A Method and System for Controlling Electrode Spacing in Aluminum Electrolytic Cells by Integrating Time-Series and Mechanistic Models

By integrating time-series and mechanistic models into an aluminum electrolytic cell electrode distance control method, the problems of model illusion and insufficient knowledge utilization in the aluminum electrolytic cell control system are solved. This method enables multi-source knowledge collaborative decision-making, improves the physical consistency of prediction results and the robustness of the system.

CN123063818APending Publication Date: 2026-11-10HUNAN SHENGXIANG CONTROLS
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
CN202611170435.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-04
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing aluminum electrolysis cell control systems suffer from problems such as model illusion, insufficient utilization of expert knowledge, lack of physical consistency in prediction results, and lack of collaborative decision-making capabilities through multi-source knowledge fusion in electrode spacing adjustment.

Method used

A method for controlling the electrode spacing of aluminum electrolyzers by integrating time-series and mechanistic models is proposed. This method achieves collaborative decision-making based on multi-source knowledge by constructing an expert knowledge base for aluminum electrolysis processes, an electrolyte mechanism model, and an enhanced PatchTST multi-scale time-series coding network, combined with knowledge distillation and physical regularization loss terms.

Benefits of technology

It significantly improves the physical consistency of predictions, breaks through the deep integration of expert knowledge and deep learning, constructs a full-link intelligent control system, and enhances the interpretability of prediction results and the robustness of the system.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123063818A_ABST
    Figure CN123063818A_ABST
Patent Text Reader

Abstract

This invention relates to the field of intelligent control technology in the aluminum electrolysis industry, and provides a method and system for controlling the electrode spacing of aluminum electrolysis cells by integrating a time-series model and a mechanistic model. The method includes: acquiring electrolysis cell operating parameter data collected by multi-source sensors; constructing an expert knowledge base for aluminum electrolysis processes; constructing an electrolyte mechanistic model; constructing and using an enhanced PatchTST multi-scale time-series coding network to perform time-series coding and dynamic prediction of the electrolysis cell operating parameter data; dynamically identifying the cell condition and predicting the anode effect trend based on knowledge graph feature injection; calculating the electrode spacing compensation amount according to the enhanced PatchTST multi-scale time-series coding network and generating a collaborative control instruction set to be sent to the actuator. This method improves the physical consistency of the prediction results and solves the problems of lag in response to electrode spacing adjustments and changes in cell condition, and the difficulty in adaptively updating the control strategy.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of intelligent control technology in the aluminum electrolysis industry, and in particular to a method and system for controlling the electrode spacing of aluminum electrolysis cells by integrating a time-series model and a mechanism model. Background Technology

[0002] Aluminum electrolysis is a typical complex industrial process characterized by nonlinearity, strong coupling, and large time delay. Electrochemical reactions, material transport, and heat exchange occur simultaneously within the electrolytic cell, with various physical fields coupled together, posing significant challenges to process control. Currently, mainstream aluminum electrolysis cell control systems primarily rely on traditional cell controllers to perform basic material feeding control and voltage regulation functions. However, existing technologies still suffer from the following prominent problems: 1. Purely data-driven models lack physical constraints and suffer from the "model illusion" problem. In recent years, significant progress has been made in deep learning-based aluminum electrolysis cell control methods. However, purely data-driven time-series prediction models have revealed a core pain point in practical industrial applications—"model illusion." Because deep learning models learn statistical correlations only from historical data and lack an understanding of the physical mechanisms of aluminum electrolysis, when operating conditions deviate from the training data distribution, the model may generate predictions that contradict physical laws. For example, the superheat prediction curve may violate the first law of thermodynamics, and the cell resistance prediction sequence may violate energy conservation constraints. This "model illusion" severely undermines industrial users' trust in AI control systems and limits their application in key process stages. Although some scholars have developed data-knowledge-driven decision support systems for aluminum electrolysis processes, integrating prediction models, multi-objective optimizers, and knowledge-guided decision modules, their knowledge bases are mainly used in the decision-making stage rather than the constraint injection stage of model training, failing to fundamentally suppress model illusion.

