An integrated mechanism reconstruction and knowledge decision-driven electrolytic cell collaborative control method

Abstract

The application discloses a kind of integrated mechanism reconfiguration and electrolytic cell collaborative control method of knowledge decision drive, it is related to metallurgical automation and industrial artificial intelligence technical field.The method comprises: first, the temperature field in the interior of electrolytic cell and side department are constructed with physical information neural network online, and the self-adaptive adjustment of model parameter is realized by introducing reinforcement learning algorithm;Through the feature selection module of fusion self-attention mechanism, the working condition data is dynamically weighted, and the global state vector representing anode polarization and thermal field fluctuation is extracted;Design composite reward function containing expert knowledge constraint and carbon emission cost, guide reinforcement learning intelligent agent to generate discharging control strategy considering energy efficiency stability and advance emission reduction.The application reduces unnecessary anode adjustment action and alumina discharging action, realizes the deep cooperation of energy balance and material balance, significantly improves current efficiency and controls greenhouse gas emission intensity.

Description

Technical Field

[0001] This invention relates to the fields of metallurgical automation and industrial artificial intelligence, and in particular to a collaborative control method for electrolytic cells that integrates mechanism reconstruction and knowledge decision-driven approaches. Background Technology

[0002] In modern aluminum electrolysis production, maintaining a suitable temperature distribution and molecular ratio within the electrolyte system is crucial for achieving efficient and stable operation. However, due to the extreme environment of high temperature and strong corrosion inside the electrolytic cell, directly measuring the internal physical field is extremely difficult and carries significant cost and risk. Traditional control methods heavily rely on the accumulated experience of operators and lagging voltage slope feedback, making it difficult to capture the dynamic correlation between thermal field evolution and greenhouse gas emissions in real time, especially when the alumina concentration is in a critically low-fluorine state, making it impossible to proactively avoid perfluorocarbon (PFC) emissions. Furthermore, existing artificial intelligence models often lack physical consistency, and parameter settings rely heavily on empirical values, making it difficult to achieve optimal performance of the algorithm in complex coupled systems.

[0003] Therefore, developing an intelligent control model with physical consistency perception capabilities and knowledge-driven collaborative capabilities has become an important direction for the industry's green transformation. Summary of the Invention

[0004] To address the aforementioned problems, the present invention aims to provide a collaborative control method for electrolytic cells that integrates mechanism reconstruction and knowledge-driven decision-making. By cross-application of deep learning, reinforcement learning, and heat transfer mechanisms, a closed-loop system covering the entire process of "perception-decision-execution" is constructed to achieve intelligent perception and precise control of key variables in the electrolysis process, thereby solving the control problem between thermal balance stability and greenhouse gas emission control in large prebaked aluminum electrolytic cells during operation.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: On the one hand, a collaborative control method for electrolyzers integrating mechanism reconstruction and knowledge decision-driven approaches is provided, the method comprising the following steps: S1. Collect multi-source operating data of the electrolytic cell using distributed sensors, construct a physical information neural network, incorporate the residual term of the heat transfer partial differential equation into the loss function, and introduce an automatic hyperparameter adjustment mechanism to achieve physical consistency reconstruction of the whole-domain temperature field and side-bonding thickness in the electrolytic cell. S2. The high-frequency harmonic amplitude features of the electrolytic cell voltage are extracted using fast Fourier transform. Combined with the reconstruction results of the physical information neural network and offline test data, a dynamically weighted global state vector is generated by a feature selection module that integrates a self-attention mechanism. S3. Construct a collaborative control engine based on deep reinforcement learning. Input the global state vector into the preset decision network and introduce an expert knowledge base constructed by fuzzy decision tree as the control boundary constraint. Based on the comprehensive reward function including voltage stability, thermal balance deviation and instantaneous carbon emission cost, solve the optimal combination command of feeding frequency and pitch adjustment. S4. The optimal combination of instructions is converted into the actions of the feeder and the anode lifting mechanism, and the abnormal response dual threshold mechanism is activated simultaneously. When the real-time operating conditions deviate from the target range or fault identification is triggered, the expert knowledge base is called to correct the strategy, so as to achieve deep collaborative control of the energy balance and material balance of the electrolytic cell.

[0006] Optionally, online multi-source operating condition data is acquired via the physical interface of the tank control box using a high-frequency sampling method, including: Electrical physical parameters: real-time tank voltage, bus current intensity, and voltage fluctuation standard deviation; Frequency domain characteristic signal: 3-10Hz high-frequency voltage noise amplitude, used to characterize the polarization degree at the bottom of the anode and the carbon emission risk; Physical field sensing data: Multi-point temperature data provided by a flexible temperature sensor array arranged on the side, bottom, and end of the cathode steel rod of the tank shell; Execution status feedback: real-time on / off status of the feeder, and current lifting / lowering stroke position of the anode busbar; Offline test data is periodically input into the laboratory information management system and used for parameter correction of the physical information neural network and deviation calculation of the reinforcement learning reward function, including: Component analysis indicators: molecular ratio, alumina concentration, calcium fluoride content; Thermodynamic references: electrolyte temperature, aluminum level, electrolyte level, and furnace bottom pressure drop.

