithm based on neural tangent kernels is introduced into the loss function to reconcile gradient conflicts between data-driven and physical constraints, thereby achieving high-fidelity prediction of instantaneous carbon emission intensity, instantaneous pollutant generation rate, and overall energy efficiency. Based on this, external economic signals such as carbon trading market prices, pollutant treatment penalty prices, and electricity spot prices are introduced to construct a multi-objective optimization function encompassing carbon emission costs, pollutant treatment penalty costs, and energy consumption costs. A multi-objective particle swarm optimization algorithm guided by the aforementioned mechanism model is used to achieve Pareto optimality. The optimal control target is selected from the solution set through the principle of maximizing fuzzy satisfaction. Finally, the optimized control parameters are converted into hardware execution signals to achieve closed-loop linkage regulation of the underlying equipment. This invention effectively prunes predictions that violate physical laws by embedding metallurgical mechanism constraints into a digital twin model. This ensures that the model's prediction accuracy remains high even under complex operating conditions such as drastic fluctuations in raw material composition. It fundamentally avoids the algorithm divergence problem caused by the lack of mechanism constraints in purely data-driven models, guaranteeing the safety and reliability of control command issuance. By constructing a multi-objective collaborative optimization framework encompassing carbon emissions, pollutant emissions, and energy consumption, and introducing dynamic economic signals such as external carbon prices and electricity prices, it breaks through the limitations of traditional isolated EMS, CEMS, and MES systems. This achieves a shift from single-objective or fragmented optimization to collaborative optimization of carbon, pollution, and energy, effectively solving the dilemma that reducing carbon emissions may increase pollution, and reducing pollution may increase energy consumption. By introducing a time-series lag factor into the optimization objective for spatiotemporal lag compensation, the system can predict peak pollutant generation in advance based on changes in preceding process operations, significantly reducing the risk of pollutant emissions exceeding standards. By employing a mechanism-guided multi-objective particle swarm optimization algorithm and a fuzzy satisfaction decision mechanism, the system achieves rapid solution and unique output of the optimal compromise solution under multiple conflicting objectives. Combined with a feedforward feedback joint control strategy, it ensures accurate tracking and disturbance-resistant execution of optimization instructions in the underlying PLC / DCS.

[0019] Additional advantages, objects, and features of the invention will be set forth in part in the description which follows, and will also become apparent in part to those skilled in the art upon studying the description, or may be learned by practice of the invention. The objects and other advantages of the invention can be realized and obtained by means of the structures specifically pointed out in the description and drawings.

[0020] Those skilled in the art will understand that the objectives and advantages achievable with the present invention are not limited to those specifically described above, and that the above and other objectives achievable with the present invention will become clearer from the following detailed description. Attached Figure Description

[0021] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, are not intended to limit the scope of the invention. In the drawings: Figure 1 This is a flowchart illustrating the dynamic collaborative optimization method for carbon pollution control parameters in steel production according to an embodiment of the present invention.

[0022] Figure 2 This is a schematic diagram of the pre-training process of the digital twin model of the entire steel production process in the dynamic collaborative optimization method for energy carbon pollution control parameters in steel production according to an embodiment of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and descriptions of this invention are used to explain the invention, but are not intended to limit the invention.

[0024] It should also be noted that, in order to avoid obscuring the invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.

[0025] It should be emphasized that the term "including / comprises" as used herein refers to the presence of a feature, element, step, or component, but does not exclude the presence or addition of one or more other features, elements, steps, or components.

[0026] In the current steel production control process, the energy management system, environmental monitoring system, and production scheduling system are isolated from each other, and carbon emission data relies on manual calculation. This results in the inability to perceive carbon emissions, pollution, and energy consumption data in real time and to utilize them collaboratively across processes. At the same time, existing intelligent scheduling methods generally lack deep integration with the physical and chemical reaction mechanisms of metallurgy, relying only on general neural network models to fit historical data. When production boundary changes such as raw material fluctuations occur, the prediction accuracy drops significantly. The control commands issued by the optimization algorithm cannot be safely and directly executed to the underlying industrial controller due to the lack of mechanism constraints, which can easily lead to safety accidents such as material suspension and furnace cooling. In addition, most existing solutions are single-objective optimizations and fail to incorporate carbon emissions, pollutant emissions, and energy consumption into a unified multi-objective collaborative optimization framework. It is difficult to take into account the dynamic balance between carbon reduction, pollution reduction, and energy conservation, and it cannot respond to real-time changes in external economic signals such as carbon trading market prices and electricity spot prices. Ultimately, this makes it difficult to implement collaborative governance of pollution reduction and carbon reduction throughout the entire steel process.

[0027] In view of this, the present invention provides a method for dynamic collaborative optimization of carbon pollution control parameters in steel production, such as... Figure 1 As shown, the method includes the following steps S101~S104: Step S101: Obtain dynamic production data for the ironmaking process, steelmaking process, rolling process and end-of-pipe treatment process in the steel production process. The dynamic production data includes process control parameters and operating status parameters.

[0028] Step S102: Obtain a digital twin model of the entire steel production process pre-trained based on Transformer or graph neural network and constrained by the physical and chemical mechanisms of the metallurgical process as a mechanism model; the digital twin model of the entire steel production process takes graph-structured dynamic production data as input and outputs predictions of instantaneous carbon emission intensity, instantaneous pollutant generation rate and comprehensive energy consumption efficiency; the dynamic production data is constructed for the steel production process, with nodes being monitoring points in the steel production process and recording the operating status, and edges being material and energy flow directions and parameters.

[0029] Step S103: Obtain dynamic prices in the carbon trading market, pollution control penalty prices, and real-time spot prices of external electricity. Construct multiple optimization objectives targeting carbon emission costs, pollution control penalty costs, and energy consumption costs. Introduce time-series lag factors for spatiotemporal lag compensation for these multiple optimization objectives. Employ a multi-objective particle swarm optimization algorithm guided by a mechanistic model to find Pareto solutions that satisfy the multiple optimization objectives for each process control parameter. Introduce multi-attribute decision-making based on the principle of maximizing fuzzy satisfaction to select the optimal solution from the Pareto solution set, thereby determining the current optimal control target position and its corresponding target parameters.

[0030] Step S104: Based on the target parameters of the current optimal control target position and the current parameters recorded in the dynamic production data, generate control signals executable by the underlying hardware and send them out.