[0003] 2. Insufficient utilization of expert knowledge and inadequate depth of knowledge-data integration. Over decades of development, the aluminum electrolysis industry has accumulated a wealth of process operation experience, expert rule bases, and mechanistic knowledge. For example, while the patent "Method and Equipment for Obtaining Process Control Parameters of Aluminum Electrolysis Cells Based on Data and Knowledge" (patent number CN108360021B) authorized by the Central South University team involves knowledge-driven thinking, expert knowledge is only used as an auxiliary reference and is not deeply integrated with the training process of deep learning models. Existing research combining knowledge distillation, knowledge graphs, and time-series prediction models is mostly focused on load forecasting and equipment fault diagnosis, and has not yet formed a knowledge injection method deeply coupled with multi-scale time-series coding networks in the field of aluminum electrolysis cell control.

[0004] 3. The prediction results lack physical consistency constraints and have poor interpretability. The contradiction between the complexity of the aluminum electrolysis cell's operating mechanism and the limitations of data sampling often results in a lack of physical consistency in data-driven predictions. The predicted cell resistance and electrolyte superheat should satisfy an energy conservation relationship, and the alumina concentration variation trend should conform to the material balance principle. Existing control systems fail to embed these physical constraints into the prediction model, leading to a lack of engineering reliability in the prediction results. In recent years, the "data + mechanism + experience" three-drive system proposed by the industry has achieved initial results in the aluminum electrolysis field, but a systematic integrated prediction-decision-making patent solution has yet to be formed.

[0005] 4. Polar distance adjustment lacks collaborative decision-making capabilities through multi-source knowledge fusion. Electrode spacing is a key process parameter that determines cell voltage and electrolysis energy consumption. In existing technologies, electrode spacing adjustment mainly relies on manual experience or simple linear superposition based on the output of deep learning prediction models. There is a lack of a multi-source knowledge collaborative decision-making mechanism that can simultaneously integrate electrochemical mechanism models, process expert rule bases, and historical optimization case knowledge bases.

[0006] Therefore, there is an urgent need for an intelligent cell condition collaborative control method and system that integrates an expert knowledge base of aluminum electrolysis process, an electrolyte mechanism model, and a time-series coding neural network. Summary of the Invention

[0007] To overcome the problems existing in the control systems of aluminum electrolytic cells in the prior art, this invention provides a method for controlling the electrode spacing of aluminum electrolytic cells that integrates timing models and mechanistic models. The method includes: The process involves acquiring electrolytic cell operating parameter data from multi-source sensors, constructing an aluminum electrolysis process expert knowledge base based on process rules, safety constraints, process expert experience, and historical best practices, and building an electrolyte mechanism model based on the Hall-Herrouhl electrolysis reaction principle and material-heat balance theory. An enhanced PatchTST multi-scale temporal coding network is then constructed and used to perform temporal coding and dynamic prediction of the electrolytic cell operating parameter data, outputting prediction deviations for alumina concentration drift trends, electrolyte superheat, cell resistance, anode effect probability, and electrode distance. The time-series coding network includes a time-series data patch partitioning module, a knowledge distillation time-frequency feature extraction module, a channel-independent embedding layer, a multi-scale patch coding layer, a physical information constraint fusion module, and a prediction output layer. The time-series data patch partitioning module divides the electrolytic cell operating parameter data collected by multiple sources of sensors into multiple time-series segments. The knowledge distillation time-frequency feature extraction module simultaneously performs time-domain feature extraction and frequency-domain feature extraction on the multiple time-series segments, and uses process rules from the aluminum electrolysis process expert knowledge base as teacher knowledge to perform knowledge distillation on the frequency-domain feature extraction process. The channel-independent... The embedding layer processes the extracted features using a channel-independent mechanism, outputting multi-channel patch embedding vectors. The multi-scale patch encoding layer inputs the patch embedding vectors from different channels into independent encoder branches, capturing the temporal dependencies within each channel's patch embedding vector. Then, a cross-scale feature interaction layer is established between the encoder branches, outputting a multi-scale joint feature representation. The physical information constraint fusion module uses the analytical values ​​of the electrolyte mechanism model as auxiliary input features and fuses them with the multi-scale joint feature representation. The model is trained using a loss function with added physical regularization loss term. The prediction layer flattens the fused features and generates the final prediction result through the prediction head; the aluminum electrolysis process expert knowledge base is transformed into a structured knowledge graph, and then the embedding vector of the corresponding state node in the knowledge graph is fused with the multi-scale joint feature representation to obtain the process expert knowledge base influence factor; the electrode distance compensation amount is calculated based on the alumina concentration drift trend prediction value, electrolyte superheat prediction value, cell resistance prediction value, anode effect probability prediction value, electrode distance prediction value, and process expert knowledge base influence factor output by the enhanced PatchTST multi-scale temporal coding network, and a collaborative control instruction set is generated and sent to the actuator.

