Sintering ignition multi-valve combined regulation and control method based on PINN-PMPC

Abstract

The invention discloses a PINN-PMPC-based sintering ignition multi-valve combined regulation and control method, and belongs to the technical field of metallurgy. According to the method, based on actual sintering ignition physical logic, a multi-stage PINN network containing temperature, flow and pressure prediction sub-networks is constructed, and physical constraints are embedded to form a system steady-state equation; and then an MPC optimization objective function considering temperature tracking precision and valve action smoothness is constructed. Forward deducing a temperature track through a particle swarm algorithm, solving a future optimal opening sequence of the valve through rolling optimization, and finally executing the control action of the current step to realize closed-loop stable control of the ignition temperature; the prediction precision of the model is improved, so that the precision control of the ignition temperature can be realized; meanwhile, the model is periodically updated to adapt to system changes, and long-term control performance is ensured.

Description

Technical Field

[0001] This invention belongs to the field of metallurgical technology, and more specifically, relates to a multi-valve joint control method for sintering ignition based on PINN-PMPC. Background Technology

[0002] As a crucial step in steel production, the sintering process directly impacts the yield, quality, and energy consumption of sintered ore due to the quality of its ignition control. A modern industrial sintering furnace ignition system is a complex, multivariable coupled system involving the coordinated control of multiple subsystems, including the gas main, air system, and ignition furnace. During ignition, the system needs to maintain the ignition furnace temperature stable around the set ignition temperature to ensure sufficient ignition of the sintering bed and the formation of a suitable sintered ore structure.

[0003] Traditional sintering ignition control relies primarily on manual experience and simple PID control strategies. However, with the increasing scale and automation of sintering processes, the complexity of ignition systems has significantly increased. In actual production, factors such as fluctuations in raw material composition, changes in ambient temperature, and equipment aging cause continuous fluctuations in pressure parameters across the system. These pressure fluctuations are transmitted layer by layer through the flow-pressure-valve coupling relationship, ultimately affecting the temperature stability of the ignition furnace. Traditional single-variable control methods cannot effectively handle this complex multi-variable coupling problem, resulting in large fluctuations in ignition temperature and impacting the quality of the sintered ore.

[0004] A search revealed a sintering combustion control system in patent CN116624886A. This application selects flow control when adjusting the sintering material thickness, automatically adjusting the opening of the gas or air regulating valve intermittently based on the deviation between the target flow rate and the feedback flow rate. When the sintering ignition temperature is stable, intensity control is selected, automatically setting the gas flow rate based on the deviation between the set intensity and the calculated actual intensity. When the temperature and gas flow rate deviation is large, the deviation between the total gas flow rate and the furnace temperature is detected to calculate the required adjustment value for the coke oven gas regulating valve, ultimately adjusting the valve setting. This application uses segmented logic judgment for valve fine-tuning by setting a gas / air flow rate deviation threshold. However, its rule-based empirical adjustment lacks modeling of the system's changing processes, making high-precision coordinated control difficult to achieve.

[0005] For example, patent CN117707070A discloses a boiler or gas turbine combustion optimization method and system based on PINN. This application applies Physical Information Neural Network (PINN) to boiler / gas turbine combustion optimization, constructing a high-fidelity model by embedding physical constraints such as conservation equations, and combining Bayesian optimization to solve for the optimal valve opening offline. However, the control model in this application is an offline model, lacking real-time control of the dynamic process, and therefore cannot guarantee the accuracy of real-time control.

[0006] The above applications all involve technical improvements to the sintering ignition control system, but due to the lack of system modeling or the lack of online control optimization mechanisms, they are unable to cope with the strong disturbances and time-varying nature in the sintering ignition process, thus failing to guarantee the accurate control of the ignition temperature. Summary of the Invention

[0007] The problem to be solved To address at least some of the problems existing in the prior art, this invention proposes a multi-valve joint control method for sintering ignition based on PINN-PMPC, which aims to solve the problem that existing sintering ignition control systems cannot accurately control the ignition temperature.

[0008] Technical solution To solve the above problems, the technical solution adopted by the present invention is as follows: This invention discloses a multi-valve joint control method for sintering ignition based on PINN-PMPC, which achieves stable temperature control of the ignition furnace by adjusting the opening degrees of the ignition furnace valve, the air valve, and the main gas pipe valve. The steps include: S1, constructing the ignition steady-state equation. S11, Data Acquisition Key parameters during the sintering ignition process are collected by PLC, including ( ), air pressure ( ), Ignition furnace pressure ( ), gas main flow rate ( ), airflow ( ), Ignition furnace flow rate ( ), Gas main pipe valve opening ( ), air valve opening ( ), Ignition furnace valve opening ( ) and real-time temperature detection ( ); S12, Hierarchical PINN Network Design Based on the progressive physical logic of the sintering ignition process, namely “gas / air flow rate → ignition furnace pressure → ignition furnace flow rate → temperature”, prediction sub-networks for the corresponding stages are constructed. The four-stage prediction sub-networks are connected in series according to the physical logic, and physical constraints are embedded to construct the total loss function, thus obtaining the PINN temperature prediction model. S13. Using the PINN temperature prediction model as the system steady-state equation, establish a state transition function that describes the relationship between the current state and the future temperature. S2, PMPC Optimization Objective Function Design A multi-objective optimization objective function is constructed, which includes a temperature tracking error term and a valve action penalty term. The temperature tracking error term is used to penalize the deviation between the predicted temperature and the set value, and the valve action penalty term is used to penalize drastic changes in valve opening. S3, PMPC rolling optimization solution Based on the state transition function and the optimization objective function, the optimal valve opening is solved using a rolling optimization algorithm.