[0007] Optionally, the physical information neural network adopts a 6-layer fully connected residual structure, with each layer containing 128 neurons, and uses tanh as the activation function to ensure the continuity of the higher-order derivatives of the temperature field distribution; its loss function L( θ The residual terms of the unsteady heat transfer partial differential equations constrained by the law of conservation of energy are incorporated into the equations. The loss function L( θ This is used to guide the training of neural networks, enabling predictions to approximate both observed data and physical laws. (1) in, This represents the total loss function value, which is minimized by the goal of training the neural network. This represents the data-driven loss, which is the mean square error between the measured and predicted values. This represents the physical mechanism loss, i.e., the physical residual term. The sum of squares at the point of configuration; This represents the boundary condition constraint loss, ensuring that the temperature field distribution conforms to the heat dissipation boundary conditions; This represents the weighting coefficient of each loss term, used to adjust the neural network's emphasis on data and mechanisms; (2) in, This is the physical residual term, which should ideally approach zero, representing the deviation in energy balance; The density of the melt; The specific heat capacity of the melt; The global scalar temperature field output by the neural network; Use time coordinates; The thermal conductivity of the material; For the Joule heat source term, from the formula Calculations show that the endogenous heat source representing the current generation is, among which It is the current density vector. The electric field intensity vector; This is a gradient operator used to calculate the rate of change of temperature in three-dimensional space.

[0008] Optionally, the hyperparameter automatic adjustment mechanism defines an agent to explore within the hyperparameter state space, using classification performance or physical residual convergence as the reward function, and automatically outputs the optimal convolution kernel size and network depth parameters for the current tank age condition.

[0009] Optionally, using automatic differentiation techniques, the temperature field inside the electrolytic cell can be reconstructed in the entire spatiotemporal domain without the need for complex mesh generation; the side heat dissipation flux is constrained by introducing Robin boundary conditions. (3) in, n The direction of the normal to the boundary surface; h The comprehensive convective heat transfer coefficient reflects the heat transfer intensity between the tank shell and the environment; For ambient reference temperature; This is the derivative of temperature along the boundary normal direction, i.e., the heat flux gradient.

[0010] Optionally, while acquiring the reconstructed temperature field, the electrolytic cell voltage sequence is analyzed in the frequency domain using Fast Fourier Transform to extract... High-frequency harmonic amplitude characteristics serve as key physical information characterizing the degree of micro-polarization at the bottom of the anode. These characteristics are input into the fusion carbon accounting module, and combined with the modified electrochemical equivalent formula, the instantaneous formation rate of non-effect perfluorinated carbon is estimated in real time. : (4) in, It is the critical voltage for local polarization, V is the instantaneous cell voltage of the electrolytic cell, and I is the electrolytic current. α、β These are empirical coefficients related to the electrolytic cell type; the instantaneous generation rate is co-mapped with the temperature field characteristics and side-slab thickness characteristics output by the physical information neural network to construct a global state vector covering the dimensions of electric field, thermal field, and carbon emissions. .

[0011] Optionally, the self-attention mechanism is implemented by: calculating the correlation weight matrix between pairs of features in the feature dimension, and then performing weighted fusion of the features after normalization to enhance the model's perception accuracy of weak polarization signals in the 3-10Hz range and avoid interference from industrial environmental noise.

[0012] Optionally, the decision network employs an Actor-Critic reinforcement learning network with an experience replay mechanism to find the optimal combination of alumina feeding frequency and electrode moment adjustment; a comprehensive reward function is designed. Its mathematical definition integrates voltage stability, thermal equilibrium stability, and emission reduction costs: (5) in, It is the value of the comprehensive reward function, which the agent seeks to maximize; due to the negative sign, it means minimizing the penalty. It is the voltage fluctuation variance of the electrolytic cell, reflecting the thermal balance stability of the electrolytic cell; This refers to the actual perceived concentration of alumina. It is the midpoint of the set target alumina concentration range; It is instantaneous carbon emission intensity, based on the instantaneous formation rate of non-effect perfluorinated carbon. Estimated environmental costs; It is the reconstructed deviation of the side joint thickness, that is, the deviation of the side joint thickness from the ideal steady state; It is the weighted penalty coefficient of the above evaluation indicators.

[0013] Optionally, carbon emission penalty weights Instead of a fixed value, the alumina concentration is dynamically adjusted nonlinearly based on whether it approaches the low-fluorine zone. When the physical information neural network detects an abnormally thinning sidewall thickness and voltage noise reflecting a high risk of carbon emissions, the penalty term in the comprehensive reward function will increase sharply, guiding the agent to solve for the preventive feeding instruction through the Bellman equation. (6) in, Q ( s, a ) is in state sTake action below a The value of the action; R It is the immediate reward obtained after performing an action, that is ; This is a discount factor used to balance current benefits with long-term expected returns; s ', a 'This represents the state and possible actions at the next moment.'