[0031] In step S101, the pre-ironization processes in steel production encompass sintering, pelletizing, and blast furnace ironmaking, involving key steps such as blast furnace charging, hot blast stove blasting, hydrogen-rich gas injection, and pulverized coal injection, primarily completing the reduction and smelting of iron ore. The steelmaking process mainly includes converter smelting and electric arc furnace smelting, involving operations such as oxygen blowing, slag formation, and alloying, converting molten iron into crude steel. The rolling process, centered on the heating furnace, heats, rolls, and cools steel billets to form various steel products. The end-of-pipe treatment process purifies the flue gas and wastewater generated throughout the entire process, mainly including flue gas desulfurization, denitrification, dust removal, and wastewater treatment to ensure pollutant emissions meet standards. These processes are closely interconnected in terms of materials, energy, and emissions, forming a complex industrial process with long cycles and significant time lags.

[0032] Correspondingly, process control parameters are core monitorable and adjustable operating parameters covering all processes in the entire steel production process, directly determining the production operation status and carbon / pollution / energy emission levels. Specifically, these include: in the ironmaking process, sintering machine trolley speed, ignition temperature, main exhaust fan frequency, and blast furnace hot blast temperature, hot blast pressure, oxygen enrichment rate, pulverized coal injection flow rate, furnace top pressure, charging chute angle, and furnace section temperature; in the steelmaking process, converter oxygen lance position, top-blown oxygen flow rate, bottom-blown argon flow rate, electric arc furnace electrode current, and blowing time; in the rolling process, furnace temperature, air-fuel ratio, gas flow rate, combustion air flow rate, rolling mill rolling pressure, and roller conveyor speed; and in the end-of-pipe treatment process, desulfurization tower slurry circulation pump frequency, denitrification ammonia injection flow rate, induced draft fan speed, and bag filter pulse jet pressure. Operating status parameters are another type of dynamic production data that complements process control parameters. They are mainly used to characterize the actual operating conditions, material characteristics, and output status of each process at a specific moment, rather than directly adjustable control quantities. Specifically, the operational status parameters include: raw material parameters, such as the grade, sulfur content, and moisture content of the iron ore fed into the furnace; the fixed carbon content and ash content of the coke; the volatile matter and calorific value of the pulverized coal; and the purity of the limestone; intermediate product parameters, such as the temperature, silicon content, and sulfur content of molten iron; the tapping temperature of the converter; the oxygen activity of the molten steel; and the surface temperature of the continuously cast billet; production and emission monitoring parameters, such as the strength of the sintering drum, the surface quality defect rate of the finished steel, and the flue gas flow rate, flue gas temperature, measured concentrations of sulfur dioxide, nitrogen oxides, and particulate matter, and the chemical oxygen demand and ammonia nitrogen content of the wastewater at the inlet of the end-of-pipe treatment facility; in addition, equipment operational status parameters, such as the vibration value of the fan bearings, the temperature of the motor windings, and the operating current of key pump groups, reflecting the health status of the equipment. These operational status parameters provide crucial data support for the real-time updating, condition identification, and constraint verification of the digital twin model.

[0033] All parameters are collected at a high frequency of seconds from the field PLC / DCS system through a distributed edge computing gateway. These parameters serve as the core input features of the digital twin model, mapping the production status in real time. They are also the direct control objects that need to be dynamically adjusted in subsequent optimization stages, covering key control nodes throughout the entire chain from raw material reaction and product processing to pollutant treatment.

[0034] In some embodiments, the method further includes: using a sliding window mid-range filtering algorithm to remove abnormal noise points in the dynamic production data; and calculating the coefficient of variation of multiple specified parameters in the dynamic production data according to a set time window to construct a supplementary feature vector reflecting the flow field fluctuations in the blast furnace or the combustion disturbances in the heating furnace.

[0035] In step S102, the digital twin model of the entire steel production process serves as the core sensing, prediction, and simulation foundation for the entire dynamic collaborative optimization scheme. Using a production topology structure to analyze multi-source production data across the entire process, it breaks down the data silos of traditional energy, environmental protection, and production systems. By embedding a high-fidelity model constrained by metallurgical physicochemical mechanisms, it replaces the traditional month-end manual carbon emission accounting and isolated monitoring of single systems. It can accurately perceive the instantaneous carbon emission intensity, pollutant generation rate, and overall energy efficiency of the entire process within seconds, and maintains prediction accuracy even under boundary scenarios such as raw material fluctuations and changes in production load, thus addressing the root cause of the problem. This avoids the safety issues associated with purely data-driven black-box models that output results that violate physical laws. Simultaneously, this model serves as a virtual simulation sandbox for control parameters, rapidly simulating carbon, pollution, and energy trends under different combinations of control parameters. It provides reliable simulation results for multi-objective optimization without requiring trial and error on actual production lines. Furthermore, it can predict pollutant and carbon emission peaks in advance, providing a basis for feedforward control. Ultimately, through continuous iterative correction based on feedback from actual production data, it becomes a reliable bridge connecting the real physical world of production with intelligent optimization algorithms, supporting the entire system's transformation from traditional static experience-based control to dynamic multi-objective collaborative control.

[0036] Specifically, such as Figure 2 As shown, the pre-training steps of the digital twin model of the entire steel production process include steps S201 to S204: Step S201: Obtain the training sample set, which consists of historical production data of the ironmaking process, steelmaking process, rolling process and end-of-pipe treatment process in the steel production process. The historical production data includes historical process control parameters and their corresponding historical supplementary feature vectors. Calculate the instantaneous carbon emission intensity, instantaneous pollutant generation rate and comprehensive energy efficiency at each moment based on the historical material and energy consumption parameters and the flue gas parameters before and after historical end-of-pipe treatment as labels.

[0037] Step S202: Construct a graph structure for the historical production data in the training sample set. The historical production data is constructed for the steel production process. The nodes are the monitoring points in the steel production process and record the operating status. The edges are the material and energy flow directions and parameters.

[0038] Step S203: Train the initial neural network model using a training sample set. The initial neural network model includes an input layer, a spatiotemporal feature extraction layer, and an output layer. The input layer is used to import structured historical production data and extract initial process features. The spatiotemporal feature extraction layer extracts spatiotemporal attention features from the initial process features based on the self-attention mechanism of the Industrial Transformer. The output layer uses a multilayer perceptron to fuse the spatiotemporal attention features and outputs predicted values ​​for instantaneous carbon emission intensity, instantaneous pollutant generation rate, and comprehensive energy consumption efficiency.