[0009] In some implementations, in step S12, constructing the four-stage prediction subnetwork specifically involves: Phase 1: Constructing a subnetwork for predicting gas / air flow rates Based on the main gas pipe pressure and the main gas pipe valve opening, the main gas pipe flow rate is predicted to independently reflect the flow characteristics of the gas system. Its steady-state relationship is as follows: ; Simultaneously, based on air pressure and air valve opening, air flow is predicted to independently reflect the flow characteristics of the air system. Its steady-state relationship is: ; Phase 2: Constructing the Ignition Furnace Pressure Prediction Subnetwork Based on the mixing effect of gas and air flow, it predicts the gas pressure inside the ignition furnace, and its steady-state relationship is as follows: ; Phase 3: Constructing the Ignition Furnace Flow Prediction Subnetwork Based on the ignition furnace pressure and the ignition furnace valve opening, the output flow rate of the ignition furnace is predicted, and its steady-state relationship is as follows: ; Phase 4: Constructing the temperature prediction subnetwork Based on the ignition furnace flow rate, the final electric furnace temperature is predicted, and its steady-state relationship is as follows: .

[0010] In some implementations, the physical constraints in step S12 include the following monotonicity constraints: Gas flow constraint: That is, the greater the gas pressure, the greater the flow rate; That is, the larger the opening of the gas valve, the greater the flow rate; Airflow constraints: In other words, the greater the air pressure, the greater the flow rate. In other words, the larger the air valve opening, the greater the flow rate. Ignition furnace pressure constraints: That is, the greater the gas flow rate, the higher the pressure inside the furnace; That is, the greater the airflow, the higher the pressure inside the furnace; Ignition furnace flow constraint: That is, the greater the pressure inside the furnace, the greater the flow rate; That is, the larger the opening of the ignition furnace valve, the greater the flow rate; Temperature constraints: That is, the higher the flow rate of the ignition furnace, the higher the combustion temperature; The five monotonicity constraints are transformed into the corresponding set of monotonic losses for the model as follows: .

[0011] In some implementations, the physical constraints in step S12 further include temperature asymptotic constraints: , The temperature asymptotic constraint is transformed into the temperature gradient loss corresponding to the model as follows:

[0012] Adding the physical loss and data loss together, PINN's total loss function is:

[0013] in, It's data loss. ; For the first Predicted values ​​for each sample, No. The true value of each sample; , , These are the weighting coefficients for data loss, monotonicity loss set, and temperature variability loss, respectively. The final ignition temperature prediction model is: .

[0014] In some implementations, the state transition function in step S13 is:

[0015] Where k represents the current time, and k+j represents j future times. This represents the predicted temperature of the furnace at the current time k, and the predicted temperature at the future time k+j.

[0016] In some embodiments, the temperature tracking error term in step S2 is, ; The temperature tracking error term is calculated and predicted in the time domain. Internal step-by-step temperature prediction With target temperature The square of the difference, then weighted by a factor The summation is performed after adjusting the weights. When the deviation between the predicted temperature and the target temperature is small, this value is extremely small and has little impact on the total loss, indicating that the temperature tracking effect is excellent. When the deviation is large, this value increases significantly, generating a strong penalty to guide the optimization direction.

[0017] In some implementations, the valve action penalty in step S2 is, , The valve action penalty term is calculated in the control time domain. Incremental opening of the three types of valves in each step: gas, air, and ignition furnace. , , The squared values, respectively weighted by the corresponding coefficients , , The summation is performed after adjusting the weights. When the valve opening increment is small, this value is small and the penalty effect is weak. When the opening increment is large, this value increases sharply, resulting in a strong penalty to avoid excessive valve wear and process fluctuations.

[0018] In some implementations, step S3 specifically involves: S31, in one cycle Read the current main gas pipe pressure air pressure The three valve opening vectors at the previous moment ; S32. Using the PINN state transition function to perform forward deduction of the future. The temperature trajectory of the step is analyzed, and the future temperature trajectory is predicted using the particle swarm optimization (PSO) algorithm. By jointly optimizing the three valve increment sequences of the step, the globally optimal particle sequence is obtained, which represents the future state under the current condition. The optimal adjustment strategy for the three valves in the step; S33. In actual execution control, only future operations are executed. The first step in the process controls the increment and updates the future through feedback correction. The second step involves controlling the incremental control to optimize the valve opening rolling.