[0014] Optionally, the optimal combination command is sent to the tank control box via industrial Ethernet or Modbus protocol. The tank control box, as the underlying command hub, drives the relay or frequency converter after receiving the optimal combination command to realize pulse control of the feeding frequency of the alumina feeder and logic drive of the motor of the anode lifting mechanism.

[0015] On the other hand, an electronic device is provided, the electronic device comprising: processor; The memory stores computer-readable instructions, which, when loaded and executed by the processor, implement the steps of the electrolytic cell collaborative control method described above, which integrates mechanism reconstruction and knowledge decision-driven control.

[0016] On the other hand, a computer-readable storage medium is provided, wherein program code is stored in the computer-readable storage medium, and the program code can be called by a processor to execute the steps of the electrolytic cell collaborative control method described above, which is based on integrated mechanism reconstruction and knowledge decision-driven method.

[0017] The beneficial effects of the technical solution provided by this invention include at least the following: This invention realizes the transformation of aluminum electrolysis production from "passive response" to "proactive prediction." Through the deep integration of mechanism-driven and data-driven approaches, it enhances the physical accuracy and real-time performance of electrolytic cell condition monitoring. Utilizing a Physical Information Neural Network (PINN) with partial differential equations of heat transfer incorporated into the loss function, it achieves high-precision reconstruction of the high-temperature melt temperature field and side-bonding thickness within the electrolytic cell in seconds, without the need for complex mesh generation. Compared to the hours-long computation time of traditional finite element analysis (FEA), this invention improves the efficiency of thermal balance diagnosis and effectively solves the industry pain point of difficulty in measuring the internal state of large electrolytic cells, providing a preliminary feedback source with in-depth physical mechanism understanding for subsequent accurate decision-making.

[0018] This invention achieves deep collaborative control of energy and material balance in electrolytic cells. By constructing integrated thermal field indicators, voltage fluctuation characteristics, and carbon emission status, the reinforcement learning agent can take preventative alumina feeding and electrode spacing adjustment actions at the initial stage of microscopic shifts in thermal balance (agglomeration thickness, superheat). This control logic avoids the concentration control distortion phenomenon that occurs in traditional resistance slope control at low superheat, effectively reducing the probability of sediment formation at the bottom of the cell. While ensuring stable current efficiency above 93.5%, it reduces the cell voltage variance to within 15% of the rated value, extending the service life of the electrolytic cell.

[0019] This invention achieves a green and low-carbon transformation of electrolytic aluminum production from the source. It innovatively uses instantaneous carbon emission intensity as a negative feedback constraint in reinforcement learning, and guides the agent to actively avoid the critical under-fluorinated range that induces the anode effect through a composite reward function, thereby effectively suppressing the emission of perfluorinated carbon (PFCs).

[0020] This invention possesses strong industrial applicability and a high level of intelligence. By introducing a hyperparameter automatic adjustment mechanism and an expert knowledge base correction layer, the control strategy can adapt to parameter drift under different cell ages and current intensities, effectively replacing the subjectivity and uncertainty of traditional manual adjustments. By transforming complex industrial mechanisms into a strategy model, this invention reduces the reliance on operator experience on the production line.

[0021] Meanwhile, this invention, based on proactive intervention, reduces unnecessary anode lifting and lowering adjustments and blind alumina feeding. This steady-state operation mode reduces wear from frequent start-stop cycles of mechanical equipment, significantly slows down the erosion and oxidation rate of the anode steel claws in high-temperature electrolyte and corrosive environments, and reduces the pathways for impurities to be introduced from a physical execution perspective, thereby improving the product quality of the primary aluminum melt. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0023] Figure 1 This is a flowchart of an electrolytic cell collaborative control method integrating mechanism reconstruction and knowledge decision-driven approach provided by an embodiment of the present invention; Figure 2 This is a diagram of the intelligent control framework for electrolytic aluminum feeding based on carbon emission constraints provided in an embodiment of the present invention; Figure 3 This is a diagram of the intelligent control framework for aluminum electrolysis cells provided in an embodiment of the present invention; Figure 4 This is a diagram of the overall architecture of a mechanism-driven physical information neural network provided in an embodiment of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0025] This invention provides a method for coordinated control of electrolyzers that integrates mechanism reconstruction and knowledge-driven decision-making. This method can be implemented by an electronic device, which can be a terminal or a server. (See reference...) Figures 1-3 As shown, the processing flow of this method may include the following steps: S1. Collect multi-source operating data of the electrolytic cell using distributed sensors, construct a physical information neural network, incorporate the residual term of the heat transfer partial differential equation into the loss function, and introduce an automatic hyperparameter adjustment mechanism to achieve physical consistency reconstruction of the whole-domain temperature field and side-bonding thickness in the electrolytic cell. S2. The high-frequency harmonic amplitude features of the electrolytic cell voltage are extracted using fast Fourier transform. Combined with the reconstruction results of the physical information neural network and offline test data, a dynamically weighted global state vector is generated by a feature selection module that integrates a self-attention mechanism. S3. Construct a collaborative control engine based on deep reinforcement learning. Input the global state vector into the preset decision network and introduce an expert knowledge base constructed by fuzzy decision tree as the control boundary constraint. Based on the comprehensive reward function including voltage stability, thermal balance deviation and instantaneous carbon emission cost, solve the optimal combination command of feeding frequency and pitch adjustment. S4. The optimal combination of instructions is converted into the actions of the feeder and the anode lifting mechanism, and the abnormal response dual threshold mechanism is activated simultaneously. When the real-time operating conditions deviate from the target range or fault identification is triggered, the expert knowledge base is called to correct the strategy, so as to achieve deep collaborative control of the energy balance and material balance of the electrolytic cell.

