Intelligent prediction and control method for energy consumption of glass melting

Abstract

The present application relates to the technical field of industrial automation control, and relates to an intelligent prediction and control method for glass melting energy consumption. The present application acquires device parameters and fluid mechanics priori knowledge, eliminates pseudo correlation to build a structure causal diagram reflecting the true physical heat transfer direction; performs causal intervention deduction on the parameters of the charging machine in the diagram, takes the minimum liquid surface thermal reflectivity as the optimization target, solves the target material blanket shape parameters, and accordingly reshapes the physical material blanket space layout; identifies the causal effect mutation between the temperature and the burner, eliminates the invalid mapping, and dynamically generates the combustion control sub-area topology layout matched with the current material blanket; calculates the required synthetic heat energy and the top flame parameters based on this, and then outputs the local following type bubbling instruction. The present application breaks the static boundary of the traditional measurement and control architecture, realizes the material blanket form optimization and the three-dimensional physical cooperation of the upper and lower heat sources, and effectively avoids the invalid energy consumption caused by the hotspot drift and global blind bubbling from the source.

Description

Technical Field

[0001] This invention relates to the field of industrial automation control technology, and specifically to an intelligent prediction and control method for energy consumption in glass melting. Background Technology

[0002] Glass melting is the core production process that melts various raw materials in a high-temperature furnace to form a homogeneous molten glass. This process involves complex heat transfer, mass transfer, and chemical reactions. It is a typical energy-intensive industrial process, with its energy consumption accounting for a very high proportion of the total energy consumption of the glass melting process. Therefore, implementing precise automated control of the glass melting process to maintain stable glass quality, reduce energy consumption per unit product, and improve energy utilization efficiency is crucial for optimizing the melting process.

[0003] Existing glass melting control technologies largely rely on experience-based settings or simple feedback control loops. Operators manually adjust the system based on visual information such as flame color and melting conditions, or employ proportional-integral-derivative (PI-DE) control strategies that pair the temperature at a specific measuring point with the burner valve opening. Some advanced systems also utilize predictive models based on statistical regression, analyzing historical data to establish mathematical relationships between input and output variables, thereby guiding control operations. These methods have, to some extent, automated the production process, replacing some of the heavy manual operations.

[0004] However, these existing technologies have inherent limitations when dealing with complex and ever-changing smelting conditions. Fixed-loop control strategies cannot adapt to hot spot drift within the furnace caused by batch variations in raw materials or disturbances in the feeding process, rendering the pre-set measurement and control mapping relationship ineffective and leading to decreased control accuracy. Furthermore, control models based on statistical correlation struggle to distinguish between genuine causal relationships and spurious synergistic relationships between variables, potentially leading to decisions based on incorrect associations and impacting control effectiveness. In addition, top combustion heating and bottom bubbling agitation are typically controlled as two independent systems, lacking an effective coordination mechanism. This prevents energy from being precisely applied to the areas most requiring enhanced melting, resulting in energy waste. Summary of the Invention

[0005] In view of this, in order to solve the problems mentioned in the background technology, an intelligent prediction and control method for glass melting energy consumption is proposed.

[0006] The objective of this invention can be achieved through the following technical solution: This invention provides an intelligent prediction and control method for energy consumption in glass melting, comprising: S1, acquiring discretized equipment operating parameters and prior knowledge of fluid mechanics in the physical space of glass melting, extracting spatial heat transfer variables, generating an initial state dataset, performing causal discovery processing on the initial state dataset, calculating the partial correlation coefficients between spatial heat transfer variables through conditional independence tests, thereby identifying and eliminating spurious correlations between spatial heat transfer variables, and constructing a structural causal graph that reflects the true physical heat transfer direction.

[0007] S2. Perform causal intervention on the operating parameters of the feeder in the structural cause-effect diagram, deduce the liquid surface thermal reflectivity under different material blanket propulsion trajectories, and solve for the target material blanket shape parameters with the minimum liquid surface thermal reflectivity as the optimization objective.

[0008] S3. Generate feeder control commands based on the target material blanket shape parameters, adjust the feeder's pushing stroke and vibration frequency, and reshape the spatial layout of the physical material blanket.

[0009] S4. Identify the abrupt change in the causal effect intensity between the glass melt temperature variable and the burner node variable in the structural causal diagram, eliminate failed measurement and control mapping relationships, and generate a combustion control sub-region topology layout that matches the current physical blanket spatial layout.

[0010] S5. Based on the topology layout of the combustion control sub-region, calculate the synthesis thermal energy parameters required to maintain the target viscosity of the molten glass in the current region, and generate the corresponding top flame extension length and direction parameters.

[0011] S6. Generate the start / stop topology diagram of the bottom bubbling valve array based on the extension length and direction parameters of the top flame, and output local follow-up bubbling control commands to realize intelligent control of glass melting energy consumption.

[0012] Compared with the prior art, the embodiments of the present invention have at least the following advantages or beneficial effects: (1) The present invention breaks through the technical barrier of the highly static and solidified bottom-level measurement and control infrastructure of traditional glass melting furnaces. By constructing a structural causal graph that reflects the real physical heat transfer direction, false temperature coordination data caused by flue gas flow is effectively eliminated. The system can identify the sudden change in the intensity of the causal effect between the glass melt temperature and the burner nodes in real time, automatically eliminate the failed measurement and control mapping relationship, and dynamically generate a combustion control sub-region topology that matches the current material blanket layout. In this process, without modifying any physical pipeline hardware, the digital infrastructure can be upgraded to generative autonomy, effectively solving the control lag and energy waste caused by local hot spot drift under variable load.

[0013] (2) This invention changes the passive situation in traditional control where "fluctuations in the physical shape of the material blanket are treated as uncontrollable disturbances leading to global temperature rise". By performing counterfactual causal intervention deduction on the operating parameters of the feeder in the structural causal model, the target material blanket shape parameters are solved with the minimum thermal reflectivity of the liquid surface as the optimization objective. By precisely adjusting the feeding stroke and vibration frequency, the spatial layout of the physical material blanket is actively reshaped, improving energy transfer efficiency from the source, maximizing the absorption of heat by the glass liquid surface, and significantly reducing the basic energy consumption of the system.

[0014] (3) This invention constructs a tight physical logic closed loop of input control, surface energy supply, and deep heat transfer. By calculating the distribution of thermal radiation flux, regions with radiation flux greater than a set threshold are extracted, and combined with regions with a stack thickness greater than a set thickness threshold, the system automatically generates the start / stop topology map and gas volume command for the bottom bubble array. This transforms the traditional "global blind bubbling" with fixed output year-round into precise bubbling that "locally follows the flame hotspots". This physical spatial coordination of upper and lower heat sources precisely guides heat transfer to the deep glass melt, not only avoiding insufficient heat penetration of the deep glass melt and excessive erosion of local refractory materials, but also significantly reducing energy waste caused by ineffective bubbling and significantly reducing natural gas consumption.

[0015] (4) This invention differs from the purely data-driven black-box models in the prior art that lack physical mechanism constraints. The control topology of this invention is based on a structural causal model constrained by prior knowledge of fluid mechanics and thermodynamics. Every control loop reconstruction and every energy allocation of the system strictly follows the real physical heat transfer causal logic rather than just data correlation. This strong physical binding mechanism endows the model with high interpretability and robustness. Combined with the dynamic feedback parameter adaptive update mechanism, it ensures that the method has good feasibility and long-term effectiveness in complex actual industrial sites. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of the method steps of the present invention.

[0018] Figure 2 This is a bar chart comparing the abrupt changes in the intensity of causal effects according to the present invention.

[0019] Figure 3 This is the response curve of the synthesis thermal energy parameters of the present invention as a function of viscosity deviation. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please see Figure 1 This invention provides an intelligent prediction and control method for energy consumption in glass melting, comprising: S1, acquiring discretized equipment operating parameters and prior knowledge of fluid mechanics in the physical space of glass melting, extracting spatial heat transfer variables, generating an initial state dataset, performing causal discovery processing on the initial state dataset, calculating the partial correlation coefficients between spatial heat transfer variables through conditional independence tests, thereby identifying and eliminating spurious correlations between spatial heat transfer variables, and constructing a structural causal graph that reflects the true physical heat transfer direction.