[0019] In some implementations, step S32 specifically involves: Substituting the predicted temperature and temperature tracking error terms, valve action penalty, and boundary penalty term into the MPC objective function as the particle fitness, and optimizing through PSO iterative search, the globally optimal particle sequence is obtained after convergence. , , ; Among them, the boundary penalty items include that the absolute value of the single-step valve opening change does not exceed 10% and the actual valve opening must be in the range of 0% to 100%.

[0020] Beneficial effects Compared with the prior art, the beneficial effects of the present invention are as follows: (1) The present invention provides a multi-valve joint control method for sintering ignition based on PINN-PMPC. For the clear causal chain of "gas / air pressure → flow rate → furnace pressure → ignition furnace flow rate → ignition temperature" during sintering ignition, a four-stage serial hierarchical PINN architecture is designed, with each layer corresponding to a physical subprocess. Unlike general black-box PINN or traditional data-driven models, the hierarchical structure of the present invention explicitly embeds the process mechanism. By introducing monotonicity constraints and temperature smoothness constraints in the loss function, the model output strictly conforms to industrial physical laws, significantly improving the model's interpretability, generalization ability, and prediction accuracy, thereby enabling precise control of the ignition temperature.

[0021] (2) The present invention provides a multi-valve joint control method for sintering ignition based on PINN-PMPC, which applies model predictive control (MPC) to the sintering ignition system and uses layered PINN as the steady-state equation; at the same time, it optimizes the opening sequence of three strongly coupled actuators: the main gas valve, the air valve, and the ignition furnace valve. This breaks through the limitations of traditional single-loop PID or rule-based switching control, and achieves precise control of ignition temperature by rolling the solution of the optimal control increment in the future time domain in each control cycle through the MPC framework.

[0022] (3) The present invention provides a multi-valve joint control method for sintering ignition based on PINN-PMPC, which uses particle swarm optimization (PSO) as the underlying solver for MPC. Under the premise of meeting practical engineering safety constraints such as valve opening range (0–100%) and single-step change limit (<10%), it efficiently solves the nonlinear, multivariable MPC optimization problem. It solves the problem of real-time performance and feasibility of complex nonlinear MPC in industrial field, avoids the failure risk of gradient-dependent optimization methods under non-smooth objective functions, ensures the optimality of online control commands, and combines theoretical advancement with engineering applicability. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of a multi-valve joint control method for sintering ignition based on PINN-PMPC. Figure 2 This is a diagram showing the relationship between sintering ignition parameters in this invention; Figure 3 The results are the temperature simulation results after PINN-PMPC control in this invention. Detailed Implementation

[0024] To further understand the content of this invention, a detailed description of the invention will be provided in conjunction with the accompanying drawings.

[0025] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0026] As mentioned in the background section, the multi-valve joint control of existing sintering ignition systems faces numerous technical challenges. From a control mechanism perspective, this multi-layered coupling relationship means that adjusting a single valve can trigger a chain reaction, affecting the stability of the entire system. For example, adjusting the main gas pipe valve not only affects the gas flow rate but also influences the ignition furnace pressure by changing the mixing ratio, thereby affecting the ignition furnace flow rate and temperature. Similarly, adjusting the air valve will affect the final temperature through a similar path.

[0027] Therefore, to achieve temperature stability at the specific target, coordinated control of the three valves is necessary. Furthermore, the system faces challenges such as high real-time requirements, frequent disturbances, and stringent control precision requirements. Sintering production is a continuous process; any control delay can cause the temperature to deviate from the setpoint, affecting product quality. Simultaneously, system disturbances are random and time-varying, making traditional static control strategies ill-suited to such dynamic changes.

[0028] To address the limitations of existing control methods, this invention proposes a multi-valve joint control method based on the combination of Physical Information Neural Network (PINN) and Model Predictive Control (MPC).

[0029] Specifically, refer to Figure 1 , Figure 2As shown, the multi-valve joint control method for sintering ignition based on PINN-PMPC of the present invention includes the following steps: Step 1: Constructing the Ignition Steady-State Equation (1) Data collection: During the sintering ignition process, the flow rate of the main gas pipe is determined by both the main gas pipe pressure and the opening degree of the main gas pipe valve, while the air flow rate is determined by both the opening degree of the air valve and the air pressure. After the gas and air are mixed, they enter the ignition furnace, and the pressure of the mixed gas is determined by the flow ratio of the gas to the air. The pressure of the ignition furnace and the opening degree of the ignition furnace valves jointly determine the flow rate of the ignition furnace, and the flow rate of the ignition furnace directly determines the temperature of the ignition furnace.

[0030] Therefore, it is necessary to collect flow and pressure information from multiple points. A PLC-based data acquisition device is installed at the locations of the equipment requiring data collection in the industrial environment to ensure that the data is collected quickly, accurately, and stably. The collected data includes... ( ), air pressure ( ), Ignition furnace pressure ( ), gas main flow rate ( ), airflow ( ), Ignition furnace flow rate ( ), Gas main pipe valve opening ( ), air valve opening ( ), Ignition furnace valve opening ( ) and real-time temperature detection ( ).