[0008] Furthermore, the formula for calculating the pole gap compensation is as follows: Δd = k1·ΔC_Pred + k2·ΔT_Pred + k3·E_vib + k4·P_AE + k5·K_KB +k6·D_physics, Wherein, Δd is the electrode distance compensation amount, ΔC_Pred is the prediction deviation of alumina concentration drift trend, ΔT_Pred is the prediction deviation of electrolyte superheat, E_vib is the predicted cell resistance value, P_AE is the predicted anode effect probability value, K_KB is the process expert knowledge base influence factor, D_physics is the deviation correction amount between the electrode distance compensation mechanism reference value calculated by the electrolyte mechanism model and the electrode distance prediction value, and k1, k2, k3, k4, k5, and k6 are weighting coefficients.

[0009] Furthermore, the time-domain feature extraction involves extracting local waveform features through a multi-layer convolutional network, while the frequency-domain feature extraction uses a short-time Fourier transform to convert the time-domain signal to the frequency domain and extract the energy distribution features of the slot resistance signal in different frequency bands.

[0010] Furthermore, the process rules in the aluminum electrolysis process expert knowledge base are used as teacher knowledge to perform knowledge distillation on the frequency domain feature extraction process. Specifically, this includes: using the process rules in the aluminum electrolysis process expert knowledge base as teacher knowledge, using each independent encoder branch as a student model, and introducing a knowledge distillation loss function to guide the time-frequency feature extraction. The teacher knowledge includes: reference distribution curves of energy in each frequency band under normal cell conditions, typical frequency domain evolution patterns of cell resistance before the occurrence of the anode effect, and characteristic patterns of cell resistance spectrum corresponding to different alumina concentration ranges.

[0011] Furthermore, the physical regularization loss function is specifically as follows: L_physics = λ1·||ΔC_Pred - ΔC_Mech||² + λ2·||ΔT_Pred - ΔT_Mech||² + λ3·||U_Pred - U_Mech||², Wherein, ΔC_Pred is the prediction deviation of alumina concentration drift trend, ΔC_Mech is the deviation of theoretical change in alumina concentration calculated by the electrolyte mechanism model, ΔT_Pred is the prediction deviation of electrolyte superheat, ΔT_Mech is the reference value of electrolyte superheat calculated by the electrolyte mechanism model, U_Pred is the predicted value of cell voltage, U_Mech is the reference value of cell voltage calculated by the electrolyte mechanism model, and λ1, λ2, and λ3 are weighting coefficients.

[0012] Furthermore, the actuator adopts a dual-track alumina feeding control strategy, including: the first track is a fixed-cycle baseline feeding to maintain the basic material supply of the electrolytic cell; the second track is a dynamically adjusted feeding, which performs independent feeding compensation for local areas based on the real-time alumina concentration deviation of each control area collected by the sensor, the alumina concentration drift trend prediction value output by the enhanced PatchTST multi-scale time-series coding network, and the optimal feeding strategy for the corresponding concentration range in the process rules.

[0013] Furthermore, while the enhanced PatchTST multi-scale temporal coding network outputs the prediction results, it also provides a mechanistic explanation of the prediction results.

[0014] Furthermore, it also includes receiving the tank condition response data after execution, and updating the prediction parameters of the enhanced PatchTST multi-scale temporal coding network and the aluminum electrolysis process expert knowledge base through a knowledge-data collaborative self-learning module. Furthermore, the operating parameter data includes: electrical parameter data, thermodynamic parameter data, and tank condition characteristic parameter data.

[0015] On the other hand, the present invention also provides an aluminum electrolytic cell electrode spacing control system that integrates a timing model and a mechanism model, for implementing the above-mentioned aluminum electrolytic cell electrode spacing control method, specifically including: The multi-source data acquisition unit is used to acquire electrolytic cell operating parameter data collected by multi-source sensors in the aluminum electrolytic cell.

[0016] The expert knowledge base construction unit includes a process rule base module, a safety constraint base module, a process expert experience base module, and a historical best case base module. It is used to store and reason about domain knowledge in the field of aluminum electrolysis cell control, and to provide prior guidance and boundary constraints for model training and control decisions.