[0026] In this embodiment of the invention, the collected parameters include real-time perceived online multi-source operating condition data and offline verification data used for model calibration, which together constitute the input features of the Physical Information Neural Network (PINN) and the state space of reinforcement learning.

[0027] Online multi-source operating condition data is mainly acquired through the physical interface of the tank control box (single tank controller) using a high-frequency sampling method, including: Electrical physical parameters: real-time tank voltage V, bus current intensity I, standard deviation of voltage fluctuation ; Frequency domain characteristic signal: specially extracted 3-10Hz high-frequency voltage noise amplitude, used to characterize the polarization degree at the bottom of the anode and the carbon emission risk; Physical field sensing data: Multi-point temperature data provided by an array of flexible temperature sensors arranged on the sides, bottom, and ends of the cathode steel rods of the tank. ; Execution status feedback: real-time on / off status of the feeder, and the current lifting / lowering stroke position of the anode bus.

[0028] Offline testing data is periodically input into the laboratory information management system and used for the correction of physical property parameters and the calculation of bias in the reinforcement learning reward function of the Physical Information Neural Network (PINN), including: Compositional analysis indicators: Molecular ratio (primary crystallization temperature, a core indicator), alumina concentration Calcium fluoride ( )content; Thermodynamic references: electrolyte temperature, aluminum level, electrolyte level, and furnace bottom pressure drop.

[0029] This invention directs physical actions through a three-tiered process of "perception-decision-execution," ensuring the simultaneous optimization of energy and material balance. The core algorithm layer runs on a host computer; the Physical Information Neural Network (PINN) first uses the collected tank shell temperature and electrical parameters to reconstruct the side-wall thickness. When the reconstruction results show that the alumina thickness deviates from the steady-state threshold, or the self-attention mechanism identifies abnormal noise in the 3-10Hz range (indicating an imbalance in alumina concentration), the reinforcement learning agent, based on the composite reward function... Find the optimal action step size in the current state, and use a penalty term to force the agent to find a balance between "stabilizing tank pressure, protecting the structure, and reducing carbon emissions".

[0030] The decision-making level determined the optimal action vector. The command route is then as follows: Material feeding command: The generated "feeding interval shortening / extending time" or "replenishment command" is sent to the cell control box (PLC / controller) via industrial Ethernet. The cell control box controls the alumina feeding valve above the electrolytic cell by driving the relay to close, achieving precise "dual-track" material compensation.

[0031] Anode Action Command: When the Physical Information Neural Network (PINN) detects insufficient thermal balance and requires increased heat input, the system outputs "anode lift stroke amount". After receiving the command, the tank control box drives the anode busbar beam lifting motor to adjust the pole pitch and compensate for the internal temperature field through the Ohmic thermal effect.

[0032] Table 1 describes the mapping relationship from data collection to action: Table 1

[0033] This invention enables deep synergy between energy balance and material balance, significantly improving current efficiency and controlling greenhouse gas emission intensity.

[0034] The architecture of this invention includes a deep virtual sensing layer based on a Physical Information Neural Network (PINN). In this embodiment, real-time heat flow data is first acquired using a finite temperature measurement array deployed outside the electrolytic cell and at the end of the cathode steel rod. To compensate for the technical problem of not being able to directly monitor the high-temperature molten material inside the cell, a Physical Information Neural Network (PINN) model is constructed. (Reference) Figure 4 As shown, the physical information neural network adopts a 6-layer fully connected residual structure, with each layer containing 128 neurons, and uses tanh as the activation function to ensure the continuity of the higher-order derivatives of the temperature field distribution; its loss function L( θ The residuals of unsteady heat transfer partial differential equations (PDEs) constrained by the law of conservation of energy are incorporated into the equation.

[0035] The loss function L( θ This is used to guide the training of neural networks, enabling predictions to approximate both observed data and physical laws. (1) in, This represents the total loss function value, which is minimized by the goal of training the neural network. This represents the data-driven loss, which is the mean square error between the measured and predicted values ​​(e.g., the mean square error between the true and predicted values ​​of the measured temperature at a temperature measurement point). This represents the physical mechanism loss, i.e., the physical residual term. The sum of squares at the point of configuration; This represents the boundary condition constraint loss, ensuring that the temperature field distribution conforms to the heat dissipation boundary conditions; This represents the weighting coefficient of each loss term, used to adjust the neural network's emphasis on data and mechanisms.