[0039] Step S204: Construct a data supervision loss based on the deviation between the predicted value and the label, and introduce a weighted total loss function by adding carbon material balance constraint, heat balance constraint, and upper limit constraint of pollutant combustion reaction kinetics. The total loss function is dynamically reconstructed in real time by using a neural tangent kernel adaptive dynamic weight algorithm. The parameters of the initial neural network model are updated by minimizing the total loss function to obtain a digital twin model of the entire steel production process.

[0040] Specifically, step S201 completes the automatic construction of training samples, covering historical production data of the entire process from ironmaking to steelmaking, rolling, and end-of-pipe treatment. In addition to historical process control parameters, supplementary feature vectors such as raw material composition fluctuations and furnace condition parameter variation coefficients are extracted. Based on historical material and energy consumption and flue gas composition before and after treatment, instantaneous carbon emission intensity, pollutant generation rate, and comprehensive energy efficiency at each moment are automatically generated as supervision labels through material balance calculations, thus constructing a large-scale, high-quality training set without manual annotation. Subsequently, step S202 completes the graph-structured organization of the data, constructing a graph according to the actual production topology. A directed graph attaches historical data to corresponding nodes based on the physical location of monitoring points, using the actual flow direction of materials and energy as directed edges and pre-storing parameter propagation lag attributes to achieve spatial alignment of multi-source heterogeneous data. After entering the model forward propagation stage, the initial neural network in step S203 is a three-layer industrial-specific network optimized for long-process time-series data in the steel industry. The input layer is responsible for dimensional normalization and fixed-length time window slicing of the graph-structured multi-source data, extracting initial process features from different monitoring points, and eliminating dimensional differences between parameters. The spatiotemporal feature extraction layer is the core structure of the model, distinct from... The fully connected self-attention computation of the general-purpose Transformer layer performs constrained attention computation based on the adjacency relationship of the production graph. Each node only calculates spatial attention weights with its upstream and downstream neighbor nodes that have material / energy associations. Simultaneously, self-attention computation is performed along the time dimension on the historical time-series window of a single node, extracting the spatial correlation features of parameters between processes and the temporal trend features of dynamic parameter changes. This approach not only aligns with the real causal logic of steel production and avoids meaningless false cross-process associations but also significantly reduces computational load. The output layer uses a multilayer perceptron to perform nonlinear fusion of spatiotemporal features, outputting instantaneous... Initial values ​​for three types of indicators—carbon emission intensity, pollutant generation rate, and comprehensive energy consumption efficiency—are predicted. Finally, step S204 completes parameter iteration with hard constraints on mechanisms. Based on the mean square error data supervision loss between the predicted values ​​and the labels, three types of mechanism residual losses—carbon element balance, thermal balance, and pollutant reaction limit—are added with weights. The weights of each loss term are dynamically adjusted through the neural tangent kernel algorithm to prevent the mechanism constraints from being submerged by the data loss gradient. The network parameters are iteratively updated by minimizing the total loss through backpropagation, ultimately obtaining a high-fidelity digital twin model that fits historical production patterns and strictly satisfies metallurgical physical rules.

[0041] In some embodiments, the carbon material balance constraint term calculates the carbon input-output residual between the total carbon amount entering the steel production process and the sum of the carbon amount of the produced material and the carbon amount emitted in the gas phase, and constructs an asymmetric residual loss term as the carbon element conservation loss based on the carbon input-output residual.

[0042] For example, during implementation, the instantaneous flow rate and carbon mass fraction of all input materials are collected and summed at each process and time step to obtain the total input carbon flow rate. Then, the flow rate and carbon content of all products and by-products are collected and summed to obtain the carbon flow rate carried out by the materials. Combined with the instantaneous carbon emission intensity and instantaneous output predicted by the model, the gas phase emission carbon flow rate is calculated, and finally, the carbon balance residual is obtained. Since a negative residual will overestimate the emission reduction potential and lead to the risk of excessive fuel reduction in the furnace, an asymmetric penalty is set: the penalty coefficient for negative residuals is 3, and the penalty coefficient for positive residuals is 1. The carbon loss is obtained by weighted averaging of the squared residuals within the batch.

[0043] The thermal balance constraint term is used in the constraint to calculate the instantaneous thermal balance residual, and an asymmetric residual loss term is constructed based on the instantaneous thermal balance residual as the thermal balance loss.

[0044] For example, during implementation, the total input heat power at each time step is first calculated, including fuel combustion heat, electrical energy, and sensible heat of the feed. Then, the fixed heat loss is calculated using furnace wall temperature and cooling water parameters. Combined with the comprehensive energy efficiency predicted by the model and the theoretically required effective heat under the current operating conditions, the input power corresponding to the effective heat is calculated to obtain the heat balance residual. Since a positive residual will overestimate the energy-saving potential and cause safety accidents such as insufficient fuel or insufficient furnace temperature, an asymmetric penalty is set: the penalty coefficient for positive residuals is 4, and the penalty coefficient for negative residuals is 1. The heat loss is obtained by batch weighted averaging.

[0045] In the upper limit constraint term of pollutant combustion reaction kinetics, max(0, The function constructs a penalty term as pollutant loss, which is activated when pollutant production exceeds the physical limit.

[0046] For example, during implementation, the total flow rate of target elements such as sulfur and nitrogen in the feed at each time step is first calculated, and multiplied by the highest kinetic conversion rate of the element to pollutants in that process and the molecular weight conversion coefficient of the element to pollutants to obtain the upper limit of physical generation of pollutants; only when the model prediction value exceeds the upper limit is the residual calculated by max(0, prediction value - upper limit), and the residual is 0 and no penalty is imposed when it does not exceed the upper limit, and the batch weighted average is used to obtain the pollutant loss.

[0047] The formula for calculating the total loss function is: ( = ( + ( + ( ( ; in, ( Represents the total loss function. ( Indicates data supervision loss, ( This represents the loss due to the conservation of carbon. ( Indicates heat balance loss, ( Indicates the loss of pollutants; , and These are the weighting coefficients for each loss term.

[0048] In some embodiments, the method further includes: adjusting the weighting coefficients. , and An adaptive dynamic weighting algorithm for neural tangent kernels is introduced for updating, and the calculation formula is as follows: ; in, Indicates weighting coefficient exist The value at time, Indicates weighting coefficient exist The value at time; Indicate the impact of various losses on network parameters The gradient; Represents the absolute value of the gradient; β This is a smoothing factor.