[0022] In a specific embodiment of the present invention, discrete equipment operating parameters and prior knowledge of fluid mechanics in the physical space of glass melting are obtained, spatial heat transfer variables are extracted, an initial state dataset is generated, causal discovery processing is performed on the initial state dataset, and partial correlation coefficients between spatial heat transfer variables are calculated through conditional independence tests to identify and eliminate spurious correlations between spatial heat transfer variables, and a structural causal graph reflecting the true physical heat transfer direction is constructed. This includes: obtaining mechanical action parameters of the feeder, valve opening parameters of the burner, gas volume parameters of the bottom bubbling array, and spatial discrete temperature measurement point parameters in the physical space of glass melting, and combining them with prior knowledge of fluid mechanics to generate an initial state dataset.

[0023] Conditional independence tests were performed on the initial state dataset to identify the correlation matrix among spatial heat transfer variables.

[0024] Based on the correlation matrix, false temperature correlation data caused by flue gas flow are eliminated, and a structural causal graph reflecting the true physical heat transfer direction is constructed.

[0025] Specifically, this step aims to construct a model that accurately describes the causal relationship of heat transfer within the melting space by analyzing operational data during the glass melting process and combining it with known physical laws. The process of building this model begins with data acquisition.

[0026] The first step is to acquire the data needed to form the initial state dataset. This data falls into two categories: operational parameters directly collected from the equipment within the physical space of glass melting, and prior fluid dynamics knowledge serving as physical constraints. The equipment operational parameters specifically include four aspects: feeder mechanical action parameters, describing the force and frequency with which the feeder pushes the raw material into the furnace; burner valve opening parameters, determining the supply of fuel and combustion air, directly affecting the flame's energy output; bottom bubbling array gas flow parameters, controlling the flow rate of gas blown in from the bottom of the furnace to agitate the molten glass and enhance heat transfer; and spatially discrete temperature measurement point parameters, i.e., real-time temperature readings measured by thermocouples installed at different locations within the furnace. Prior fluid dynamics knowledge refers to universally accepted physical heat transfer formulas such as the first law of thermodynamics, Fourier's law of heat conduction, and Newton's law of cooling. In the system algorithm, this prior knowledge is pre-encoded as a directed edge constraint rule base, for example, specifying that the node direction of the heat transfer path can only point from a high-temperature node to a low-temperature node. This knowledge is used in subsequent steps to determine the plausibility of the causal relationship between variables, such as the possibility that heat can only be transferred from a high-temperature region to a low-temperature region. Integrating these continuously collected equipment operating parameters with spatially discrete temperature measurement point parameters constitutes the initial state dataset. This dataset is a multi-dimensional time series matrix that records the system's state at different times.

[0027] After obtaining the initial state dataset, it is necessary to identify the inherent relationships between spatial heat transfer variables. Spatial heat transfer variables refer to variables in the dataset that directly or indirectly affect heat distribution, mainly including burner valve opening parameters, gas flow parameters of the bottom bubbling array, and spatial discrete temperature measurement point parameters. To quantify the strength of the relationships between these variables, conditional independence tests need to be performed on the initial state dataset. A conditional independence test is a statistical method used to determine whether two variables remain statistically independent after excluding the influence of one or more other variables. For example, testing variables... and variables In a given set of variables Independence under the condition of is denoted as In practice, this is typically achieved by calculating the partial correlation coefficient. The partial correlation coefficient measures the strength of the linear relationship between two variables after controlling for other variables. and In control variables Partial correlation coefficient after It can be calculated using the following formula: In the formula, , and Representing variables respectively and , and , and The Pearson correlation coefficient between the variables is calculated. A correlation matrix can be generated by performing pairwise tests on all pairs of spatial heat transfer variables in the initial state dataset and progressively increasing the dimension of the condition variable set. The element values ​​of this matrix reflect the strength of the conditional dependence between the variable pairs; values ​​close to 0 indicate that the two variables are independent given other variables.

[0028] The final step is to use the correlation matrix obtained in the previous step to construct a structural causal graph reflecting the true physical heat transfer direction. The high-temperature flue gas in the glass furnace flows throughout the space, simultaneously heating multiple temperature measuring points along its path. This causes the readings of these points to show a strong correlation, which is spurious temperature correlation data caused by the flue gas flow. Without proper screening, this could be incorrectly interpreted as a direct heat transfer between these temperature measuring points. Eliminating this spurious correlation is crucial for building an accurate model. Performing a conditional independence test can effectively identify such cases. For example, if two temperature measuring points... and Temperature readings are highly correlated, but when we consider the burner parameters that affect flue gas temperature... After being used as a condition variable, it was found that and The partial correlation coefficient is close to 0, that is... Then it can be concluded that and The correlation between them is due to a common cause The spurious correlations are driven by factors that lack a direct causal path. By systematically applying this principle, all spurious connections caused by common causes are eliminated. Then, by combining this with a directed edge direction constraint rule base corresponding to prior knowledge of fluid mechanics, the causal direction is determined. For example, increasing the burner valve opening is the cause, and increasing the temperature at the measuring point is the effect. Finally, a directed acyclic graph is obtained. This graph is the structural causal graph, where the nodes represent spatial heat transfer variables, and the directed edges represent the actual physical heat transfer directions and causal relationships between the variables.

[0029] For example, assuming a simplified glass melting physical space, we focus on three spatial heat transfer variables: the burner valve opening parameter located near burner A. Parameters of spatially discrete temperature measurement points located on the left side of the molten pool And the spatial discrete temperature measurement point parameters located on the right side of the molten pool. The flue gas produced by burner A will simultaneously flow through and Location.

[0030] First, we collected data for these three variables over 10 consecutive time units, with the collection frequency set to 1 minute based on the industrial sensor response time, forming an initial state dataset, as shown in Table 1 below: Table 1 Example of initial state dataset .

[0031] Next, we perform a conditional independence test on the initial state dataset. We first calculate the Pearson correlation coefficient between the variables. and correlation coefficient The calculation result is 0.99962. and correlation coefficient The calculation result is 0.9998. and correlation coefficient The calculated result is 0.9998. These highly correlated values ​​initially constitute the correlation matrix.

[0032] Then, spurious temperature correlation data caused by flue gas flow needs to be removed. We suspect... and The high correlation between them is due to their shared flue gas heat source. This is caused by. Therefore, we calculate in the case of After being used as a condition variable and Partial correlation coefficient between .

[0033] Substitute into the formula: .

[0034] To make the example clearer, the data here has been manipulated; the partial correlation coefficient calculated from actual industrial data would be closer to 0. Assuming more precise calculations or the use of a larger dataset, the value would be 0.05, below the preset independence threshold of 0.1. This indicates that once the burner valve opening parameter is considered... The impact, and The statistical correlation between them disappears. Therefore, we conclude that... and The strong correlation between them is a pseudo-correlation.

[0035] Finally, a structural cause-effect graph is constructed. Initially, since all variable pairs are highly correlated, a fully connected graph may exist. Based on the conditional independence test results above, we remove... and The direct connection between them. Furthermore, based on prior knowledge of fluid mechanics, the burner is the heat source, its state change is the cause, and the temperature change is the effect; therefore, the cause-effect direction is from... point to and The final constructed cause-effect graph is as follows: This diagram accurately reflects the actual physical direction of heat transfer, that is, burner A heats both the left and right regions simultaneously, but there is no direct heat transfer between the two regions.

[0036] S2. Perform causal intervention on the operating parameters of the feeder in the structural cause-effect diagram, deduce the liquid surface thermal reflectivity under different material blanket propulsion trajectories, and solve for the target material blanket shape parameters with the minimum liquid surface thermal reflectivity as the optimization objective.