[0017] The electrolyte mechanism model unit includes a material balance mechanism module, a thermal balance mechanism module, and a combined electrode distance-voltage-thermal balance mechanism module, which are used to perform analytical calculations on the key state variables of the aluminum electrolysis cell.

[0018] The enhanced PatchTST cell condition dynamic identification unit includes a time-series data patch partitioning module, a knowledge distillation time-frequency feature extraction module, a channel independent embedding module, a multi-scale encoding module, a physical information constraint fusion module, and a prediction output layer module. It performs time-series encoding and dynamic prediction on electrolytic cell operating parameter data, and outputs prediction deviations for alumina concentration drift trend, electrolyte superheat, cell resistance, anode effect probability, and electrode distance, enabling early perception of cell condition evolution.

[0019] The knowledge graph feature injection unit includes a knowledge graph construction module, a feature fusion module, and a mechanism explanation output module. The knowledge graph construction module is used to transform the aluminum electrolysis process expert knowledge base into a structured knowledge graph. The feature fusion module is used to perform attention fusion between the embedding vectors of the corresponding state nodes in the knowledge graph and the multi-scale joint feature representation to obtain the process expert knowledge base influence factor. The mechanism explanation output module is used to provide a mechanism explanation of the prediction results while the enhanced PatchTST multi-scale temporal coding network outputs the prediction results.

[0020] The extreme distance adaptive adjustment unit includes a prediction-knowledge-mechanism fusion decision module and an extreme distance actuator control module. The prediction-knowledge-mechanism fusion decision module takes the prediction results output by the enhanced PatchTST multi-scale time-series coding network, the best practice guidance provided by the process expert knowledge base, and the physical constraint benchmark provided by the electrolyte mechanism model as multi-source inputs, and generates an extreme distance control strategy through fusion. The extreme distance actuator control module is used to calculate the extreme distance compensation amount and generate a collaborative control instruction set to be issued to the actuator.

[0021] The material feeding coordination control unit is used to execute the dual-track alumina feeding control strategy.

[0022] The collaborative self-learning update unit includes a self-supervised pre-training module, an online incremental fine-tuning module, a reinforcement learning weight optimization module, and a knowledge base incremental update module. It is used to continuously learn new data, optimize model parameters, and update rule knowledge during system operation.

[0023] The central control unit connects all units, unifies the management and control timing and command priority, and ensures the coordinated and orderly operation of multiple units.

[0024] The aluminum electrolytic cell electrode spacing control method and system that integrates time-series and mechanistic models constructed in this invention have the following significant technical advantages and beneficial effects: (1) It fundamentally suppresses “model illusion” and significantly improves the physical consistency of predictions. This invention embeds an expert knowledge base on aluminum electrolysis processes and an electrolyte mechanism model into the entire PatchTST temporal coding network. The knowledge distillation time-frequency feature extraction module ensures the model learns frequency domain features with clear physical meaning. The physical information constraint fusion module applies physical constraints to the network output through a dual mechanism of physical regularization loss term and mechanism knowledge feedforward injection, fundamentally solving the "model illusion" problem that easily arises in complex industrial scenarios where purely data-driven models violate physical laws. The prediction results of this invention possess verified physical consistency.

[0025] (2) Deep integration of expert knowledge and deep learning to break through the bottleneck of shallow knowledge utilization. This invention seamlessly integrates process rules, safety constraints, expert experience, and best practices from an aluminum electrolysis process expert knowledge base—integrating them at different depths and in different formats: teacher knowledge during knowledge distillation, feature injection points from the knowledge graph, and decision-making references from historical best practices—into the entire process of deep learning training, prediction, and decision-making. Compared to existing technologies where knowledge bases are only used as auxiliary references in the decision-making stage (such as CN108360021B "A Method for Obtaining Process Control Parameters of Aluminum Electrolysis Cells Based on Data and Knowledge"), the knowledge-data fusion of this invention is achieved at the core training layer of the model rather than the surface decision layer, representing a fundamental breakthrough in the depth of knowledge injection and the quality of fusion.