[0049] In step S103, three time-varying cost targets are constructed. The first is energy consumption cost, which is the real-time cost of all energy consumption, including electricity cost multiplied by the real-time spot electricity price, with the weight of electricity consumption automatically increased during peak electricity price periods. The second is environmental protection cost, including penalties for exceeding pollutant standards and the cost of desulfurization and denitrification agents. The third is carbon emission cost, which is the cost of purchasing carbon in excess of the quota multiplied by the real-time carbon trading price, with the weight of carbon reduction automatically increased when carbon quotas are insufficient at the end of the quarter.

[0050] In some embodiments, multiple optimization objectives are constructed for carbon emission costs, pollutant treatment penalty costs, and energy consumption costs, and the calculation formula is as follows: ； in, This refers to the carbon emission cost calculated using dynamic prices from the carbon trading market for each operation X in the process control parameters. This represents the cost of pollution treatment penalties calculated by introducing the pollution treatment penalty price for each operation X in the process control parameters. This represents the energy cost calculated using the real-time spot price of electricity from external sources for each operation X in the process control parameters; a time lag factor is introduced for the variables in this calculation. , is used to constrain the delayed response time of the operation adjustment of the j detection point by the i-th process.

[0051] Furthermore, a multi-objective particle swarm optimization algorithm guided by a mechanistic model is employed to find the Pareto solution set satisfying multiple optimization objectives for each item in the process control parameters, including: Let the position of each particle represent a set of process control parameters for steel production, and its velocity update formula is: The position update formula is: .

[0052] in, This represents the velocity of particle i in the d-th dimension of the control variable during the k-th iteration. Let w represent the position of particle i in the d-th dimension of the control variable during the k-th iteration, and w represent the inertia weight. and For individual and social learning factors, and A random number between 0 and 1. Let i be the optimal value of the control variable in the d-th dimension for the k-th generation particle i. It is the globally optimal solution selected from the external Pareto non-dominated solution set by the crowding distance up to generation k; Mechanism-guided correction terms; ; in, Indicating mechanism guiding factor, A random number between 0 and 1. This represents the theoretically optimal parameter.

[0053] In some embodiments, multi-attribute decision-making is introduced to select the optimal solution in the Pareto solution set based on the principle of maximizing fuzzy satisfaction, including: For the m-th objective function of the j-th Pareto solution Establish membership function The calculation formula is: ; The current optimal control target position satisfies: ; in, This represents the lower bound of the m-th objective function. This represents the upper bound of the m-th objective function. Let be the weighting coefficient, satisfying .

[0054] In step S104, a preset industrial mapping function is used. The abstract optimal control target position is converted into hardware setpoints that can be recognized by the PLC / DCS, such as valve opening, inverter frequency, and equipment speed. Then, based on the actual values ​​fed back by the sensors in real time, the PID calculation feedback correction is improved to offset random disturbances in the field such as pipeline pressure fluctuations and valve jamming, and to avoid large jumps in parameters.

[0055] Specifically, based on the target parameters of the current optimal control target position and the current parameters recorded in the dynamic production data, control signals executable by the underlying hardware are generated and issued, including: The feedforward control benchmark is constructed using the following formula: ; in, The target parameters for the current optimal control target position. Represents the industrial control mapping function. This indicates the feedforward control reference.

[0056] The formula for calculating the feedback deviation correction value is as follows: ; e(t) = X opt (t) -Y real (t) ; in, This indicates the feedback deviation correction value. Y real (t) This represents the current actual parameters of the optimal control target position. e(t) Indicates parameter deviation. This indicates the cumulative amount of parameter deviation. This represents the rate of change of parameter deviation. , and These are the weighting coefficients.

[0057] The final control signal is obtained based on the feedforward control reference and the feedback deviation correction value. The calculation formula is as follows: ; in, This indicates the final control signal issued.

[0058] On the other hand, the present invention also provides a dynamic collaborative optimization device for control parameters of the entire steel production process, including a processor, a memory, and a computer program or instructions stored in the memory. The processor is used to execute the computer program or instructions, and when the computer program / instructions are executed, the device implements the steps of the above method.

[0059] On the other hand, the present invention also provides a computer-readable storage medium having a computer program or instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.

[0060] On the other hand, the present invention also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the steps of the above-described method.

[0061] The present invention will now be described with reference to a specific embodiment: This embodiment provides a dynamic collaborative optimization method for control parameters throughout the entire steelmaking process, specifically including the following steps: Step 1: Real-time acquisition and feature edge processing of multi-source heterogeneous dynamic production data Utilizing a distributed edge computing gateway, dynamic production data from the ironmaking, steelmaking, rolling, and end-of-pipe treatment processes are collected in real time. Data types include: process control parameters such as gas flow rate, blast temperature, furnace pressure, and rotational speed; energy consumption parameters such as instantaneous coke weighing, instantaneous power consumption of electricity meters, and hydrogen-rich injection flow rate; and flue gas parameters before and after end-of-pipe treatment facilities, such as instantaneous concentrations of SO2, NOx, particulate matter, and CO, and flue gas velocity collected by real-time monitoring instruments. Edge processing: A sliding window midpoint filtering algorithm is used to remove abnormal noise points, and the coefficient of variation within a specific time window is calculated as a feature vector reflecting flow field fluctuations within the blast furnace or combustion disturbances in the heating furnace.

[0062] Step 2: Construct a high-fidelity digital twin model of carbon, pollution, and energy that integrates metallurgical mechanism constraints. A deep learning model based on Industrial Transformer or Graph Neural Network (GNN) will be established as a digital twin model of the entire steel production process to predict the instantaneous carbon emission intensity under the current operating state of the process. instantaneous formation rate of pollutants and overall energy efficiency During the training process, the digital twin model's loss function is forcibly embedded with a set of metallurgical physicochemical mechanism constraint equations, including carbon material balance constraints, iron oxide reduction thermodynamic balance constraints, and pollutant combustion reaction kinetic constraints. Through these mechanism constraints, algorithmic divergent outputs that do not conform to physical laws are pruned, ensuring that the digital twin model maintains a prediction accuracy of over 96% even when raw material composition fluctuates drastically.