[0036] (2)

[0037] in, This is the physical residual term, which should ideally approach zero, representing the deviation in energy balance; The density of the melt (electrolyte and molten aluminum); The specific heat capacity of the melt; The global scalar temperature field output by the neural network; Use time coordinates; Thermal conductivity is the coefficient of thermal conductivity of the material (including electrolyte, carbon block, lining, etc.). For the Joule heat source term, from the formula Calculations show that the endogenous heat source representing the current generation is, among which It is the current density vector. The electric field intensity vector; This is a gradient operator used to calculate the rate of change of temperature in three-dimensional space.

[0038] The physical information neural network introduces an automatic hyperparameter adjustment mechanism. By defining an agent to explore within the hyperparameter state space, and using classification performance or physical residual convergence as the reward function, it automatically outputs the optimal convolution kernel size and network depth parameters for the current tank age condition.

[0039] Using automatic differentiation techniques, the model can reconstruct a high-resolution temperature field inside the electrolyzer in the entire spatiotemporal domain without the need for complex mesh generation. The side heat dissipation flux is constrained by introducing Robin boundary conditions. (3) in, n The direction of the normal to the boundary surface; h The comprehensive convective heat transfer coefficient reflects the heat transfer intensity between the tank shell and the environment; For ambient reference temperature; This is the derivative of temperature along the boundary normal, i.e., the heat flux gradient. The model can accurately identify the dynamic melting or growth state of the side joint thickness, thus providing physically consistent real-time state feedback for subsequent decisions.

[0040] The architecture of this invention includes a feature alignment layer for multi-source heterogeneous data and a carbon emission accounting layer. While acquiring the reconstructed temperature field, a Fast Fourier Transform (FFT) is used to perform frequency domain analysis on the electrolyzer voltage sequence to extract... High-frequency harmonic amplitude characteristics serve as key physical information characterizing the degree of micro-polarization at the bottom of the anode. These high-frequency harmonic amplitude characteristics are input into the fused carbon accounting module, and combined with the modified electrochemical equivalent formula, the instantaneous formation rate of non-effect perfluorinated carbon (PFCs) is estimated in real time. : (4) in, It is the critical voltage for local polarization, V is the instantaneous cell voltage of the electrolytic cell, and I is the electrolytic current. α、βThese are empirical coefficients related to the electrolytic cell type. Through this step, the present invention transforms environmental information that is originally difficult to quantify into state data that can be perceived in real time by the control system. By co-mapping the instantaneous generation rate with the temperature field characteristics and side-slab thickness characteristics output by the physical information neural network, a global state vector covering the dimensions of electric field, thermal field, and carbon emissions is constructed. .

[0041] In this embodiment of the invention, the self-attention mechanism is implemented as follows: the correlation weight matrix between pairs of features is calculated on the feature dimension, and the features are weighted and fused after normalization to enhance the model's perception accuracy of weak polarization signals in the 3-10Hz range and avoid interference from industrial environmental noise.

[0042] The architecture of this invention includes a collaborative decision-making layer based on multi-constraint reinforcement learning. The decision network employs an Actor-Critic reinforcement learning network with an experience replay mechanism to find the optimal combination of alumina feeding frequency and amplitude adjustment. This invention designs a comprehensive reward function. Its mathematical definition integrates voltage stability, thermal equilibrium stability, and emission reduction costs: (5) in, It is the value of the comprehensive reward function, which the agent seeks to maximize; due to the negative sign, it means minimizing the penalty. It is the voltage fluctuation variance of the electrolytic cell, reflecting the thermal balance stability of the electrolytic cell; This refers to the actual perceived concentration of alumina. It is the midpoint of the set target alumina concentration range; It is instantaneous carbon emission intensity, based on the instantaneous formation rate of non-effect perfluorinated carbon. Estimated environmental costs; It is the reconstructed deviation of the side joint thickness, that is, the deviation of the side joint thickness from the ideal steady state; It is the weighted penalty coefficient of the above evaluation indicators.

[0043] Under this framework, carbon emission penalty weights Instead of a fixed value, the alumina concentration is dynamically adjusted nonlinearly based on whether it approaches the low-fluorine zone. When the physical information neural network detects an abnormally thinning sidewall thickness and voltage noise reflecting a high carbon emission risk, the penalty term in the comprehensive reward function will increase sharply, guiding the agent to solve for the preventative refueling command through the Bellman equation. (6) in, Q ( s, a ) is in state s Take action belowa The value of the action; R It is the immediate reward obtained after performing an action, that is ; This is a discount factor used to balance current benefits with long-term expected returns; s ', a 'This represents the state and possible actions at the next moment.'

[0044] This mechanism ensures that the control strategy can guide the electrolyzer away from the high-emission risk zone before the alumina concentration imbalance causes an effect.

[0045] The architecture of this invention includes a closed-loop feedback optimization and a technology layer. The optimal combination command is sent to the tank control box via industrial Ethernet or Modbus protocol. The tank control box, as the underlying command hub, drives a relay or frequency converter after receiving the optimal combination command to achieve precise pulse control of the alumina feeder's feeding frequency and logical drive of the anode lifting mechanism motor (anode bus beam motor).