[0037] In a specific embodiment of the present invention, a causal intervention operation is performed on the operating parameters of the feeder in the structural causal graph to infer the thermal reflectivity of the liquid surface under different material blanket propulsion trajectories. The goal is to minimize the thermal reflectivity of the liquid surface and solve for the shape parameters of the target material blanket. This includes setting the operating parameters of the feeder as intervention variables in the structural causal graph and constructing a counterfactual inference model.

[0038] The intervention variables are input into the counterfactual model to calculate the liquid surface thermal reflectivity corresponding to different material blanket propulsion trajectories.

[0039] With minimizing the thermal reflectivity of the liquid surface as the optimization objective, the counterfactual inference model is solved, and the target material blanket shape parameters are output.

[0040] Specifically, after constructing a structural causal graph reflecting the actual physical heat transfer direction in step S1, this step uses this graph to perform proactive and exploratory calculations to determine the optimal feeding strategy. The core of this process is to perform causal intervention operations on the feeding machine's operating parameters. This is not an actual adjustment of the physical equipment, but a simulation and deduction within the established causal model to predict the impact of different operations on the system's final thermal efficiency.

[0041] First, a mathematical model for deduction needs to be constructed. In the structural causal graph, the operating parameters of the feeding machine are set as intervention variables. These parameters are directly controllable engineering quantities, such as the feeding stroke and vibration frequency. Changes in these parameters directly cause changes in a series of subsequent physical quantities. Intervention variables are given special status in the causal model; their values ​​are forcibly set by us, rather than determined by their parent nodes in the graph, thus breaking the original data generation process and simulating the effect of human intervention. Based on this intervention idea, a counterfactual deduction model is constructed. This model is essentially a quantitative expression of the causal chain represented by the structural causal graph. It consists of a set of structural equations, each describing how a node variable is determined by its direct cause parent node.

[0042] For example, the operating parameters of the feeder determine the trajectory of the molten glass blanket, which in turn determines the surface thermal reflectivity of the molten glass. The molten glass blanket trajectory is a physical state describing the spread and speed of the newly added material on the molten surface, and can be characterized by parameters such as the average coverage length of the blanket. Surface thermal reflectivity is a dimensionless parameter that measures the proportion of heat reflected from the flame by the molten glass surface; the lower the value, the more heat absorbed by the surface. The counterfactual model mathematizes these relationships. Assume the feeder's pushing stroke is... The vibration frequency is These parameters collectively determine the key characteristic parameters of the blanket's propulsion trajectory, such as the average length of the blanket. This relationship can be represented as: In the formula, It is a function whose specific form is determined through regression analysis of the initial state dataset in step S1, specifically including multinomial regression fitting, support vector machine regression, or multilayer perceptron algorithm. This function captures the physical relationship between the feeding operation and the shape of the material blanket. Next, the average length of the material blanket... It will affect the thermal reflectivity of the liquid surface This causal relationship can be expressed as: In the formula, This is another function, also learned from the initial state dataset, which reflects the effect of the blanket covering on the heat absorption characteristics of the molten pool surface. These two equations together form the core of the counterfactual inference model.

[0043] Once the model is built, extrapolation calculations can be performed. This involves a series of intervention variables within the equipment's allowable range, namely different operating parameters of the feeder. The combinations are then fed one by one into the counterfactual inference model. For each set of inputs... The model will first calculate the corresponding blanket propulsion trajectory parameters. Then based on Calculate the final liquid surface thermal reflectivity By iterating through all reasonable parameter combinations, a mapping table of different material blanket propulsion trajectories and corresponding liquid surface thermal reflectivity is obtained.

[0044] Finally, it is necessary to find the operating parameters that minimize the thermal reflectivity of the liquid surface. This transforms into an optimization problem. (The question then shifts to the question of how to minimize the thermal reflectivity of the liquid surface.) Minimize the objective function of the counterfactual inference model to solve the problem. The constraint is the feeding stroke of the feeder. and vibration frequency The physical operating range. By solving this optimization problem, a set of optimal operating parameters for the feeder can be obtained. The parameters corresponding to this set of parameters are the material feeding trajectory parameters. This refers to the target blanket shape parameters. These target blanket shape parameters will be passed to the next step to generate specific equipment control commands.

[0045] For example, suppose that in step S1, by analyzing the initial state dataset generated based on 200 sets of industrial sensor measured data, we obtained a structural causal graph and constructed the following counterfactual inference model based on it.

[0046] First, set the operating parameters of the feeder, namely the pushing stroke. (Unit: mm) and vibration frequency (Unit: Hertz), set as the intervention variable. The constructed counterfactual inference model consists of the following two equations obtained through data regression:

[0047] The trajectory of the blanket advance is determined by its average length (Unit: meters) Characterized by its relationship with the intervention variable: In this formula, the coefficient 0.05 is in meters per millimeter (m / mm) and 0.1 is in meters per hertz (m / Hz), ensuring dimensional consistency. Liquid surface thermal reflectivity. (Dimensionless) and average length of blanket The relationship is: In this formula, the coefficient 0.02 is in units of... The unit of -0.2 is 0.8 is a dimensionless constant, which also ensures dimensional consistency.

[0048] Next, different intervention variables are input into the counterfactual inference model for calculation. Assume the feeding stroke of the feeder. The feasible range is [60, 120] mm, vibration frequency The feasible range is [5, 25] Hertz. We selected two sets of intervention variables for extrapolation: the first set of intervention variables: pushing distance. millimeters, vibration frequency Hertz. Calculate the corresponding blanket propulsion trajectory parameters: Meters. Calculate the corresponding liquid surface thermal reflectivity: The second intervention variable: pushing stroke. millimeters, vibration frequency Hertz. Calculate the corresponding blanket propulsion trajectory parameters: Meters. Calculate the corresponding liquid surface thermal reflectivity: Calculations revealed that the second set of intervention variables corresponded to a lower surface thermal reflectivity.

[0049] Finally, with minimizing the thermal reflectivity of the liquid surface as the optimization objective, the counterfactual inference model is solved. We need to solve the function. The minimum value of . This is a quadratic function that opens upwards, and its minimum value is obtained at the vertex. Differentiate and set it to 0: The solution is: rice.

[0050] Therefore, when the average length of the blanket is 5 meters, the liquid surface thermal reflectivity is lowest, and heat absorption is maximized. Thus, the target blanket shape parameter is 5 meters.

[0051] S3. Generate feeder control commands based on the target material blanket shape parameters, adjust the feeder's pushing stroke and vibration frequency, and reshape the spatial layout of the physical material blanket.

[0052] In a specific embodiment of the present invention, a feeding machine control command is generated based on the target material blanket shape parameters, the feeding machine's pushing stroke and vibration frequency are adjusted, and the spatial layout of the physical material blanket is reshaped. This includes: obtaining the target material blanket shape parameters, and calling the structural equation between the feeding machine operating parameters and the material blanket shape parameters in the counterfactual inference model to construct an inverse solution model.

[0053] A reference vibration frequency based on the equipment energy efficiency curve is introduced as a constraint. The constraint is substituted into the inverse solution model, and the structural equation is simplified into a single-variable equation about the pushing stroke.

[0054] Solve the single-variable equation to obtain the target pushing stroke corresponding to the shape parameters of the target blanket.

[0055] The target feeding stroke is combined with the reference vibration frequency and converted into a standard industrial communication protocol format to generate feeder control commands.

[0056] The control commands of the feeder are sent to the underlying programmable logic controller to drive the feeder's actuators to adjust mechanical actions and reshape the spatial layout of the physical material blanket.

[0057] Specifically, this step transforms the abstract optimization result calculated in step S2, i.e., the target blanket shape parameters, into specific operational instructions that can be directly executed by physical equipment. This is a reverse solution process, the core of which is to deduce the control input feeder operating parameters required to achieve the desired system state based on the optimal blanket shape.

[0058] The input to this process is the target blanket shape parameter that maximizes heat absorption, output from the previous step. This parameter is based on the average length of the blanket. For example, the ideal spatial layout of the physical blanket on the molten pool surface is described. To achieve this layout in the real world, the feeding stroke and vibration frequency of the feeder need to be precisely adjusted. These two controllable physical quantities are defined as intervention variables in step S2, and their functional relationship with the blanket shape is described. It has been established through counterfactual reasoning modeling.