[0026] (3) Construct a three-in-one intelligent control system integrating data, mechanism and knowledge across the entire chain. The proposed extreme distance compensation formula incorporates six dimensions of decision variables, achieving for the first time in the field of aluminum electrolysis cell control a deep integration and collaborative decision-making of data-driven (predictive output), mechanism-driven (material / thermal balance equations), and knowledge-driven (expert rule base + excellent case library). This three-drive fusion design allows the control system to still rely on the mechanism model to provide safe operating boundaries and the expert knowledge base to provide rule-based response strategies when encountering extreme cell conditions not covered by training data. The system's robustness and adaptability to operating conditions are significantly superior to purely data-driven solutions. Extreme distance adjustment has transitioned from "post-event correction" to a multi-dimensional collaborative optimization of "advance prediction + physical constraints + knowledge guidance."

[0027] (4) The interpretability of the prediction results is substantially improved. The mechanistic explanation output module, combined with the causal reasoning chain of the knowledge graph, enables the system to clearly explain to operators "which observational features, which knowledge rules, and with what probability predict what tank condition evolution trend." This innovation fundamentally solves the trust bottleneck of the "black box" problem of traditional deep learning models, which makes large-scale deployment difficult in the high-risk and critical industrial process of aluminum electrolysis. The organic connection of the expert knowledge base enables the control system to not only know "what," but also "why."

[0028] (5) Continuous evolution and knowledge accumulation of expert knowledge base The incremental update module of the expert knowledge base enables process knowledge to be continuously accumulated, verified, and refined as the system operates. In the initial stages of industrial deployment, an initial expert knowledge base can be relied upon. With the continuous accumulation of excellent case studies, the coverage and reasoning capabilities of the knowledge base constantly improve, and the control system exhibits a self-evolving characteristic of "getting smarter with use." This is of significant importance for preserving and increasing the value of knowledge assets in aluminum electrolysis production enterprises with high personnel turnover. Attached Figure Description

[0029] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 This is an overall flowchart of an aluminum electrolytic cell electrode spacing control method that integrates timing and mechanistic models according to an embodiment of the present invention.

[0031] Figure 2 This is a schematic diagram of the hierarchical structure of the aluminum electrolysis process expert knowledge base in one embodiment of the present invention.

[0032] Figure 3 This is a block diagram of the feeding concentration determination mechanism model in one embodiment of the present invention.

[0033] Figure 4 This is a schematic diagram of the structure of the enhanced PatchTST multi-scale temporal coding network in one embodiment of the present invention.

[0034] Figure 5 This is a schematic diagram of the knowledge graph feature injection process in one embodiment of the present invention.

[0035] Figure 6 This is a schematic diagram of the prediction-knowledge-mechanism fusion decision-making framework in one embodiment of the present invention.

[0036] Figure 7 This is a schematic diagram of the structure of a self-learning module based on knowledge-data collaboration in one embodiment of the present invention.

[0037] Figure 8 This is a flowchart illustrating the calculation of the physical regularization loss function in one embodiment of the present invention.

[0038] Figure 9 This is a structural block diagram of an aluminum electrolytic cell electrode spacing control system that integrates a timing model and a mechanism model according to an embodiment of the present invention. Detailed Implementation

[0039] To facilitate understanding of this application, the following description will be more comprehensive and detailed in conjunction with the accompanying drawings and preferred embodiments, but the scope of protection of this application is not limited to the following specific embodiments.

[0040] Unless otherwise defined, all technical terms used herein have the same meaning as commonly understood by those skilled in the art. The technical terms used herein are for the purpose of describing particular embodiments only and are not intended to limit the scope of this application.

[0041] This embodiment uses a 500kA large prebaked anode aluminum electrolytic cell as the application object to implement the aluminum electrolytic cell electrode spacing control method and system of the present invention, which integrates timing model and mechanism model. The following is in conjunction with... Figure 1 The specific implementation methods of the embodiments of the present invention will be further described below.

[0042] like Figure 1 As shown, the aluminum electrolytic cell electrode spacing control method in this embodiment includes multi-source data acquisition, expert knowledge base construction, electrolyte mechanism model deployment, and enhanced PatchTST network construction and training, as detailed below: S1: Real-time acquisition of multi-source data. By deploying a multi-source sensor array in the aluminum electrolysis cell, the operating parameters of the electrolysis cell are acquired in real time with a millisecond-level sampling period, resulting in the electrolysis cell operating parameter data collected by the multi-source sensors.