[0046] In this embodiment of the invention, the final generated control commands (optimal combination commands) are implemented by the PLC actuator to the alumina feeder and anode lifting system. The system monitors the voltage response of the electrolytic cell after adjustment and the changes in the thermal field (temperature field) reconstructed by PINN in real time, and uses online measured data to periodically fine-tune the hyperparameters of the model to adapt to the parameter drift caused by the increase of the electrolytic cell age. The implementation of this invention can significantly reduce the equivalent carbon dioxide emissions per ton of aluminum while maintaining a high current efficiency of over 93%, and effectively reduce the sedimentation at the bottom of the cell caused by thermal imbalance, thus extending the service life of the electrolytic cell and achieving a deep synergy between industrial production benefits and environmental protection goals.

[0047] In a more specific embodiment, a series of 500kA large prebaked anode electrolytic cells are used as the application object to demonstrate the technical superiority of the present invention under complex working conditions.

[0048] The system-defined input data layer consists of online monitoring data and offline testing data. Online monitoring data: Real-time acquisition of cell voltage V, electrolytic current I, cell voltage high-frequency noise, anode current distribution, and bus stroke control parameters via distributed sensors. Offline testing data: Key process indicators such as electrolyte temperature, molecular ratio, calcium fluoride content, aluminum level, and electrolyte level, periodically input by the laboratory information management system. Output control commands: Decision vectors generated by the system. This includes the alumina feeding cycle (seconds), single feeding amount (kg), aluminum fluoride addition dosage, and anode lifting stroke command.

[0049] To address sensor jitter or input errors in the raw data, this invention introduces an outlier removal mechanism based on the interquartile range (IQR) method: calculating the 25th percentile Q1 and 75th percentile Q3 for various operating conditions to determine the interquartile range. Set the judgment interval. The isolated points outside the interval are removed, and the missing time series are filled in using a linear interpolation algorithm.

[0050] Implementation of Virtual Sensing and Thermal Field Reconstruction Based on PINN. After acquiring cleaned data, the system enters the internal state reconstruction stage based on a Physical Information Neural Network (PINN). Model Input: Define the three-dimensional spatial coordinates (x, y, z) of the area to be monitored and the time step t. Physical Constraint Fusion: During the training iteration of the neural network, its loss function explicitly incorporates the residual terms of the partial differential equation (PDE) for heat transfer constrained by the law of energy conservation. The gradient of the temperature field with respect to coordinates is calculated using automatic differentiation, forcing the model to adhere to Fourier's law of thermal conduction even in the sensor-deficient electrolyte interior region. Parameter adaptive adjustment: A deep reinforcement learning-based network hyperparameter automatic adjustment mechanism is introduced. The agent dynamically optimizes the hidden layer depth and number of neurons in the physical information neural network based on the convergence speed of the physical residuals, ensuring the model exhibits extremely high robustness under unsteady conditions such as polarization fluctuations.

[0051] Feature dynamic weighting and carbon footprint accounting implementation. To extract weak carbon emission early warning features from complex electrical signals, this invention performs multimodal feature enhancement. Frequency domain analysis: Fast Fourier Transform (FFT) is used to separate the 3-10Hz harmonic components from the voltage signal. These components are related to the bubble removal kinetics at the bottom of the anode and the non-effect perfluorocarbon (PFC) formation rate. It exhibits a strong correlation; when the alumina concentration approaches the critical under-fluorine state but has not yet triggered a significant voltage surge, the energy fluctuations in this frequency band will be significantly enhanced. Self-attention mechanism weighting: The system receives feature vectors generated by the gradient boosting decision tree and calculates the correlation weight matrix along the feature dimensions through a self-attention layer. For high-frequency noise, reconstructed temperature gradient, and molecular ratio deviation, the system dynamically allocates weighting coefficients, thereby ensuring prediction accuracy while enabling the model to prioritize capturing the critical polarization signal that induces the anode effect.

[0052] A closed-loop decision-making process driven by both knowledge and data is implemented. The decision network operates based on a deep reinforcement learning architecture, aiming to solve for the optimal combination of instructions (i.e., the optimal action sequence). Integrating the global temperature field reconstructed by PINN, lateral strut thickness, measured trench pressure noise, and instantaneous carbon emission intensity indicators, a comprehensive reward function is designed that includes physical consistency constraints and environmental costs. When the system detects that the tank conditions are cooling down and the risk of carbon emissions is increasing, it guides the agent to perform adjustments by strengthening penalties. That is, while appropriately lengthening the feeding interval to restore heat recovery, it suppresses the generation of PFCs by fine-tuning the electrode spacing. Expert rule correction: An expert knowledge base constructed by fuzzy decision trees is introduced as a safe zone for the decision layer to ensure that the instructions generated by the model do not exceed the safe operating limits of the aluminum electrolysis process.