[0059] Now, the task is based on what is known. Solve for the corresponding pusher stroke. and vibration frequency This solution process is based on the same set of structural equations established in step S2. Specifically, we need to solve the following equations: Since the equation contains two unknown variables and There are usually multiple solutions. To obtain a unique, executable control command, additional constraints need to be introduced. A common engineering practice is to prioritize combinations of operating parameters with lower energy consumption or less wear and tear on equipment, while ensuring that key performance indicators are met. For example, the vibration frequency can be fixed. A baseline value set based on equipment energy efficiency curves and long-term operating experience. Then solve for the unique pushing stroke. At this point, the equation simplifies to about The single-variable equation can be solved directly.

[0060] The calculated pusher stroke and the set vibration frequency Combined, these actions constitute the feeder control command. This command, in the format of a standard industrial communication protocol (such as Modbus-TCP, PROFINET, or OPC-UA), is sent to the furnace's underlying control system, such as a programmable logic controller (PLC). Upon receiving the command, the controller drives the feeder's actuators to precisely adjust their mechanical movements, thereby changing the feeder's speed and mode. The ultimate result of this series of actions is to reshape the spatial layout of the physical material blanket, making its shape and distribution approach the preset target state described by the target material blanket's shape parameters, thus laying an optimized physical foundation for subsequent combustion and bubbling coordinated control.

[0061] For example, following the example from step S2, we have already solved for the target blanket shape parameter as the average length of the blanket. Meters. Now, we need to generate feeder control instructions based on this parameter.

[0062] In the example of step S2, we established the relationship between the operating parameters of the feeder and the average length of the material blanket: .in, This refers to the material pushing stroke (unit: millimeters). It is the vibration frequency (unit: Hertz).

[0063] Now, we will determine the shape parameters of the target blanket. Substituting the meter into the equation: This is an equation with two unknowns, resulting in an infinite number of solutions. To obtain a definite set of operating parameters, we adopt a strategy of fixing one of the variables. Based on the analysis of the feeder's energy consumption characteristics, we set the vibration frequency... For example, a benchmark value with high energy efficiency Hertz. Substitute Hertz into the equation and solve for the required pusher stroke. : First, calculate the known terms on the right side of the equation: Then, move the constant term to the left side of the equation: , Finally, solve... : The calculated pusher stroke is 70 mm.

[0064] Therefore, the generated control command for the feeder is: set the pushing stroke to 70 mm and the vibration frequency to 15 Hz. Once this command is issued and executed, the feeder will operate according to the new parameters, resulting in a reshaping of the spatial layout of the physical blanket, achieving an average length of 5 meters, thereby realizing efficient heat absorption conditions.

[0065] See Figure 2 S4. Identify the abrupt change in the causal effect intensity between the glass melt temperature variable and the burner node variable in the structural causal diagram, eliminate the failed measurement and control mapping relationship, and generate a combustion control sub-region topology layout that matches the current physical blanket spatial layout.

[0066] In a specific embodiment of the present invention, the information on the sudden change in the intensity of causal effect between the glass melt temperature variable and the burner node variable in the structural causal graph is identified, the failed measurement and control mapping relationship is eliminated, and a combustion control sub-region topology layout that matches the current physical blanket spatial layout is generated, including: collecting the glass melt temperature parameters of the current specific area of ​​the glass melting physical space located according to the current physical blanket spatial layout.

[0067] Input the current glass melt temperature parameters of a specific region into the structural cause-effect diagram, and calculate the real-time causal effect intensity value between the glass melt temperature variable and the node variables of each burner.

[0068] By comparing real-time causal effect strength values ​​with historical causal effect strength values, information on abrupt changes in causal effect strength can be extracted.

[0069] Based on the information on sudden changes in the intensity of causal effects, the initial measurement and control mapping relationship that failed due to hotspot drift can be located.

[0070] Remove the failed initial measurement and control mapping relationships from the structural cause-effect graph to obtain the remaining valid measurement and control mapping relationships.

[0071] Based on the effective measurement and control mapping relationship and the current physical blanket space layout, the corresponding closed loops of the burner and thermocouple are redefined to generate the combustion control sub-region topology layout.

[0072] Specifically, after reshaping the spatial layout of the physical blanket according to the optimal heat absorption target in step S3, the heat transfer path inside the furnace changes accordingly. The original thermodynamic response relationship between the burner and the temperature measuring point may no longer be applicable. The purpose of this step is to dynamically identify this change and adjust the internal structure of the combustion control accordingly to ensure the accuracy and effectiveness of the control.

[0073] This process begins with sensing the current state of the system. First, thermocouples installed within the physical space of the glass melting process are used to collect temperature parameters of the molten glass in a specific area. These parameters provide direct evidence of the thermal field distribution under the new blanket layout.

[0074] Next, the newly acquired data is analyzed using the structural causal graph constructed in step S1 to quantify the actual impact of the burner on temperature. The current glass melt temperature parameters in a specific area, along with synchronously acquired burner valve opening parameters, are input into the mathematical model represented by the structural causal graph. The real-time causal effect strength value between the glass melt temperature variable and each burner node variable is calculated. The real-time causal effect strength value is a quantitative indicator representing the degree to which a change in unit burner output energy causes a temperature change at the corresponding temperature measurement point. For a given burner node variable... and glass melt temperature variables The causal pair formed is the real-time causal effect strength value. The parameters can be calculated using online parameter estimation methods. For example, a linear regression of data within a time window can be performed using the least squares method, and the calculation formula is as follows: In the formula, Indicates the current moment The calculated real-time causal effect intensity value is in units of temperature per burner parameter, such as degrees Celsius / valve opening percentage. and They represent past moments. The first collection The burner and the first Data from each temperature measurement point. and In the length of The average value within the time window. The settings are based on the sensor sampling frequency and the thermal inertia of the furnace, and are usually set to cover the number of data points for 3 to 5 major operating condition fluctuation cycles. For example, based on historical data analysis, 60 data points are set.

[0075] After obtaining real-time values, it is necessary to determine whether the changes are significant. This is done by comparing the real-time causal effect strength values. Compared with the historical causal effect strength values ​​stored in the system This involves extracting information on abrupt changes in the intensity of causal effects. Historical causal effect intensity values ​​are baseline values ​​calculated and calibrated under initial or steady-state conditions. The difference between these two values ​​is then analyzed. This constitutes the core of mutation information: when Exceeding a preset tolerance threshold At that point, a significant abrupt change in the strength of the causal effect is considered to have occurred. This threshold... It was determined based on process requirements and statistical analysis of historical data.

[0076] In a specific embodiment of the present invention, the tolerance threshold The typical range of values ​​is set as the historical causal effect strength benchmark value. 20% to 30% (e.g., 25%). This value range is set based on the following: Firstly, during normal steady-state operation of a glass melting furnace, the natural fluctuations in heat transfer caused by thermocouple sensor noise, minor fluctuations in natural gas pipeline pressure, and differences in the batch properties of conventional raw materials are typically distributed within... Within this range, conforming to the historical data distribution The principle of normal deviation; on the other hand, when a real "large displacement of the material blanket space blocks flame radiation" or "sudden change in combustion conditions causes hot spot physical drift" occurs, the resulting cliff-like decrease in single-point heat transfer efficiency is usually greater than 20%.

[0077] Based on information about abrupt changes in the intensity of causal effects, the failure of the control loop can be precisely located. Due to hot spot drift caused by blanket movement, the heat generated by a burner may no longer be effectively transferred to the temperature measurement point it was originally responsible for, which will lead to a disconnect between them. A sharp drop, thus Exceeding the threshold. This correspondence, i.e., the initial measurement and control mapping relationship, is identified as having failed due to hotspot drift.

[0078] Subsequently, the system removes failed initial measurement and control mapping relationships from the structural causal graph. At the model level, this means disconnecting or significantly reducing the weight of the corresponding causal edges. After removal, the remaining connections in the graph represent the still valid and reliable causal paths under the current operating conditions; these paths constitute the remaining valid measurement and control mapping relationships.