[0031] (2) Hierarchical PINN network design: To address the progressive physical logic of the sintering ignition process—"gas / air flow rate → ignition furnace pressure → ignition furnace flow rate → temperature"—a four-stage hierarchical PINN network was designed. Each physical step was decomposed into an independent sub-network, and by connecting these sub-modules, a complete prediction from input (pressure, valve) to output (temperature) was achieved. Each stage strictly corresponds to the actual causal relationship, ensuring the model's interpretability and physical consistency. Specifically, Phase 1: Gas / Air Flow Prediction Subnetwork Based on the main gas pipe pressure and the main gas pipe valve opening, the main gas pipe flow rate is predicted to independently reflect the flow characteristics of the gas system. Its steady-state relationship is shown in equation (1).

[0032] Simultaneously, based on air pressure and air valve opening, air flow rate is predicted to independently reflect the flow characteristics of the air system. Its steady-state relationship is shown in equation (2).

[0033] in, , These correspond to the flow-pressure-valve relationships of the gas and air systems, respectively.

[0034] Phase 2: Ignition Furnace Pressure Prediction Subnetwork Based on the mixing effect of gas and air flow, the gas pressure inside the ignition furnace is predicted, and its steady-state relationship is shown in equation (3).

[0035] in, This corresponds to the relationship between the mixing flow rate and the ignition furnace pressure.

[0036] Phase 3: Ignition Furnace Flow Prediction Subnetwork Based on the ignition furnace pressure and the ignition furnace valve opening, the output flow rate of the ignition furnace is predicted, and its steady-state relationship is shown in equation (4).

[0037] in, This corresponds to the relationship between the ignition furnace pressure, valve opening, and flow rate.

[0038] Phase 4: Temperature Prediction Subnetwork Based on the ignition furnace flow rate, the final electric furnace temperature is predicted, and its steady-state relationship is shown in equation (5).

[0039] in, This corresponds to the relationship between the ignition furnace flow rate and the predicted ignition furnace temperature.

[0040] Traditional data-driven models rely solely on the statistical regularities of historical data for modeling, making them susceptible to noise and anomalies in industrial data. This can lead to the learning of erroneous correlations that contradict actual physical logic, severely impacting the model's final prediction and control accuracy. Therefore, this invention, based on standard data fitting models, transforms the mechanistic changes in electric furnace operation into physical constraints and embeds these constraints into the PINN model. The PINN model is forced to learn the causal relationships during the ignition process, effectively avoiding physical biases in model predictions. Specifically, the physical constraints are transformed into a loss function, and this physical loss is embedded into the model to guide optimization in a direction consistent with physical laws.

[0041] This invention introduces two types of physical constraints, as detailed below: First, during the data acquisition process in actual sintering production, abnormal samples that violate the actual process logic, such as "gas pressure increases but flow rate decreases," are easily generated due to factors such as sensor malfunction and signal transmission delay. Therefore, this invention introduces five monotonicity constraints in the actual sintering ignition process to constrain the relationship between pressure, valve opening, and flow rate during sintering, as follows: ①Gas flow constraint: (The higher the gas pressure, the greater the flow rate.) (The larger the opening of the gas valve, the greater the flow rate); ② Airflow constraints: (The greater the air pressure, the greater the flow rate.) (The larger the air valve opening, the greater the flow rate); ③ Ignition furnace pressure constraint: (The higher the gas flow rate, the higher the pressure inside the furnace). (The greater the airflow, the higher the pressure inside the furnace); ④ Ignition furnace flow constraint: (The greater the pressure inside the furnace, the greater the flow rate.) (The larger the opening of the ignition furnace valve, the greater the flow rate); ⑤ Temperature constraint: (The higher the flow rate of the ignition furnace, the higher the combustion temperature).

[0042] The above five physical constraints are transformed into the physical loss corresponding to the model, which is defined as shown in formula (6).

[0043] in, This is the set of physical losses corresponding to the five monotonicity constraints of the model. By calculating the degree to which each sample violates the positive correlation constraint (i.e., the value of the partial derivative being negative), the ReLU function is used to penalize violations of positive correlation, and the squared values ​​are then summed and averaged. Its core function is to reduce the weight of samples that violate the positive correlation constraint: a non-zero penalty term is generated when the constraint is violated, and the more severe the violation, the larger the penalty term and the lower the sample weight; when the constraint is met, there is no penalty term, and the weight is unaffected. This guides the model to optimize in a direction consistent with physical causal logic, ensuring that the parameter correlation is consistent with actual process laws.

[0044] Secondly, during the sintering ignition process, the valve opening value of the gas will change instantaneously, but the core of the temperature change is the transfer and accumulation of heat, and heat transfer is a continuous physical process.

[0045] To ensure that the model predicts the cumulative change in temperature, this invention introduces a gradual temperature constraint to guarantee the cumulative change rather than a step change in temperature, i.e., a gradual temperature constraint. .

[0046] The above temperature asymptotic constraint is transformed into the physical loss corresponding to the model, which is defined as shown in formula (7).