[0063] To capture the dynamic temporal characteristics and spatial topological relationships of complex industrial sites, the process flow is constructed as a graph structure G = (V, E), where V is a key monitoring point, such as the furnace top, injection port, and tapping port, and E is the material and energy flow direction.

[0064] At time t, a multidimensional process parameter vector matrix with a historical window length of T. Represented as: , where the feature vector Material input vector This can include raw material flow rate, carbon and sulfur content, moisture content, etc. Thermal state vector This can include parameters such as temperature, pressure, oxygen enrichment rate, and air volume for each section. Production rate parameters. This can include instantaneous output, hourly output, etc.

[0065] The spatiotemporal feature extraction layer in the deep learning model, combined with the self-attention mechanism of the industrial Transformer, shows the spatiotemporal attention coefficients between the center node i and its neighboring nodes j and k at time t. The calculation formula is: ; Among them, the Leaky ReLU linear unit is an activation function in deep learning. , , This is the initial feature vector of a node, such as a vector composed of the original process parameters of node i at the top of the blast furnace at time t, including temperature, pressure, and carbon monoxide concentration.

[0066] This is a learnable linear mapping matrix, i.e., a spatial projection scale. Due to the different dimensions of the original process parameters, Its function is to project and transform these low-level data features into a high-dimensional feature space that can characterize the carbon, pollution, and energy spaces. This is the vector after feature enhancement. This indicates a feature splicing operation. This represents the learnable weight vector for the attention mechanism. This represents the set of spatial neighbors of node i. For example, in a process flow, it represents the set of all upstream and downstream equipment directly connected to blast furnace i via materials and pipelines. The high-level fully connected layer (MLP) outputs purely data-driven instantaneous prediction initial values: .

[0067] in, These are the weight parameters of the neural network.

[0068] The constraint equations based on metallurgical physicochemical mechanisms are introduced, including the following three categories: 1) The mass conservation constraint term for carbon element, for set up.

[0069] According to the law of conservation of mass, the total carbon content entering the process is equal to the sum of the carbon content of the output material and the carbon content emitted in the gas phase.

[0070] ; in, , denoted as the instantaneous mass flow rate (t / h) of input raw material i and output product / byproduct j, respectively. , This represents the corresponding mass fraction of carbon. Let be the instantaneous product output (t / h) at time t. Under ideal conditions, the mechanism residual... .

[0071] 2) The first law of thermodynamics and the thermal equilibrium constraint terms, for set up.

[0072] The comprehensive energy efficiency (ECEt) is defined as the proportion of useful heat to the total input energy of the system. Useful heat includes sensible heat and effective enthalpy change from endothermic reactions. The total input energy includes fuel chemical energy, electrical energy, and sensible heat. The heat balance residual equation is: ; in, The instantaneous heat released by fuel combustion depends on the air volume and the calorific value of the fuel: = LHV, of which LHV represents the amount of fuel at time t, and LHV is the calorific value of the fuel. This refers to the electrical input power (kW). (t) represents the instantaneous theoretically required enthalpy change calculated based on the metallurgical reaction mechanism. The theoretical heat loss due to heat dissipation from the furnace surface and the cooling water carried away is usually calculated analytically based on the physical formula of furnace wall temperature.

[0073] 3) Elemental limits and upper limits of reaction kinetics constraints, for Settings, taking SO2 as an example In addition to being controlled by the fluid dynamics of the reaction, the rate of pollutant formation also has an absolute material limit (such as the conservation of sulfur).

[0074] ; in, Input the sulfur mass fraction of material s; This represents the maximum conversion rate limit for sulfur to gaseous SO2 in this process. = =2.0, which is the molecular weight conversion factor from sulfur to sulfur dioxide. This constraint utilizes max(0, The function acts as a penalty, activating the penalty when the network's predicted value exceeds the physical limit.

[0075] Furthermore, a physical information-driven joint loss function is constructed. During the training of a multi-task physical information neural network (PINN), ill-conditioned gradients frequently occur: the network refuses to converge due to conflicts or significant differences in magnitude between the gradient of the data loss and the gradient of the physical constraints. To address this issue, this model introduces a Neural Tangent Kernel (NTK) adaptive dynamic weight algorithm to dynamically reconstruct the total loss function in real time: the model's total loss function... It is no longer just label error, but is composed of both data supervision loss and mechanistic residual loss: ( = ( + ( + ( ( ; Dynamic weights The adaptive update formula is as follows: ; in, Indicate the impact of various losses on network parameters The gradient; β represents the absolute value of the gradient. This is a smoothing factor. When the gradient of a certain mechanism loss is too small and is ignored by the network, the algorithm will automatically increase its penalty coefficient, forcing the deep learning model to converge in a direction that satisfies the laws of thermodynamics.

[0076] Step 3: Constructing a multi-objective spatial equation based on spatiotemporal lag compensation and external boundary constraints Given the long steelmaking process, there is a significant time lag between changes in upstream control parameters and changes in end-stage flue gas and carbon emissions. This step constructs a multi-objective collaborative optimization spatial equation that considers spatiotemporal lag compensation: ； in, This refers to the carbon emission cost calculated using dynamic prices from the carbon trading market for each operation X in the process control parameters. This represents the cost of pollution treatment penalties calculated by introducing the pollution treatment penalty price for each operation X in the process control parameters. This represents the energy cost calculated using the real-time spot price of electricity from external sources for each operation X in the process control parameters; a time lag factor is introduced for the variables in this calculation. Implement spatiotemporal lag compensation to constrain the delayed response time of the i-th process's operation adjustment to the j-th detection point.

[0077] Step 4: Pareto optimal solution and underlying hardware closed-loop linkage control An improved multi-objective particle swarm optimization (MOPSO) algorithm based on mechanism guidance is used to solve the above multi-objective spatial equations online in the whole space, dynamically approximating the Pareto optimal solution set.