[0079] Finally, based on these effective measurement and control mapping relationships and the known current physical blanket spatial layout, the system reorganizes its control structure. Specifically, for each key temperature measurement point, from all burners with a valid causal relationship, the burner with the strongest causal effect and the most suitable physical location is selected as its main control unit, and the corresponding closed loops between the burners and thermocouples are redefined. This new pairing set is the combustion control sub-region topology layout that matches the current physical blanket spatial layout, providing a dynamically updated and effective basis for subsequent precise temperature control.

[0080] For example, suppose there are two burner node variables in the glass melting physical space. , and two glass melt temperature variables , In the initial state, based on the structural cause-effect graph constructed in step S1, the initial measurement and control mapping relationship is determined as follows: Main control , Main control Its historical causal effect strength value is calibrated based on long-term stable operating data as follows: Celsius / Opening % Celsius / opening degree%.

[0081] The cross-historical causal effect strength values ​​are relatively small. Celsius / Opening % Celsius / opening degree%.

[0082] After step S3 is executed, a new physical blanket spatial layout is formed, and the blanket moves towards... The area near the temperature measurement point extends, blocking the flow of... Part of the thermal radiation.

[0083] First, collect the current molten glass temperature parameters and burner parameters for a specific area after a period of time following the implementation of the new layout. Next, input this new data into the structural cause-effect graph and use the formula described above to calculate the real-time causal effect intensity value. Assume the calculation result is: Celsius / Opening % Celsius / Opening % Celsius / opening percentage, due to heat being reflected by the blanket, some energy is diverted to area. Celsius / opening percentage. Then, compare real-time values ​​with historical values ​​to extract information on abrupt changes in the intensity of causal effects. Set a tolerance threshold. This is 25% of the historical value.

[0084] for 25% of the historical value is Because 1.7 > 0.625, a mutation occurs. For 25% of the historical value is Because 0.1 < 0.6, no mutation occurred. For 25% of the historical value is Because 0.9 > 0.15, a mutation also occurs.

[0085] Based on mutation information The initial telemetry and control mapping relationship that failed due to hotspot drift was located. arrive The control loop. Remove this failure relationship from the structural cause-effect diagram. The remaining valid measurement and control mapping relationships include... , , .

[0086] Finally, based on the effective measurement and control mapping relationship and the current physical blanket spatial layout, the control closed loop is redefined. For temperature measurement points... The burner that affects it is (strength For temperature measurement points The burner that affects it is (strength )and (strength ).

[0087] Considering the effect of B1 on T2 ( The residual effect on T1 is much greater than that on T1. Furthermore, B2 has the strongest effect on T2. ​​To avoid overloading a single actuator and achieve optimal temperature control, the system performs the following allocation: control of T2 is assigned to B2, which has the strongest effect on it; control of T1 is assigned to B1, while simultaneously adjusting the newly decayed causal connection weights (…). The control gain parameter corresponding to B1 was significantly increased using a PID algorithm to compensate for the heat conduction loss caused by the blanket obstruction. Ultimately, the generated combustion control sub-region topology is: B2 is controlled by T2, and B1 is controlled by T1.

[0088] right In terms of the strongest effect, it comes from ( ).right In terms of the strongest effect, it comes from ( Considering right The effect ( ) much greater than The effect ( The final combustion control sub-region topology layout is adjusted as follows: control , Switch to auxiliary control Alternatively, its output can be adjusted based on the global optimization objective. A well-defined topology layout could be: Main controller , Main controller gradient or with Collaborative control. A clearer distinction is to... Assigned to the one with the strongest effect ,Will Assign to the one with the strongest effect on it However, to avoid overloading a single actuator, the system will choose to... Assigned to the second strongest effect (Although the effect has diminished). Ultimately, the new combustion control sub-region topology may be... control , control (However, the gain parameters of the control algorithm will be based on the new...) (Major adjustments needed). Or, a more dynamic layout is... control , control Marked as weakly coupled, its control priority is reduced. The most reasonable generation result is: control , Now changed to control (because (It is its current strongest output), and Then by Auxiliary control (due to) Finally, the generated combustion control sub-region topology layout is as follows: , , .

[0089] See Figure 3 S5. Based on the topology layout of the combustion control sub-region, calculate the synthesis thermal energy parameters required to maintain the target viscosity of the glass melt in the current region, and generate the corresponding top flame extension length and direction parameters.

[0090] In a specific embodiment of the present invention, the synthesis thermal energy parameters required to maintain the target viscosity of the molten glass in the current region are calculated based on the topological layout of the combustion control sub-region, and the corresponding top flame extension length and direction parameters are generated, including: obtaining the real-time viscosity parameters and target viscosity parameters of the molten glass in the current region as described above, and calculating the viscosity deviation value.

[0091] Based on the viscosity deviation value and the topology of the combustion control sub-region, the synthesis thermal energy parameters required to compensate for the viscosity deviation are calculated.

[0092] The synthetic thermal energy parameters are converted into the fuel flow rate and combustion air volume of the burner, generating the corresponding top flame extension length and direction parameters.

[0093] Specifically, after generating the combustion control sub-region topology layout that matches the current physical blanket spatial layout in step S4, the goal of this step is to perform precise energy regulation based on this layout to maintain the stability of key process parameters of the molten glass. The core process parameter here is viscosity, which directly affects the clarification and homogenization of the glass.

[0094] The process first requires quantifying the deviation between the current process state and the target state. By collecting the temperature parameters of the molten glass in a specific area, the real-time viscosity parameters of the molten glass in that area can be calculated. There is a clear physical relationship between the viscosity and temperature of molten glass, typically described by the Vogel-Fulcher-Tammann equation. Therefore, the first step is to obtain the real-time viscosity parameters of the molten glass in the current area. Target viscosity parameters required by the process Real-time viscosity parameters The real-time temperature is obtained through the following formula. The calculation shows that: In the formula, , and These are empirical constants for specific glass components, obtained through laboratory measurements or fitting of historical data. They are set based on the physical property data of more than 500 components in a glass formulation database. This is the real-time temperature of the current region. Target viscosity parameter. These are pre-set process values ​​based on the quality requirements of the glass products. Then, the viscosity deviation between the two is calculated. : This viscosity deviation value It is an error signal that drives subsequent control actions.

[0095] In a specific embodiment of the present invention, parameters Physically, this represents the lower limit of the initial viscosity of a system at the high-temperature limit, and its typical range is as follows: to Pascal-second, taken as 1 Pascal-second in this example; parameter It is a pseudo-temperature constant related to the activation energy of glass network structure rearrangement, which determines the sensitivity of the glass melt viscosity to temperature changes. The typical value range is 3000 to 5000 degrees Celsius, with a value of 3500 degrees Celsius. (Parameter) Physically, it corresponds to the ideal glass transition temperature (Vogel temperature), which is the asymptote temperature when the viscosity approaches infinity. The typical range is 200 to 500 degrees Celsius, with a value of 450 degrees Celsius.

[0096] Next, based on the viscosity deviation value and the combustion control sub-region topology determined in step S4, the total heat required to compensate for this deviation needs to be calculated. This calculation process typically employs a proportional-integral-derivative (PID) control algorithm. The combustion control sub-region topology plays a crucial role here, as it clarifies which burners should collaboratively provide energy to adjust the viscosity of the current region. The calculated result is the combined thermal parameters. This represents the total thermal power that all relevant burners within this control sub-zone need to output together. Its calculation formula is: In the formula, At any moment The required synthesis thermal energy parameters are expressed in kilowatts. It is a moment The viscosity deviation value. , and These are the proportional, integral, and derivative gain coefficients of the PID controller. These coefficients are tuned using system identification methods such as step response testing of the furnace to ensure the response speed and stability of the control system.