[0047] in, The physical loss corresponding to the model's asymptotic temperature constraint is calculated by determining the temperature difference between adjacent time steps. The square of the summation over all adjacent time-time samples is averaged to quantify the degree to which the model's predicted temperature deviates from the gradual change pattern. Its core function is to penalize and correct the predicted results of abrupt temperature changes: when the predicted temperature difference between adjacent time-times is small, the square value of the temperature difference is extremely small, having a negligible impact on the total loss, indicating that the prediction results conform to the physical law of gradual temperature change.

[0048] When the predicted temperature difference between adjacent time points is large (resulting in a step-like abrupt change), the squared value of the temperature difference will be amplified sharply, generating a significant penalty term that guides the optimization process away from model parameters that would cause abrupt temperature changes. Ultimately, through this mechanism, the trajectory of the model's predicted temperature is ensured to be smooth and continuous, consistent with the actual physical logic of heat transfer and accumulation during sintering ignition.

[0049] The total loss function of PINN, which combines the physical loss mentioned above with the traditional data loss, is shown in equation (8).

[0050]

[0051]

[0052] in, For the total loss, It is the data fitting loss. For the first Predicted values ​​for each sample, No. Each sample's true value, using This measures the difference between the model-predicted temperature value and the actual measured temperature value. The total loss is a weighted sum of the data fitting loss and the physical constraint loss. , , The weighting coefficients for data loss, monotonicity loss, and temperature gradient loss, respectively. The adjustment logic for each of the above weighting coefficients needs to be set specifically according to the actual problems encountered during model training: for example, when the overall deviation of the data fitting is large, the weighting coefficients can be increased. Parameter weights are used to enhance the model's data fitting ability, making the model more focused on reducing the deviation between predicted and actual temperature values.

[0053] When the temperature changes in a large step, the increase should be increased. Parameter weights, by amplifying the penalty for losses caused by temperature abrupt changes, force the model to optimize parameters to output a smooth and continuous temperature trajectory.

[0054] If the model still produces predictions that violate the monotonicity constraint (such as an erroneous correlation where gas pressure increases but flow rate decreases), then the model needs to be strengthened. Parameter weights strengthen the constraint on the positive correlation between variables, ensuring that the model learns causal relationships that conform to physical logic.

[0055] The four-stage prediction subnetworks are connected in series according to physical logic, and the physical loss designed above is added to the network to form a complete temperature prediction chain. That is, an ignition temperature prediction model based on PINN network.

[0056]

[0057] Where T represents the predicted temperature of the ignition furnace.

[0058] (3) Design of state transition function: The sintering ignition process is a typical complex nonlinear system. Traditional formulaic state transition equations derived from mechanisms are difficult to accurately characterize the nonlinear relationships between variables and the complex disturbances in industrial scenarios, resulting in insufficient model description accuracy. Therefore, this invention directly uses the PINN-based ignition temperature prediction model as the steady-state equation of the system to establish the system's state transition equation, as shown in equation (11).

[0059]

[0060] Where k represents the current time, and k+j represents j future times. This represents the predicted temperature of the furnace at the current time k, and the predicted temperature at the next time k+j. The core function of this state transition function is to replace the equations derived from traditional mechanisms, and to accurately capture the complex dynamic characteristics of the sintering ignition process by leveraging the nonlinear fitting capability of the PINN model, thus achieving a precise mapping from the current system state to the future temperature state.

[0061] Step 2: Design of PMPC Optimization Objective Function After establishing the steady-state equations of the system, the next step is to design the system's optimization objective function. In designing the objective function for ignition furnace temperature control, attention needs to be paid to: 1. Temperature deviation; 2. Valve adjustment range.

[0062] The objective function designed in this invention takes into account both temperature tracking accuracy and valve action smoothness. Specifically, a temperature tracking error term is used to penalize the deviation between the predicted temperature and the set value to optimize the valve opening and control the ignition furnace temperature; a valve action penalty term is used to penalize drastic changes in valve opening to avoid valve wear or process fluctuations. This is specifically shown in equation (12).

[0063] in, The total loss value of the objective function for optimizing furnace temperature control is a weighted sum of the temperature tracking error term and the valve action penalty term. This loss value measures the quality of the control strategy by separately quantifying the degree of temperature tracking deviation and the degree of drastic change in valve opening and then weighting them together.

[0064] Its core function is to collaboratively achieve precise temperature tracking and smooth valve regulation: the temperature tracking error term is predicted in the time domain through calculation. Internal step-by-step temperature prediction With target temperature The square of the difference, then weighted by a factor Summing is performed after adjusting the weights.

[0065] When the deviation between the predicted and target temperatures is small, this term is extremely small and has negligible impact on the total loss, indicating excellent temperature tracking performance. When the deviation is large, this term increases significantly, generating a strong penalty to guide the optimization direction. The valve action penalty term is calculated in the control time domain. Incremental opening of the three types of valves in each step: gas, air, and ignition furnace. , , The squared values, respectively weighted by the corresponding coefficients , , Summing is performed after adjusting the weights.