[0078] The basic steps of the algorithm are as follows: 1) Determine the model inputs, which are carbon emissions, pollutant emissions, energy consumption, and product quality parameters, with constraints such as material balance and energy conservation. 2) Initialize the population by randomly generating an initial population P(0) according to the range of values ​​of the optimization variables, t=0, with a population size of N, and t being the generation number of the population, initially 0. 3) Determine whether the initial population meets the constraints. If not, return to step 2 to continue generating the initial population. 4) Perform genetic operations to generate a child population Q(t) of size N, and merge P(t) and Q(t) into R(t). 5) Calculate the fitness function of individuals in R(t) and perform non-dominated sorting. Here, the fitness function is the objective function. 6) Calculate the crowding function D and select individuals with high crowding to enter the new parent population P(t+1). 7) Select individuals with high fitness to enter the new parent population P(t+1), perform genetic operations, and select, crossover, and mutate individuals in the parent generation to generate the next generation population. 8) Repeat the above process continuously. When the termination condition is met, output the individuals in the population as the Pareto optimal solution. Since the algorithm can infinitely approach the optimal solution set in practical applications, and the number of iterations is preset by the program, it is necessary to determine whether the fitness function converges after a certain number of iterations. Generally, it is considered that the average rate of change of the fitness function over several consecutive generations is small. This process of determining the convergence of the fitness function is the termination condition of model optimization.

[0079] 1) Constructing the particle state update formula In the multi-object space, the position of each particle i represents a set of steel control parameters. Its velocity and position update formulas are: , .

[0080] in, This represents the velocity of particle i in the d-th dimension of the control variable during the k-th iteration. Let w represent the position of particle i in the d-th dimension of the control variable during the k-th iteration, and w represent the inertia weight. and For individual and social learning factors, and A random number between 0 and 1. Let i be the optimal value of the control variable in the d-th dimension for the k-th generation particle i. It is the globally optimal solution selected from the external Pareto non-dominated solution set by the crowding distance up to generation k; Mechanism-guided correction terms; ; in, Indicating mechanism guiding factor, A random number between 0 and 1. This represents the theoretically optimal parameter.

[0081] 2) Pareto non-dominated solution set maintenance and congestion distance For multi-objective optimization problems During the iteration process, the non-dominated solution set is updated to ensure a uniform distribution of solutions and prevent all solutions from clustering in a single local area. By prioritizing the retention of particles with large crowding distances, the solution set covers the entire preference range from extremely low cost to extremely environmentally friendly, ensuring that suitable parameters can be found regardless of adjustments to the production strategy. The crowding distance for particles in the solution set is calculated using the following formula: 3) Solution selection based on dynamic approximation and external decision-making mechanisms After a digital twin system dynamically approximates the Pareto solution set P online, it cannot send a bunch of solutions to the hardware. Instead, it must introduce multi-attribute decision-making to output a single, currently optimal control target position. The principle of maximizing fuzzy satisfaction is used for screening, and the m-th objective function of the j-th Pareto solution is selected. Establish membership function The calculation formula is: ; The current optimal control target position satisfies: ; in, This represents the lower bound of the m-th objective function. This represents the upper bound of the m-th objective function. Let be the weighting coefficient, satisfying .

[0082] 4) Underlying hardware closed-loop linkage control Optimized control target position This needs to be converted into control signals executable by the underlying PLC / DCS, and closed-loop feedback should be used for disturbance rejection. The system automatically selects a set of optimal control target vectors from the solution set based on preset operating strategies, such as environmental priority mode or extreme low-cost mode, and directly converts them into control parameter adjustment instructions. These instructions are then sent to the corresponding underlying industrial controller (PLC / DCS) via the OPC-UA protocol for hardware closed-loop adjustment. Furnace control: Dynamically change the setpoint (SP) of the air-fuel ratio valve opening control loop to optimize combustion, reduce CO2 and suppress the formation of high-temperature NOx from natural gas; Blast furnace control: The flow controller of the hydrogen-rich gas injection valve at the tuyere is linked to adjust the angle matrix of the top charging chute to suppress carbon emissions while ensuring stable furnace conditions; End-of-pipe treatment linkage: Based on the feedforward prediction of the peak sulfide formation within the next 15 minutes, the frequency converter frequency and spray flow rate of the desulfurization tower slurry circulation pump are increased in advance to avoid lag exceeding the standard, while preventing the increase in power consumption caused by excessive spraying during off-peak periods.

[0083] 5) Dynamic feedforward-feedback joint control model Combining digital twin predictive response, dynamic matrix control (DMC) or improved PID control is used to generate control increments for the underlying actuators. : The feedforward control benchmark is constructed using the following formula: ; in, The target parameters for the current optimal control target position. Represents the industrial control mapping function. This indicates the feedforward control reference.

[0084] The formula for calculating the feedback deviation correction value is as follows: ; e(t) = X opt (t) -Y real (t) ; in, This indicates the feedback deviation correction value. Y real (t) This represents the current actual parameters of the optimal control target position. e(t) Indicates parameter deviation. This indicates the cumulative amount of parameter deviation. This represents the rate of change of parameter deviation. , and These are the weighting coefficients.

[0085] The final control signal is obtained based on the feedforward control reference and the feedback deviation correction value. The calculation formula is as follows: ; in, This indicates the final control signal issued.

[0086] 6) Closed-loop timing of intelligent control for pollution reduction and carbon reduction throughout the entire process The mathematical logic of the entire process is as follows, forming a high-frequency closed loop: a digital twin is established based on production data, and after spatial alignment, fuzzy decision-making is performed to obtain the optimal target position. The closed-loop control solves the multi-objective Pareto optimization to obtain the final adjustment.

[0087] High-frequency data injection: Read sensor data streams from the plant's distributed control system (DCS) and online flue gas analyzer (CEMS).

[0088] Spatial alignment and causal inference: Automatic correction and feature alignment of production delays are performed through spatiotemporal graph networks.

[0089] Mechanism-constrained correction prediction: The PINN model output simultaneously satisfies the laws of mass conservation and thermodynamics. .

[0090] Multi-objective Pareto online reverse optimization: Based on the current high-fidelity prediction model, with the goals of minimizing carbon emissions, maximizing energy efficiency, and optimizing output value, the non-dominated sorting genetic algorithm (NSGA-III) is used to perform millisecond-level simulation optimization in the digital space. The optimal control parameters such as oxygen enrichment rate, pulverized coal injection rate, and rotary kiln speed are generated in reverse and directly sent to the PLC physical loop to achieve intelligent closed-loop management of carbon, pollution, and energy.

[0091] As can be seen, this embodiment breaks the isolation barrier and achieves collaborative decoupling, realizing for the first time the direct mapping of external economic / policy indicators to the physical control parameters of underlying industrial equipment. Experimental data shows that, compared with traditional independently operating EMS and CEMS systems, the overall energy consumption of the entire process is effectively reduced. Emission intensity is reduced, and the cost of end-of-pipe desulfurization and denitrification agents is also reduced.