[0097] The final step is to extract this abstract synthetic thermal energy parameter. This translates into physical quantities that the burner can perform. This includes calculating the fuel flow rate and combustion air volume for each relevant burner. Synthetic thermal parameters. The weights need to be assigned to each burner within the combustion control sub-zone according to the weights defined in the sub-zone topology layout. For burners assigned to the... Thermal energy parameters of each burner Its required fuel flow The calculation is as follows: In the formula, It is the calorific value of the fuel used. For example, the calorific value of natural gas is 36,000 kJ / m³, which is a known physical constant. It is the first The combustion efficiency of each burner under current operating conditions, calibrated based on burner performance curves and historical operating data. Combustion air volume. The fuel flow rate and combustion air volume are determined based on the stoichiometric ratio of the fuel type and the combustion chemical reaction, typically with a slightly excess air coefficient to ensure complete combustion. This set of calculated fuel flow rate and combustion air volume jointly determines the physical morphology of the flame, namely, the corresponding top flame extension length and direction parameters. Higher fuel and combustion air flow rates mean higher exit momentum, resulting in a longer flame; while the ratio of these two, along with the burner nozzle design, determines the flame direction and shape. These parameters are ultimately issued as control commands.

[0098] For example, following the example from step S4, a new combustion control sub-region topology layout has been generated, in which the burner Designated as the temperature measurement point The main actuator for the temperature and viscosity of the region.

[0099] First, obtain the real-time viscosity parameters and target viscosity parameters of the molten glass in the current region. Assuming that... The current temperature measured at the temperature measuring point is 1450 degrees Celsius. The VFT equation parameters for the glass used are: Pascal's second Celsius Celsius.

[0100] Calculate real-time viscosity parameters : Pascal-second. Assuming the target viscosity parameter required by the process. The value is 25 Pascal-second. Calculate the viscosity deviation value. : Pascal-seconds. A positive value indicates that the current viscosity is too high and the temperature is too low, requiring heating. Next, the synthesis thermal energy parameters are calculated based on the viscosity deviation value and the topology layout of the combustion control sub-region. In this example, the control sub-region mainly consists of the burner. Responsible. We use a simplified proportional-integral (PI) controller for calculation, whose gain coefficient is tuned based on historical data. kilowatts per (Pascal-second), kilowatts per (Pascal-second-second). Assume the cumulative value of the integral term is 20 Pascal-second-second.

[0101] Calculate the required synthesis thermal energy parameters : The calculated result is positive, indicating that the system needs to add 505.75 kilowatts of heat to this area. Finally, the synthesized thermal energy parameters are converted into the burner's fuel flow rate and combustion air volume. Since this sub-region is composed of... The main controller, we allocate all the heat to it, that is Kilowatts. Assuming the fuel is natural gas, its calorific value... The burner efficiency is 36,000 kJ / m³. It is 0.8.

[0102] Calculate fuel flow : cubic meters per second. Assuming the theoretical air requirement for complete combustion of natural gas is 10 cubic meters of air per cubic meter of fuel, and setting a 10% excess air, then the combustion air volume... for: cubic meters per second.

[0103] The calculated fuel flow rate (0.01756 cubic meters per second) and combustion air volume (0.19316 cubic meters per second) are then sent to the burner. The controller will generate a flame with specific energy and momentum, the shape of which is determined by the corresponding top flame extension length and direction parameters. This flame will precisely direct... The area provides the necessary heat.

[0104] S6. Generate the start / stop topology diagram of the bottom bubbling valve array based on the extension length and direction parameters of the top flame, and output local follow-up bubbling control commands to realize intelligent control of glass melting energy consumption.

[0105] In a specific embodiment of the present invention, a start-stop topology map of the bottom bubbling valve array is generated based on the extension length and direction parameters of the top flame, and a local follow-up bubbling control command is output to realize intelligent control of glass melting energy consumption. This includes: calculating the thermal radiation flux distribution through a set three-dimensional computational fluid dynamics (CFD) mesh model, and extracting the coordinates of the region with radiation flux greater than a set threshold as the first target coordinates.

[0106] The target bubbling area is determined based on the coordinates of the first target and the coordinates of the area in the current physical blanket spatial layout where the accumulated thickness is greater than the set thickness threshold.

[0107] Based on the mapping of the target bubbling action area to the bottom bubbling valve array, a start-stop topology diagram of the bottom bubbling valve array is generated.

[0108] Analyze the start / stop topology of the bottom bubbling valve array and extract the node information of the bubbling valve in the open state.

[0109] Based on the bubbling valve node information and the synthesized thermal energy parameters, the target output gas volume required for each bubbling valve in its open state is calculated.

[0110] The target output air volume is packaged into a local follow-up bubbling control command and sent to the underlying programmable logic controller.

[0111] Specifically, this step aims to achieve intelligent coordination between top combustion heating and bottom bubbling agitation, transforming the flame energy parameters calculated in step S5 into a physical operation of precisely agitating the bottom layer of the molten glass. The core idea is to allow the bubbles to "follow" the hot spots of the flame, applying agitation to the areas where heat transfer is most needed, thereby maximizing energy utilization efficiency.

[0112] The process begins with the analysis of the energy distribution of the top flame. From the top flame extension length and direction parameters generated in step S5, a calibrated combustion model is used. This model is either a pre-calculated and stored three-dimensional computational fluid dynamics (CFD) mesh model or a spatial thermal radiation mapping matrix trained based on historical multi-band infrared thermal imager data. The three-dimensional temperature field and thermal radiation flux distribution of the flame above the molten pool can then be calculated. The coordinates of the region with a radiation flux greater than a set threshold are extracted as the first target coordinates. These coordinates represent the location where the top heat input is most concentrated.

[0113] Simultaneously, it is necessary to determine the locations where heat most needs to be absorbed. Based on the current physical blanket spatial layout reshaped in step S3, the coordinates of areas where the raw material accumulation thickness exceeds a set thickness threshold can be determined. The ideal heat transfer path is to efficiently transfer heat from areas with high thermal radiation to the molten glass beneath the insufficiently melted raw material. Therefore, the first target coordinates are spatially aligned with the coordinates of the areas where the accumulation thickness exceeds the set thickness threshold; that is, the geometric centroid distance between the two is calculated in the same horizontal projection coordinate system. When the distance is less than the preset threshold, an overlapping or adjacent area is determined as the target bubbling effect area. This area is the key location for implementing enhanced heat transfer.

[0114] After determining the target bubbling action area, it needs to be mapped onto the actual physical actuators. A grid-like array of bottom bubbling valves is installed at the bottom of the glass furnace. The geometry of the target bubbling action area is projected onto the bottom bubbling valve array. All bubbling valves whose coordinates fall within this projection range are set to the "open" state, and the rest to the "closed" state. This generates a binary topology diagram of the bottom bubbling valve array's start / stop, which visually shows which bubbling valves need to be activated in this coordinated operation.

[0115] Subsequently, this topology graph is analyzed to extract the identifiers and location information of all bubbling valves in the "open" state, forming the node information of the bubbling valves in the open state. This determines "where" bubbling occurs.

[0116] Next, the "bubbling intensity" needs to be determined. The bubbling intensity, i.e., the gas flow rate, should be matched to the heat received by the region. The synthesis thermal energy parameters calculated in step S5 are then used. It can calculate the total output air volume required for all bubbling valves in the open state. The relationship between the two is determined by a heat transfer enhancement coefficient. To establish: In the formula, The unit is standard cubic meters per hour. It is a dimensionless coefficient or a coefficient with specific units, whose dimensions must be consistent with those of a given coefficient. and Unit matching is used to ensure consistency. For example, if The unit is kilowatt. The unit is cubic meters per hour. The unit is (cubic meters per hour) per kilowatt. This coefficient... The setting is based on regression analysis of experimental data on heat transfer enhancement effects under more than 150 sets of different bubbling gas volumes. It balances the relationship between improved heat transfer efficiency and potential glass defects caused by bubbling. The calculated total output gas volume It will be evenly distributed to all open bubbling valves. If there are If each bubbling valve is opened, the target output air volume required for each valve is... for: Finally, the information of each open bubbling valve node and its corresponding target output gas volume are recorded. The data is combined and packaged into a data frame that conforms to the communication protocol of the underlying controller. This data frame is the local follow-up bubbling control command, which is sent to the underlying programmable logic controller (PLC) to precisely control the opening degree and air volume of the corresponding valve, completing a full bottom agitation operation in coordination with the top heating.