[0066] When the valve opening increment is small (smooth operation), this value is small, and the penalty effect is weak. When the opening increment is large (drastic change), this value increases sharply, generating a strong penalty to prevent excessive valve wear and process fluctuations. Ultimately, through the synergistic effect of the two penalties, the total loss is reduced. Minimize controls to ensure optimal performance. This achieves precise tracking of the ignition furnace temperature to the target value while ensuring smooth and stable valve operation, meeting the actual operational requirements of the sintering production process.

[0067] Step 3: PMPC Rolling Optimization Solution After determining the steady-state change process of the system and the optimization objective function, the next step is to perform a rolling optimization design of the system, the purpose of which is to solve for the optimal valve opening through the rolling optimization algorithm.

[0068] Therefore, this invention introduces a rolling optimization solution scheme based on the particle swarm optimization (PSO) algorithm. Specifically, In a cycle The system first reads the current main gas pipe pressure. air pressure The three valve opening vectors at the previous moment ; Subsequently, the future is derived by forward propagation using the PINN state transition function. The temperature trajectory of each step is calculated, and the predicted temperature, temperature tracking error term, valve action penalty, and boundary penalty term are substituted into the MPC objective function as the particle fitness. Optimization is achieved through PSO iterative search, and the globally optimal particle sequence is obtained after convergence. , That is, to represent the future in the current state. The optimal adjustment strategy for the three valves in the step; Among them, the boundary penalty item includes and That is, the valve changes by less than 10% each time, and the valve's change boundary is 0-100%.

[0069] After solving the optimization problem, the future is obtained. The optimal adjustment sequence for the three valves. , This represents the control execution strategy at the current moment. However, in actual control execution, only the future strategy is executed. The first step in the process is incremental control, which means only executing the control quantity at the current moment and updating the next cycle (future) through feedback correction. The second step in the process is to control the increment.

[0070] In summary, the PINN-PMPC-based multi-valve joint control method for sintering ignition of this invention combines physical constraints with data-driven approaches. It directly embeds the system's physical constraints (such as the monotonic relationship between valve opening and flow rate, the positive correlation between pressure and flow rate, and the positive correlation between flow rate and temperature) into the neural network. This ensures the physical rationality of the model while fully utilizing the implicit information in historical data. This method overcomes the shortcomings of traditional neural networks in lacking physical constraints and solves the problem of traditional MPC relying on precise mathematical models.

[0071] Furthermore, by combining predictive control with optimal decision-making, the MPC controller utilizes a high-precision predictive model established by the PINN model to find the optimal sequence of control parameters through a multi-objective optimization algorithm. This method can not only predict the future state of the system but also consider control constraints and performance indicators, achieving coordinated control of multiple valves.

[0072] In summary, the multi-valve joint control method based on the PINN-MPC model of this invention can effectively solve problems such as multivariable coupling, nonlinearity, time-varying nature, and disturbances in sintering ignition control, providing a new technical path for achieving stable, efficient, and intelligent sintering ignition control. This method not only has significant theoretical value but also remarkable engineering application prospects, and is of great importance for improving the automation level of sintering processes and product quality.

[0073] The present invention will be further described below with reference to specific embodiments.

[0074] First, 20,218 data points were collected from the field using a PLC. The first 20,118 data points were used to train the PINN model to establish the steady-state equations, and the last 100 data points were used for prediction and control simulation. The specific characteristics of the data are shown in the table below. Table 1 Feature Table

[0075] Subsequently, the above data relationships are used to establish temperature prediction subnetwork, ignition furnace flow prediction subnetwork, ignition furnace pressure prediction subnetwork, and gas / air flow prediction subnetwork in sequence. Then, the networks of the four different stages are merged to construct a multi-stage ignition furnace prediction model based on PINN, and the steady-state equation of the ignition furnace is constructed.

[0076] To verify the advantages of the PINN-based model in prediction, it was compared with other models, and the metrics are shown in Table 2 below.

[0077] Table 2. Performance metrics of different models in ignition prediction

[0078] The experimental results in Table 2 above show that the PINN model with added physical information has the best performance, with a MAE of only 14.32, which is 49.67% lower than that of random forest and 67.27% lower than that of polynomial fitting. The RMSE is 19.73, which is 50.15% lower than that of random forest. The R² is as high as 0.85, and the fitting effect is far superior to other models.

[0079] The above analysis reveals the limitations of traditional multinomial fitting and support vector machines in modeling multi-parameter coupling during ignition. While random forests and the basic PINN model can learn the nonlinear laws directly from the data, they lack physical constraints and are prone to predictive biases that defy common sense. In contrast, the PINN model, incorporating physical information, matches actual physical processes through sub-networks and embeds monotonicity and temperature smoothness constraints, ensuring both fitting accuracy and physical plausibility.