[0092] This embodiment achieves safety and controllability, preventing algorithm divergence. Because the digital twin model incorporates inequality constraints based on metallurgical mechanisms, the control commands issued by the optimization algorithm (such as the material distribution matrix and hydrogen-rich injection limits) all fall within the safe operating window. During a continuous six-month operation test, neither the blast furnace nor the electric furnace experienced any abnormal furnace conditions due to algorithm misadjustment.

[0093] This embodiment has a high dynamic feedforward capability. By utilizing a spatiotemporal lag compensation mechanism, it transforms the end-of-pipe environmental compliance control from a passive catch-up after exceeding the standard to a feedforward adjustment 15 to 20 minutes in advance, which greatly reduces the risk of pollutant exceeding the standard.

[0094] In summary, the dynamic collaborative optimization method and device for carbon pollution control parameters in steel production described in this invention first relies on a distributed edge computing gateway to collect dynamic production data in real time, including process control parameters, material and energy consumption parameters, and flue gas emission parameters in the ironmaking, steelmaking, and rolling processes. Secondly, it constructs a digital twin model based on an industrial Transformer or graph neural network. This model embeds metallurgical mechanism constraint equations such as carbon material balance, thermodynamic balance, and reaction kinetics into a graph structure, and introduces an adaptive dynamic weighting algorithm based on neural tangent kernels into the loss function to effectively coordinate data-driven and physical constraints. The gradient conflict between the control bundles is addressed to achieve high-fidelity prediction of instantaneous carbon emission intensity, instantaneous pollutant generation rate, and overall energy efficiency. Furthermore, external economic signals such as carbon trading market prices, pollutant treatment penalty prices, and electricity spot prices are introduced to establish a multi-objective optimization function encompassing carbon emission costs, pollutant treatment penalty costs, and energy consumption costs. A multi-objective particle swarm optimization algorithm guided by a mechanistic model is used to solve for the Pareto optimal solution set, and the current optimal control target is selected from these solutions using the fuzzy satisfaction maximization principle. Finally, the optimized control parameters are converted into hardware execution signals to achieve closed-loop linkage regulation of the underlying equipment.

[0095] This invention achieves several technical advantages: First, by embedding metallurgical mechanism constraints into the digital twin model, it effectively suppresses prediction results that violate physical laws, ensuring a high level of model prediction accuracy even under complex operating conditions such as drastic fluctuations in raw material composition. This fundamentally avoids the algorithm divergence problem caused by the lack of mechanism constraints in purely data-driven models, effectively guaranteeing the safety and reliability of control command issuance. Second, by constructing a multi-objective optimization framework that coordinates carbon, pollution, and energy, and introducing dynamic external economic signals such as carbon prices and electricity prices, it breaks the limitations of the isolation between traditional EMS, CEMS, and MES systems, realizing a shift from single-objective or fragmented optimization to collaborative optimization, effectively resolving the dilemma that carbon reduction may increase pollution and pollution reduction may increase energy consumption. Third, by introducing a time-series lag factor into the optimization objective for spatiotemporal lag compensation, the system can predict the peak of end-of-pipe pollutant generation in advance based on changes in preceding process operations, significantly reducing the risk of pollutant emissions exceeding standards. Fourth, through the mechanism-guided multi-objective particle swarm optimization algorithm and fuzzy satisfaction decision mechanism, the optimal compromise solution under multiple conflicting objectives is solved quickly and output uniquely. Combined with the feedforward-feedback joint control strategy, the optimization instructions are accurately tracked and executed against disturbances in the underlying PLC / DCS.

[0096] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this invention are programs or code segments used to perform the desired tasks. The programs or code segments can be stored in a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried in a carrier wave.

[0097] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.

[0098] In this invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or in place of features of other embodiments.

[0099] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations of the embodiments of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for dynamic collaborative optimization of carbon pollution control parameters in steel production, characterized in that, The method includes the following steps: Acquire dynamic production data for the ironmaking process, steelmaking process, rolling process and end-of-pipe treatment process in the steel production process. The dynamic production data includes process control parameters and operating status parameters. A digital twin model of the entire steel production process, pre-trained based on Transformer or graph neural network and constrained by the physicochemical mechanism of the metallurgical process, is obtained as the mechanistic model. The digital twin model of the entire steel production process takes graph-structured dynamic production data as input and outputs predictions of instantaneous carbon emission intensity, instantaneous pollutant generation rate, and comprehensive energy consumption efficiency. The dynamic production data is constructed for the steel production process, with nodes representing monitoring points in the steel production process and recording their operating status, and edges representing material and energy flow directions and parameters. The system acquires dynamic prices from the carbon trading market, pollution control penalty prices, and real-time spot prices from external electricity markets. It then constructs multiple optimization objectives targeting carbon emission costs, pollution control penalty costs, and energy consumption costs. These multiple optimization objectives incorporate a time-series lag factor for spatiotemporal lag compensation. A multi-objective particle swarm optimization algorithm, guided by the aforementioned mechanism model, is used to find Pareto solutions that satisfy the multiple optimization objectives for each of the process control parameters. Multi-attribute decision-making is then introduced, based on the principle of maximizing fuzzy satisfaction, to select the optimal solution from the Pareto solution set, thereby determining the current optimal control target and its corresponding target parameters. Based on the target parameters of the current optimal control target position and the current parameters recorded in the dynamic production data, a control signal executable by the underlying hardware is generated and sent out.

2. The method for dynamic collaborative optimization of carbon pollution control parameters in steel production according to claim 1, characterized in that, The method further includes: An abnormal noise point in the dynamic production data is removed using a sliding window mid-value filtering algorithm. In addition, a supplementary feature vector reflecting the flow field fluctuations or combustion disturbances in the heating furnace is constructed by calculating the coefficient of variation of multiple specified parameters in the dynamic production data according to a set time window.