[0117] For example, following the example from step S5, we have already calculated the amount that needs to be supplied by the burner. Provide to target area The combined thermal energy parameters of kilowatts were obtained, and the corresponding top flame extension length and direction parameters were generated.

[0118] First, according to the burner The fuel flow rate and combustion air volume are calculated using a combustion simulation model. The coordinates of regions with radiant flux exceeding a set threshold are extracted as the first target coordinates. The unit is meters.

[0119] Next, based on the current spatial layout of the physical blanket, the coordinates of the area where the stacked thickness exceeds the set thickness threshold are determined. The two coordinates are spatially close, therefore the target bubbling area is determined to be a square region with a side length of 2 meters centered at point (5.2, 7.9). The x-coordinate range of this area is [4.2, 6.2] meters, and the y-coordinate range is [6.9, 8.9] meters.

[0120] Then, Mapping to the bottom bubbling valve array. Assume the bottom bubbling valve array contains valve nodes located at (5,8) and (6,8). Node (5,8): x=5 is within [4.2,6.2], y=8 is within [6.9,8.9]. This node is active. Node (6,8): x=6 is within [4.2,6.2], y=8 is within [6.9,8.9]. This node is active. Assume there is also node (5,9), whose y=9 is not within [6.9,8.9], therefore it is not active. The generated bottom bubbling valve array start / stop topology marks nodes (5,8) and (6,8) as "on", and the rest as "off". Parse this topology and extract the information of the bubbling valve nodes in the open state as {node (5,8), node (6,8)}. Number of open valves. .

[0121] Calculate the target output gas volume based on the synthetic thermal energy parameters. Set the heat transfer enhancement coefficient. (cubic meters / hour) / kilowatt. Calculate the total output gas volume. : cubic meters per hour. Allocate the total gas volume to the two open valves and calculate the target output gas volume for each valve. : cubic meters per hour.

[0122] Finally, the calculation results are packaged into a local follow-up bubbling control instruction. This instruction contains two pieces of information: {valve node (5,8), target gas flow rate: 0.50575 cubic meters / hour} and {valve node (6,8), target gas flow rate: 0.50575 cubic meters / hour}. This instruction is then sent to the underlying programmable logic controller for execution.

[0123] In a specific embodiment of the present invention, after outputting the local follow-up bubbling control command, the process includes: collecting updated glass melt temperature parameters and updated energy consumption parameters after executing the local follow-up bubbling control command.

[0124] The updated glass melt temperature parameters and updated energy consumption parameters are fed back to the structural cause-effect graph for model verification, and the prediction error value is calculated.

[0125] When the prediction error exceeds the tolerance threshold, the parameter update mechanism of the structural causal graph is triggered to correct the causal connection weights between spatial heat transfer variables.

[0126] Specifically, after outputting and executing the local follow-up bubbling control command, the entire control system forms a complete closed loop. This step is the feedback and adaptive learning link in this closed loop, and its purpose is to continuously verify and optimize the structural causal graph constructed in step S1, ensuring that the model can accurately reflect the dynamic changes in the physical space of glass melting over a long period of time.

[0127] The process begins with data acquisition. After the local follow-up bubbling control command in step S6 has been executed for a preset time—the duration of which is set according to the thermal inertia of the furnace to ensure the system reaches a new quasi-steady state—the system will collect two types of new data through a sensor network. The first type is the updated glass melt temperature parameters, i.e., the latest temperature readings measured at various spatially discrete temperature measurement points. The second type is the updated energy consumption parameters, including the total fuel consumption rate of all burners and the total power consumption of the bubbling system, etc. These data reflect the actual cost of executing the control command.

[0128] After acquiring new data, model validation is required. The updated molten glass temperature parameters and energy consumption parameters, along with the control command parameters that generated these results, such as feeder parameters, burner valve opening, and bubbling gas volume, are fed back into the structural cause-effect graph. The structural cause-effect graph, as a quantitative predictive model, can predict the system's expected state based on the input control parameters. The model-predicted temperature values ​​are compared with the actually collected updated molten glass temperature parameters, and the prediction error is calculated for threshold determination. For the first... Prediction error value of each temperature measurement point The calculation formula is as follows: In the formula, It is the first Updated glass melt temperature parameters collected from each temperature measuring point It is the temperature value of the temperature measurement point predicted by the structural cause-effect graph based on the control input. The calculations depend on the structural equations in the model, for example, for a burner. Affected temperature measurement points Its predicted value can be expressed as: In the formula, It represents the burner. Temperature measurement point Causal connection weights that influence strength It is a burner Valve opening parameters, It is a bias term representing the influence of other background heat sources.

[0129] After calculating the prediction error value, the system will determine whether the model needs correction. When the prediction error value at a certain temperature measurement point... Exceeding the preset tolerance threshold When this occurs, it indicates that the model's description of the current physical process is no longer accurate, triggering the parameter update mechanism of the structure-causal graph. Tolerance threshold. It is set according to the precision requirements of glass production processes. For example, based on the analysis of its impact on product quality, it is set to 5 degrees Celsius.

[0130] In a specific embodiment of the present invention, the tolerance threshold The typical value range is set to 5 to 10 degrees Celsius (e.g., preferably 5 degrees Celsius). This threshold range is primarily based on two hard engineering constraints: firstly, the inherent noise filtering of the underlying sensor hardware, as platinum-rhodium thermocouples commonly used in glass melting furnaces... to Under high-temperature operating conditions, its inherent measurement errors and background noise caused by high-frequency thermal disturbances in the environment are usually... to Fluctuations between.

[0131] Once the parameter update mechanism is activated, its core task is to correct the causal linkage weights between space heat transfer variables. The causal linkage weights are those specified in the aforementioned formula. , is the core parameter for quantifying the strength of causal relationships. The correction process employs an online learning algorithm, using a loss function constructed based on mean squared error (MSE) to fine-tune the current data. A commonly used update rule is gradient descent, with the following weight update formula: In the formula, It is the updated causal connection weight. This is the weight before the update. It is a learning rate parameter with specific units, used to ensure dimensional balance on both sides of the equation. Its units must be the reciprocal of the square of the corresponding burner parameter unit (e.g., opening degree). The numerical values ​​are set to ensure stable convergence of the learning process, and are typically selected between 0.0001 and 0.01 through simulation experiments. Through this iterative process, the parameters of the structural causal graph are continuously optimized, enabling it to adapt to slow changes such as furnace aging and fluctuations in raw material composition, thereby maintaining the long-term effectiveness of the entire intelligent control system.

[0132] For example, following the example of step S6, the system has sent to the burner Control commands were issued to the bottom bubbling valve array. Assuming that in step S4, the burner... With temperature measurement point The causal connection weights between them were determined as follows: Celsius / Opening %, Offset Term Degrees Celsius. In step S5, for The valve opening is set to 65%.

[0133] First, after the command is executed and stabilized, the temperature measurement points are collected. The updated glass melt temperature parameters are as follows: Celsius.

[0134] Next, model validation will be performed. The model will predict based on the input parameters. Temperature: Degrees Celsius. Calculate the prediction error value. : Celsius.

[0135] Then, determine whether an update is needed. The set tolerance threshold. The temperature is set to 5 degrees Celsius. The prediction error is compared to the tolerance threshold: 5.5 > 5. Since the prediction error exceeds the tolerance threshold, the parameter update mechanism of the structural causal graph is triggered. The learning rate is then set. .

[0136] Correcting the causal connection weights between space heat transfer variables : , , , , .

[0137] After this update, the burner Temperature measurement point The causal connection weights have been revised from 2.3 to 2.3715 degrees Celsius / opening percentage. This updated structural causal graph will be used in the next round of control calculations, making the model's predictions and control more closely resemble the real physical processes.

[0138] The above content is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined by the present invention, and all such modifications and additions should fall within the protection scope of the present invention.