[0080] In summary, the comparative experiments fully demonstrate that the PINN model, based on hierarchical design and the integration of physical constraints, can accurately capture the multi-parameter coupling relationship and physical laws of the sintering ignition process. Its predictive performance far exceeds that of traditional models and the basic PINN model, providing high-precision and high-reliability steady-state equation support for the subsequent PINN-PMPC multi-valve joint control, and laying the core foundation for achieving stable closed-loop control of ignition temperature.

[0081] After establishing the steady-state equations, the process moves to the PMPC optimization objective function construction and parameter design stage. Specifically, in this embodiment... Meanwhile, the core parameters are determined as follows: Prediction Time Domain Step (corresponding to a 6-second sampling cycle in industrial systems, i.e., predicting the temperature trajectory for the next 60 seconds). Control Time Domain Step (rolling optimization of valve opening in the next 4 steps), balancing control response speed and computational cost; Temperature tracking error term weighting coefficient Prioritize temperature control; Penalty coefficient for the increase in opening degree of main gas pipe valve, air valve, and ignition furnace valve , , All values ​​are set to 4 to strengthen smoothing constraints on the strong influence of ignition furnace valves on flow rate and temperature. Simultaneously, valve operation constraints are clearly defined: the opening range is strictly limited to... Single adjustment increment To avoid mechanical wear and drastic fluctuations in the process.

[0082] Then, a rolling optimization solution based on the particle swarm optimization (PSO) algorithm was initiated, with a population size of 40, 80 iterations, and inertia weights. Learning factor , It balances global search efficiency with local convergence efficiency.

[0083] In the control simulation with 100 test data points, the future temperature trajectory is deduced through the PINN state transition function. The temperature deviation and valve action are substituted into the objective function to calculate the particle fitness. The optimal valve increment sequence is obtained through PSO iteration. Only the current step control quantity is executed and feedback correction is performed.

[0084] The final control result is as follows Figure 3 As shown: Figure 3In (a), the gray curve represents the original temperature change when the valve is not adjusted. Its fluctuations are drastic and show a clear, continuous downward trend, with a maximum deviation of over 150°C from the set value of 950°C, failing to meet the temperature stability requirements of industrial production. The red curve represents the temperature change after MPC control. It can be seen that after MPC control, the ignition furnace temperature consistently fluctuates slightly around the target value of 950°C, with the steady-state deviation strictly controlled within 50°C. The temperature fluctuation amplitude is significantly narrowed, effectively suppressing the temperature fluctuations in the original state.

[0085] Figure 3 In the figure, (b), (c), and (d) correspond to the dynamic changes in the adjustment amount of the main gas pipe valve, the opening degree of the air valve, and the opening degree of the ignition furnace valve, respectively. All three curves are adjusted slightly at the original valve opening, constraining the incremental valve action. This ensures temperature stability while avoiding mechanical wear and process fluctuations. It significantly improves temperature control accuracy, ultimately achieving precise and stable control of the ignition furnace temperature.

[0086] The present invention and its embodiments have been described above illustratively. This description is not restrictive, and the figures shown are only one embodiment of the present invention; the actual structure is not limited thereto. Therefore, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the present invention, such designs should fall within the protection scope of the present invention.

Claims

1. A PINN-PMPC-based sintering ignition multi-valve joint regulation method, characterized in that: Stable temperature control of the ignition furnace is achieved by adjusting the opening of the ignition furnace valve, the air valve, and the main gas pipe valve. The steps include: S1, constructing the ignition steady-state equation. S11, Data Acquisition Key parameters during the sintering ignition process are collected by PLC, including ( ), air pressure ( ), Ignition furnace pressure ( ), gas main flow rate ( ), airflow ( ), Ignition furnace flow rate ( ), Gas main pipe valve opening ( ), air valve opening ( ), Ignition furnace valve opening ( ) and real-time temperature detection ( ); S12, Hierarchical PINN Network Design Based on the progressive physical logic of the sintering ignition process, namely "gas / air flow rate → ignition furnace pressure → ignition furnace flow rate → temperature", prediction sub-networks for the corresponding stages are constructed. The four-stage prediction sub-networks are connected in series according to the physical logic, and physical constraints are embedded to construct the total loss function, thus obtaining the PINN temperature prediction model. S13. Using the PINN temperature prediction model as the system steady-state equation, establish a state transition function that describes the relationship between the current state and the future temperature. S2, PMPC Optimization Objective Function Design A multi-objective optimization objective function is constructed, which includes a temperature tracking error term and a valve action penalty term. The temperature tracking error term is used to penalize the deviation between the predicted temperature and the set value, and the valve action penalty term is used to penalize drastic changes in valve opening. S3, PMPC rolling optimization solution Based on the state transition function and the optimization objective function, the optimal valve opening is solved using a rolling optimization algorithm.