[0053] Taking a series of 500kA large prebaked electrolytic cells as an example. During implementation, when the alumina concentration decreased from 2.3% to 1.7%, the self-attention mechanism module detected a surge in the weighting coefficient of the 3-10Hz harmonic amplitude in the voltage sequence (weighting increased from 0.12 to 0.45), indicating a risk of polarization at the bottom of the anode. Simultaneously, the PINN reconstruction module sensed that the side junction thickness deviated from the steady-state threshold by 8%. Traditional control logic often maintains the original state due to lag in resistance slope feedback. After applying the method of this invention, the PINN reconstruction module sensed a potential melting and damage trend in the side junction, and the self-attention mechanism module identified an abnormal high-frequency harmonic weight. The reinforcement learning agent then received an instruction driven by the carbon emission penalty term, proactively adjusting the feeding interval from 120 seconds to 85 seconds and issuing an incremental instruction for aluminum fluoride. Production data over 60 consecutive days showed that this method increased the molecular ratio qualification rate to 93.5%, reduced the average CO2 emissions per ton of aluminum by 250 kg, and significantly enhanced the carbon competitiveness and thermal stability of the production line.

[0054] In an exemplary embodiment, the present invention also provides an electronic device, the electronic device comprising: processor; The memory stores computer-readable instructions, which, when loaded and executed by the processor, implement the steps of the electrolytic cell collaborative control method described above, which integrates mechanism reconstruction and knowledge decision-driven control.

[0055] In an exemplary embodiment, the present invention also provides a computer-readable storage medium storing at least one instruction, which is loaded and executed by a processor to implement the steps of the electrolytic cell collaborative control method described above, which integrates mechanism reconstruction and knowledge decision-driven operation. For example, the computer-readable storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, or optical data storage device.

[0056] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Unless otherwise specified, an element defined by the phrase "comprising..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0057] The use of terms such as "an embodiment," "an embodiment," "an exemplary embodiment," and "some embodiments" in the specification indicates that the described embodiment may include a specific feature, structure, or characteristic, but not every embodiment necessarily includes that specific feature, structure, or characteristic. Furthermore, when a specific feature, structure, or characteristic is described in connection with an embodiment, implementing such a feature, structure, or characteristic in conjunction with other embodiments (whether explicitly described or not) should be within the knowledge of those skilled in the art.

[0058] It should be understood that, in various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0059] In the embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0060] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0061] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0062] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0063] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0064] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for coordinated control of an electrolyzer that integrates mechanism reconstruction and knowledge-driven decision-making, characterized in that, Includes the following steps: S1. Collect multi-source operating data of the electrolytic cell using distributed sensors, construct a physical information neural network, incorporate the residual term of the heat transfer partial differential equation into the loss function, and introduce an automatic hyperparameter adjustment mechanism to achieve physical consistency reconstruction of the whole-domain temperature field and side-bonding thickness in the electrolytic cell. S2. The high-frequency harmonic amplitude features of the electrolytic cell voltage are extracted using fast Fourier transform. Combined with the reconstruction results of the physical information neural network and offline test data, a dynamically weighted global state vector is generated by a feature selection module that integrates a self-attention mechanism. S3. Construct a collaborative control engine based on deep reinforcement learning. Input the global state vector into the preset decision network and introduce an expert knowledge base constructed by fuzzy decision tree as the control boundary constraint. Based on the comprehensive reward function including voltage stability, thermal balance deviation and instantaneous carbon emission cost, solve the optimal combination command of feeding frequency and pitch adjustment. S4. The optimal combination of instructions is converted into the actions of the feeder and the anode lifting mechanism, and the abnormal response dual threshold mechanism is activated simultaneously. When the real-time operating conditions deviate from the target range or fault identification is triggered, the expert knowledge base is called to correct the strategy, so as to achieve deep collaborative control of the energy balance and material balance of the electrolytic cell.

2. The method according to claim 1, characterized in that, Online multi-source operating condition data is acquired via the physical interface of the tank control box using a high-frequency sampling method, including: Electrical physical parameters: real-time tank voltage, bus current intensity, and voltage fluctuation standard deviation; Frequency domain characteristic signal: 3-10Hz high-frequency voltage noise amplitude, used to characterize the polarization degree at the bottom of the anode and the carbon emission risk; Physical field sensing data: Multi-point temperature data provided by a flexible temperature sensor array arranged on the side, bottom, and end of the cathode steel rod of the tank shell; Execution status feedback: real-time on / off status of the feeder, and current lifting / lowering stroke position of the anode busbar; Offline test data is periodically input into the laboratory information management system and used for parameter correction of the physical information neural network and deviation calculation of the reinforcement learning reward function, including: Component analysis indicators: molecular ratio, alumina concentration, calcium fluoride content; Thermodynamic references: electrolyte temperature, aluminum level, electrolyte level, and furnace bottom pressure drop.