Claims

1. A method for intelligent prediction and control of energy consumption in glass melting, characterized in that, include: S1. Obtain the discretized equipment operating parameters and prior knowledge of fluid mechanics in the physical space of glass melting, extract spatial heat transfer variables, generate an initial state dataset, perform causal discovery processing on the initial state dataset, calculate the partial correlation coefficient between spatial heat transfer variables through conditional independence test, thereby identifying and eliminating pseudo-correlation relationships between spatial heat transfer variables, and constructing a structural causal graph that reflects the real physical heat transfer direction. S2. Perform causal intervention on the operating parameters of the feeder in the structural cause-effect diagram, deduce the liquid surface thermal reflectivity under different material blanket propulsion trajectories, take the minimum liquid surface thermal reflectivity as the optimization objective, and solve the target material blanket shape parameters. S3. Generate feeding machine control commands based on the target material blanket shape parameters, adjust the feeding machine's pushing stroke and vibration frequency, and reshape the spatial layout of the physical material blanket. S4. Identify the abrupt change in the causal effect intensity between the glass melt temperature variable and the burner node variable in the structural causal diagram, eliminate the failed measurement and control mapping relationship, and generate a combustion control sub-region topology layout that matches the current physical blanket spatial layout. S5. Based on the topological layout of the combustion control sub-region, calculate the synthesis thermal energy parameters required to maintain the target viscosity of the molten glass in the current region, and generate the corresponding top flame extension length and direction parameters. S6. Generate the start / stop topology diagram of the bottom bubbling valve array based on the extension length and direction parameters of the top flame, and output local follow-up bubbling control commands to realize intelligent control of glass melting energy consumption.

2. The intelligent prediction and control method for glass melting energy consumption according to claim 1, characterized in that: The process involves obtaining discretized equipment operating parameters and prior knowledge of fluid mechanics in the physical space of glass melting, extracting spatial heat transfer variables, generating an initial state dataset, performing causal discovery processing on the initial state dataset, calculating partial correlation coefficients between spatial heat transfer variables through conditional independence tests to identify and eliminate spurious correlations between spatial heat transfer variables, and constructing a structural causal graph reflecting the true physical heat transfer direction, including: The mechanical motion parameters of the feeder, the opening parameters of the burner valve, the gas volume parameters of the bubble array at the bottom of the tank, and the spatial discrete temperature measurement point parameters of the glass melting physical space are obtained. Combined with prior knowledge of fluid mechanics, an initial state dataset is generated. Perform conditional independence tests on the initial state dataset to identify the correlation matrix among spatial heat transfer variables; Based on the correlation matrix, false temperature correlation data caused by flue gas flow are eliminated, and a structural causal graph reflecting the true physical heat transfer direction is constructed.

3. The intelligent prediction and control method for energy consumption in glass melting according to claim 1, characterized in that: The process involves performing causal intervention on the operating parameters of the feeding machine in the structural causal graph, extrapolating the liquid surface thermal reflectivity under different material blanket propulsion trajectories, and using the minimum liquid surface thermal reflectivity as the optimization objective to solve for the target material blanket shape parameters, including: In the structural causal diagram, the operating parameters of the feeder are set as intervention variables to construct a counterfactual inference model; Input the intervention variables into the counterfactual model to calculate the liquid surface thermal reflectivity corresponding to different material blanket propulsion trajectories; With minimizing the thermal reflectivity of the liquid surface as the optimization objective, the counterfactual inference model is solved, and the target material blanket shape parameters are output.

4. The intelligent prediction and control method for energy consumption in glass melting according to claim 1, characterized in that: The process of generating control commands for the feeding machine based on the shape parameters of the target material blanket, adjusting the feeding machine's pushing stroke and vibration frequency, and reshaping the spatial layout of the physical material blanket includes: Obtain the target material blanket shape parameters, and call the structural equation between the feeder operating parameters and the material blanket shape parameters in the counterfactual inference model to construct the inverse solution model; A reference vibration frequency based on the equipment energy efficiency curve is introduced as a constraint condition. The constraint condition is substituted into the inverse solution model, and the structural equation is simplified into a single variable equation about the pushing stroke. Solve the single-variable equation to obtain the target pushing stroke corresponding to the shape parameters of the target blanket; The target feeding stroke is combined with the reference vibration frequency and converted into a standard industrial communication protocol format to generate feeder control commands. The control commands of the feeder are sent to the underlying programmable logic controller to drive the feeder's actuators to adjust mechanical actions and reshape the spatial layout of the physical material blanket.

5. The intelligent prediction and control method for glass melting energy consumption according to claim 1, characterized in that: The information on abrupt changes in the intensity of causal effects between the molten glass temperature variable and the burner node variable in the identified causal graph includes: Collect the temperature parameters of the molten glass in a specific area of ​​the physical space of glass melting, based on the current physical material blanket spatial layout. Input the current glass melt temperature parameters of a specific region into the structural cause-effect diagram, and calculate the real-time causal effect strength value between the glass melt temperature variable and the variables of each burner node; By comparing real-time causal effect strength values ​​with historical causal effect strength values, information on abrupt changes in causal effect strength can be extracted.

6. The intelligent prediction and control method for glass melting energy consumption according to claim 1, characterized in that: The elimination of failed measurement and control mapping relationships generates a combustion control sub-zone topology layout that matches the current physical blanket spatial layout, including: Based on the information on abrupt changes in the intensity of causal effects, the initial measurement and control mapping relationship that failed due to hotspot drift can be located; Remove the failed initial measurement and control mapping relationships from the structural cause-effect graph to obtain the remaining valid measurement and control mapping relationships; Based on the effective measurement and control mapping relationship and the current physical blanket space layout, the corresponding closed loops of the burner and thermocouple are redefined to generate the combustion control sub-region topology layout.

7. The intelligent prediction and control method for energy consumption in glass melting according to claim 1, characterized in that: The calculation of the synthesis thermal energy parameters required to maintain the target viscosity of the molten glass in the current region based on the combustion control sub-region topology layout generates the corresponding top flame extension length and direction parameters, including: Obtain the real-time viscosity parameters and target viscosity parameters of the molten glass in the current area as described above, and calculate the viscosity deviation value; Based on the viscosity deviation value and the topology of the combustion control sub-region, calculate the synthesis thermal energy parameters required to compensate for the viscosity deviation; The synthetic thermal energy parameters are converted into the fuel flow rate and combustion air volume of the burner, generating the corresponding top flame extension length and direction parameters.

8. The intelligent prediction and control method for energy consumption in glass melting according to claim 1, characterized in that: The process of generating the start / stop topology diagram of the bottom bubbling valve array based on the extension length and direction parameters of the top flame includes: The thermal radiation flux distribution is calculated using a set three-dimensional computational fluid dynamics (CFD) mesh model, and the coordinates of regions with radiation flux greater than a set threshold are extracted as the first target coordinates. The target bubbling area is determined based on the coordinates of the first target and the coordinates of the area in the current physical blanket spatial layout where the accumulated thickness is greater than the set thickness threshold. Based on the mapping of the target bubbling action area to the bottom bubbling valve array, a start-stop topology diagram of the bottom bubbling valve array is generated.

9. The intelligent prediction and control method for energy consumption in glass melting according to claim 1, characterized in that: The output local follow-type bubbling control command realizes intelligent control of glass melting energy consumption, including: Analyze the start / stop topology of the bottom bubbling valve array and extract the node information of the bubbling valve in the open state; Based on the bubbling valve node information and the synthesized thermal energy parameters, calculate the target output gas volume required for each bubbling valve in its open state. The target output air volume is packaged into a local follow-up bubbling control command and sent to the underlying programmable logic controller.

10. The intelligent prediction and control method for energy consumption in glass melting according to claim 1, characterized in that: After outputting the local follow-type bubbling control command, it includes: Collect updated glass melt temperature parameters and updated energy consumption parameters after executing the local follow-up bubbling control command; The updated glass melt temperature parameters and updated energy consumption parameters are fed back to the structural cause-effect graph for model verification, and the prediction error value is calculated. When the prediction error exceeds the tolerance threshold, the parameter update mechanism of the structural causal graph is triggered to correct the causal connection weights between spatial heat transfer variables.

Images