3. The method for dynamic collaborative optimization of carbon pollution control parameters in steel production according to claim 2, characterized in that, The pre-training steps of the digital twin model of the entire steel production process include: A training sample set is obtained, which consists of historical production data of the ironmaking process, steelmaking process, rolling process and end-of-pipe treatment process in the steel production process. The historical production data includes historical process control parameters and their corresponding historical supplementary feature vectors. Based on historical material and energy consumption parameters and historical flue gas parameters before and after end-of-pipe treatment, the instantaneous carbon emission intensity, instantaneous pollutant generation rate and comprehensive energy consumption efficiency at each moment are calculated as labels. The historical production data in the training sample set is structured into a graph. The historical production data is constructed for the steel production process. The nodes are the monitoring points in the steel production process and record the operating status. The edges are the material and energy flow directions and parameters. The initial neural network model is trained using the training sample set. The initial neural network model includes an input layer, a spatiotemporal feature extraction layer, and an output layer. The input layer is used to import the structured historical production data and extract initial process features. The spatiotemporal feature extraction layer extracts spatiotemporal attention features from the initial process features based on the self-attention mechanism of the Industrial Transformer. The output layer uses a multilayer perceptron to fuse the spatiotemporal attention features and outputs predicted values ​​for instantaneous carbon emission intensity, instantaneous pollutant generation rate, and comprehensive energy consumption efficiency. A data supervision loss is constructed based on the deviation between the predicted value and the label, and a total loss function is obtained by weighting and introducing carbon material balance constraint term, heat balance constraint term and pollutant combustion reaction kinetic upper limit constraint term; the total loss function is dynamically reconstructed in real time using a neural tangent kernel adaptive dynamic weight algorithm; the parameters of the initial neural network model are updated by minimizing the total loss function to obtain the digital twin model of the entire steel production process.

4. The method for dynamic collaborative optimization of carbon pollution control parameters in steel production according to claim 3, characterized in that, In the carbon material balance constraint term, the carbon input-output residual between the total carbon amount entering the steel production process and the total carbon amount of the produced material and the carbon amount emitted in the gas phase is calculated, and an asymmetric residual loss term is constructed based on the carbon input-output residual as the carbon element conservation loss. The thermal balance constraint term is used in the constraint to calculate the instantaneous thermal balance residual, and an asymmetric residual loss term is constructed based on the instantaneous thermal balance residual as the thermal balance loss; In the upper limit constraint term of the pollutant combustion reaction kinetics, max(0, The function constructs a penalty term as pollutant loss, which is activated when pollutant production exceeds the physical limit. The formula for calculating the total loss function is: ( = ( + ( + ( ( ； in, ( Represents the total loss function, ( This represents the data supervision loss. ( This represents the carbon element conservation loss. ( This represents the heat balance loss. ( This indicates the loss of the pollutants; , and These are the weighting coefficients for each loss term.

5. The method for dynamic collaborative optimization of carbon pollution control parameters in steel production according to claim 4, characterized in that, The method further includes: For weighting coefficients , and An adaptive dynamic weighting algorithm for neural tangent kernels is introduced for updating, and the calculation formula is as follows: ； in, Indicates weighting coefficient exist The value at time, Indicates weighting coefficient exist The value at time; Indicate the impact of various losses on network parameters The gradient; Represents the absolute value of the gradient; β This is a smoothing factor.

6. The method for dynamic collaborative optimization of carbon pollution control parameters in steel production according to claim 1, characterized in that, Construct multiple optimization objectives targeting carbon emission costs, pollutant treatment and penalty costs, and energy consumption costs, with the following calculation formula: ； in, This refers to the carbon emission cost calculated using the dynamic price of the carbon trading market for each operation X in the process control parameters. This refers to the cost of pollutant treatment penalty calculated by introducing the pollutant treatment penalty price for each operation X in the process control parameters. This represents the energy cost calculated using the real-time spot price of electricity for each operation X in the process control parameters; a time lag factor is introduced for the variables therein. , is used to constrain the delayed response time of the operation adjustment of the j detection point by the i-th process.

7. The method for dynamic collaborative optimization of carbon pollution control parameters in steel production according to claim 1, characterized in that, A multi-objective particle swarm optimization algorithm guided by the aforementioned mechanism model is used to find the Pareto solution set that satisfies the multiple optimization objectives for each item in the process control parameters, including: Let the position of each particle represent a set of process control parameters for steel production, and its velocity update formula is: ； The position update formula is: ; in, This represents the velocity of particle i in the d-th dimension of the control variable during the k-th iteration. Let w represent the position of particle i in the d-th dimension of the control variable during the k-th iteration, and w represent the inertia weight. and For individual and social learning factors, and A random number between 0 and 1. Let i be the optimal value of the control variable in the d-th dimension for the k-th generation particle i. It is the globally optimal solution selected from the external Pareto non-dominated solution set by the crowding distance up to generation k; Mechanism-guided correction terms; ； in, Indicating mechanism guiding factor, A random number between 0 and 1. This represents the theoretically optimal parameter.

8. The method for dynamic collaborative optimization of carbon pollution control parameters in steel production according to claim 1, characterized in that, Multi-attribute decision-making is introduced to select the optimal solution from the Pareto solution set based on the principle of maximizing fuzzy satisfaction, including: For the m-th objective function of the j-th Pareto solution Establish membership function The calculation formula is: ； The current optimal control target position then satisfies: ； in, This represents the lower bound of the m-th objective function. This represents the upper bound of the m-th objective function. For the weighting coefficients, satisfying .

9. The method for dynamic collaborative optimization of carbon pollution control parameters in steel production according to claim 1, characterized in that, Based on the target parameters of the current optimal control target position and the current parameters recorded in the dynamic production data, generate and send control signals executable by the underlying hardware, including: The feedforward control benchmark is constructed using the following formula: ； in, The target parameters are those for the current optimal control target position. Represents the industrial control mapping function. This represents the feedforward control reference; The formula for calculating the feedback deviation correction value is as follows: ； e(t) = X opt (t) -Y real (t) ； in, This represents the feedback deviation correction value. Y real (t) This represents the current actual parameters of the current optimal control target position. e(t) Indicates parameter deviation. This indicates the cumulative amount of parameter deviation. This indicates the rate of change of parameter deviation. , and These are the weighting coefficients; The final control signal is obtained based on the feedforward control reference and the feedback deviation correction value, calculated as follows: ； in, This indicates the final control signal issued.

10. A dynamic collaborative optimization device for control parameters throughout the entire steel production process, comprising a processor, a memory, and a computer program or instructions stored in the memory, characterized in that, The processor is configured to execute the computer program or instructions, and when the computer program / instructions are executed, the device implements the steps of the method as described in any one of claims 1 to 9.

11. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method as described in any one of claims 1 to 9.

12. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 9.

Images