2. The method for joint control of sintering ignition using multiple valves based on PINN-PMPC according to claim 1, characterized in that: In step S12, constructing the four-stage prediction sub-network specifically involves: Phase 1: Constructing a subnetwork for predicting gas / air flow rates Based on the main gas pipe pressure and the main gas pipe valve opening, the main gas pipe flow rate is predicted to independently reflect the flow characteristics of the gas system. Its steady-state relationship is as follows: ; Simultaneously, based on air pressure and air valve opening, air flow is predicted to independently reflect the flow characteristics of the air system. Its steady-state relationship is: ; Phase 2: Constructing the Ignition Furnace Pressure Prediction Subnetwork Based on the mixing effect of gas and air flow, it predicts the gas pressure inside the ignition furnace, and its steady-state relationship is as follows: ; Phase 3: Constructing the Ignition Furnace Flow Prediction Subnetwork Based on the ignition furnace pressure and the ignition furnace valve opening, the output flow rate of the ignition furnace is predicted, and its steady-state relationship is as follows: ； Phase 4: Constructing the temperature prediction subnetwork Based on the ignition furnace flow rate, the final electric furnace temperature is predicted, and its steady-state relationship is as follows: .

3. The method for joint control of sintering ignition using multiple valves based on PINN-PMPC according to claim 2, characterized in that: The physical constraints in step S12 include the following monotonicity constraints. Gas flow constraint: That is, the greater the gas pressure, the greater the flow rate; That is, the larger the opening of the gas valve, the greater the flow rate; Airflow constraints: In other words, the greater the air pressure, the greater the flow rate. In other words, the larger the air valve opening, the greater the flow rate. Ignition furnace pressure constraints: That is, the greater the gas flow rate, the higher the pressure inside the furnace; That is, the greater the airflow, the higher the pressure inside the furnace; Ignition furnace flow constraint: That is, the greater the pressure inside the furnace, the greater the flow rate; That is, the larger the opening of the ignition furnace valve, the greater the flow rate; Temperature constraints: That is, the higher the flow rate of the ignition furnace, the higher the combustion temperature; The five monotonicity constraints are transformed into the corresponding set of monotonic losses for the model as follows: 。 4. The method for joint control of sintering ignition using multiple valves based on PINN-PMPC according to claim 3, characterized in that: The physical constraints in step S12 also include temperature asymptotic constraints: , The temperature asymptotic constraint is transformed into the temperature gradient loss corresponding to the model as follows: Adding the physical loss and data loss together, PINN's total loss function is: in, It's data loss. ; For the first Predicted values ​​for each sample, No. The true value of each sample; , , These are the weighting coefficients for data loss, monotonicity loss set, and temperature variability loss, respectively. The final ignition temperature prediction model is: .

5. The method for joint control of sintering ignition using multiple valves based on PINN-PMPC according to claim 4, characterized in that: The state transition function in step S13 is: Where k represents the current time, and k+j represents j future times. This represents the predicted temperature of the furnace at the current time k, and the predicted temperature at the future time k+j.

6. A multi-valve joint control method for sintering ignition based on PINN-PMPC according to any one of claims 1-5, characterized in that: The temperature tracking error term in step S2 is: ; The temperature tracking error term is calculated and predicted in the time domain. Internal step-by-step temperature prediction With target temperature The square of the difference, then weighted by a factor The summation is performed after adjusting the weights. When the deviation between the predicted temperature and the target temperature is small, this value is extremely small and has little impact on the total loss, indicating that the temperature tracking effect is excellent. When the deviation is large, this value increases significantly, generating a strong penalty to guide the optimization direction.

7. The method for joint control of sintering ignition using multiple valves based on PINN-PMPC according to claim 6, characterized in that: The valve action penalty item in step S2 is... ， The valve action penalty term is calculated in the control time domain. Incremental opening of the three types of valves in each step: gas, air, and ignition furnace. , , The squared values, respectively weighted by the corresponding coefficients , , Summation after adjusting weights; When the valve opening increment is small, this value is small and the penalty effect is weak; When the opening increment is large, this value increases sharply, resulting in a strong penalty to avoid excessive valve wear and process fluctuations.

8. The method for joint control of sintering ignition using multiple valves based on PINN-PMPC according to claim 7, characterized in that: Specifically, step S3 is as follows: S31, in one cycle Read the current main gas pipe pressure air pressure The three valve opening vectors at the previous moment ; S32. Using the PINN state transition function to perform forward deduction of the future. The temperature trajectory of the step is analyzed, and the future temperature trajectory is predicted using the particle swarm optimization (PSO) algorithm. By jointly optimizing the three-valve incremental sequence, the globally optimal particle sequence is obtained, which represents the future state under the current condition. The optimal adjustment strategy for the three valves in the step; S33. In actual execution control, only future operations are executed. The first step in the process controls the increment and updates the future through feedback correction. The second step involves controlling the incremental control to optimize the valve opening rolling.

9. The method for joint control of sintering ignition using multiple valves based on PINN-PMPC according to claim 8, characterized in that: Specifically, step S32 is as follows: Substituting the predicted temperature and temperature tracking error terms, valve action penalty, and boundary penalty term into the MPC objective function as the particle fitness, and optimizing through PSO iterative search, the globally optimal particle sequence is obtained after convergence. , , ; Among them, the boundary penalty items include that the absolute value of the single-step valve opening change does not exceed 10% and the actual valve opening must be in the range of 0% to 100%.

Images