3. The method according to claim 1, characterized in that, The physical information neural network adopts a 6-layer fully connected residual structure, with each layer containing 128 neurons, and uses tanh as the activation function to ensure the continuity of the higher-order derivatives of the temperature field distribution; its loss function L( θ The residual terms of the unsteady heat transfer partial differential equations constrained by the law of conservation of energy are incorporated into the equations. The loss function L( θ This is used to guide the training of neural networks, enabling predictions to approximate both observed data and physical laws. (1) in, This represents the total loss function value, which is minimized by the goal of training the neural network. This represents the data-driven loss, which is the mean square error between the measured and predicted values. This represents the physical mechanism loss, i.e., the physical residual term. The sum of squares at the point of configuration; This represents the boundary condition constraint loss, ensuring that the temperature field distribution conforms to the heat dissipation boundary conditions; This represents the weighting coefficient of each loss term, used to adjust the neural network's emphasis on data and mechanisms; (2) in, This is the physical residual term, which should ideally approach zero, representing the deviation in energy balance; The density of the melt; The specific heat capacity of the melt; The global scalar temperature field output by the neural network; Use time coordinates; The thermal conductivity of the material; For the Joule heat source term, from the formula Calculations show that the endogenous heat source representing the current generation is, among which It is the current density vector. The electric field intensity vector; This is a gradient operator used to calculate the rate of change of temperature in three-dimensional space.

4. The method according to claim 1, characterized in that, The hyperparameter automatic adjustment mechanism defines an agent to explore within the hyperparameter state space, using classification performance or physical residual convergence as the reward function, and automatically outputs the optimal convolution kernel size and network depth parameters for the current tank age condition.

5. The method according to claim 1, characterized in that, Using automatic differentiation technology, the temperature field inside the electrolytic cell is reconstructed in the entire spatiotemporal domain without the need for complex mesh generation; the side heat dissipation flux is constrained by introducing Robin boundary conditions. (3) in, n The direction of the normal to the boundary surface; h The comprehensive convective heat transfer coefficient reflects the heat transfer intensity between the tank shell and the environment; For ambient reference temperature; This is the derivative of temperature along the boundary normal direction, i.e., the heat flux gradient.

6. The method according to claim 1, characterized in that, While acquiring the reconstructed temperature field, the electrolytic cell voltage sequence was analyzed in the frequency domain using Fast Fourier Transform to extract... High-frequency harmonic amplitude characteristics serve as key physical information characterizing the degree of micro-polarization at the bottom of the anode. These characteristics are input into the fusion carbon accounting module, and combined with the modified electrochemical equivalent formula, the instantaneous formation rate of non-effect perfluorinated carbon is estimated in real time. : (4) in, It is the critical voltage for local polarization, V is the instantaneous cell voltage of the electrolytic cell, and I is the electrolytic current. α、β These are empirical coefficients related to the electrolytic cell type; the instantaneous generation rate is co-mapped with the temperature field characteristics and side-slab thickness characteristics output by the physical information neural network to construct a global state vector covering the dimensions of electric field, thermal field, and carbon emissions. .

7. The method according to claim 1, characterized in that, The self-attention mechanism is implemented as follows: the correlation weight matrix between pairs of features is calculated on the feature dimension, and the features are weighted and fused after normalization to enhance the model's perception accuracy of weak polarization signals in the 3-10Hz range and avoid interference from industrial environmental noise.

8. The method according to claim 1, characterized in that, The decision network employs an Actor-Critic reinforcement learning network with an experience replay mechanism to find the optimal combination of alumina feeding frequency and amplitude adjustment; a comprehensive reward function is designed. Its mathematical definition integrates voltage stability, thermal equilibrium stability, and emission reduction costs: (5) in, It is the value of the comprehensive reward function, which the agent seeks to maximize; due to the negative sign, it means minimizing the penalty. It is the voltage fluctuation variance of the electrolytic cell, reflecting the thermal balance stability of the electrolytic cell; This refers to the actual perceived concentration of alumina. It is the midpoint of the set target alumina concentration range; It is instantaneous carbon emission intensity, based on the instantaneous formation rate of non-effect perfluorinated carbon. Estimated environmental costs; It is the reconstructed deviation of the side joint thickness, that is, the deviation of the side joint thickness from the ideal steady state; It is the weighted penalty coefficient of the above evaluation indicators.

9. The method according to claim 8, characterized in that, Carbon emission penalty weight Instead of a fixed value, the alumina concentration is dynamically adjusted nonlinearly based on whether it approaches the low-fluorine zone. When the physical information neural network detects an abnormally thinning sidewall thickness and voltage noise reflecting a high risk of carbon emissions, the penalty term in the comprehensive reward function will increase sharply, guiding the agent to solve for the preventive feeding instruction through the Bellman equation. (6) in, Q ( s, a ) is in state s Take action below a The value of the action; R It is the immediate reward obtained after performing an action, that is ; This is a discount factor used to balance current benefits with long-term expected returns; s ', a 'This represents the state and possible actions at the next moment.' 10. The method according to claim 1, characterized in that, The optimal combination command is sent to the tank control box via industrial Ethernet or Modbus protocol. The tank control box, as the underlying command hub, drives the relay or frequency converter after receiving the optimal combination command to realize pulse control of the feeding frequency of the alumina feeder and logic drive of the motor of the anode lifting mechanism.

Images