[0043] The multi-source sensor array mainly includes a high-precision cell voltage sensor, a fiber optic current sensor, a thermocouple temperature sensor, and an anode current distribution monitoring device. The high-precision cell voltage sensor and fiber optic current sensor are deployed at the short-circuit port of the aluminum electrolysis cell. The voltage drop measurement device is arranged on each anode guide rod. Twelve thermocouple temperature measurement points are evenly arranged along the longitudinal direction of the electrolysis cell. The entire cell is divided into six spatial control zones, each containing four pairs of anodes and equipped with an independent feeding actuator. The data sampling period is set to 10 milliseconds.

[0044] The operating parameters include: (1) Electrical parameters: tank voltage, series current, equidistant voltage drop of each anode conductor, cathode current distribution, short circuit and column bus voltage drop; (2) Thermodynamic parameters: electrolyte temperature, tank shell steel temperature, flame color; (3) Material parameters: alumina feeding rate, aluminum fluoride addition amount, molecular ratio; (4) Tank condition characteristic parameters: tank resistance, tank resistance change rate, voltage swing amplitude and frequency, needle amplitude.

[0045] S2: Aluminum Electrolysis Process Expert Knowledge Base. An aluminum electrolysis process expert knowledge base is constructed based on process rules, safety constraints, process expert experience, and historical best practices. This knowledge base stores and infers domain knowledge in the field of aluminum electrolysis cell control, serving as a source of knowledge guidance and constraints for subsequent deep learning and control decisions. The hierarchical structure of the expert knowledge base is illustrated in the diagram below. Figure 2 As shown, it specifically includes the following four sub-libraries: (1) Process rule library. It contains a total of 127 core process rules, including rules on electrode distance-voltage relationship, alumina concentration control, cell temperature control and anode effect treatment. The rules are sourced from national standard GB / T 34480-2017 "Safety Production Specification for Aluminum Electrolysis", industry operating procedures and the experience of senior process experts. The rules for electrode spacing-voltage relationship are as follows: for every 1mm increase in electrode spacing, the cell voltage increases by approximately 35-40mV; the rules for alumina concentration control are as follows: when the cell resistance enters the low resistance zone (high concentration zone), the feed rate should be reduced or the feed rate should be suspended until the resistance returns to the normal zone; when the cell resistance enters the high resistance zone (low concentration zone), the feed rate should be increased; the rules for cell temperature control are as follows: when the superheat is too high (ΔT>15℃), the voltage setting value should be reduced or the amount of aluminum fluoride added should be increased to maintain thermal balance; when the superheat is too low (ΔT<5℃), the voltage setting value should be increased to increase heat input; the rules for handling anode effects are as follows: after the anode effect occurs, first check whether the alumina feed system is normal, and then perform appropriate effect extinguishing operations according to the duration of the effect.

[0046] Each rule is assigned a credibility weight and an applicable slot type range identifier. The rule format is IF <preconditions> THEN <conclusion / action>.

[0047] (2) Safety Constraint Library. Stores the safety boundary conditions and prohibited processes for operating aluminum electrolytic cells. Includes the following constraint categories: Electrode spacing safety constraints: The electrode spacing shall not be lower than the minimum safety threshold (preferably 38mm) to avoid serious accidents caused by anode contact with molten aluminum; the electrode spacing shall not be higher than the maximum design threshold (preferably 55mm) to avoid excessively high cell voltage leading to a surge in energy consumption; Safety boundary of superheat: The superheat of electrolyte should be maintained within the set safety range and should not exceed the preset upper limit (preferably 20°C) and lower limit (preferably 0°C) to avoid tank abnormalities or precipitation. Voltage swing tolerance threshold: When the voltage swing amplitude exceeds the first preset threshold, an early warning is triggered and further reduction of the voltage setting value is prohibited; Molecular ratio operation constraints: The rate of adjustment of the molecular ratio must not exceed the upper limit of the process to avoid drastic fluctuations in the tank conditions.

[0048] In this embodiment, constraints are set such as a safety threshold of 38-55 mm for electrode spacing, a voltage swing tolerance threshold of 30 mV / 5 min, and an upper limit of 0.02 / day for molecular ratio adjustment rate. Operations that violate the safety constraints are automatically vetoed.

[0049] (3) Process Expert Experience Database. This database stores the implicit operational experience of senior aluminum electrolysis process experts in the form of fuzzy rules and neural network weights. It includes: Typical value range of empirical parameters: empirical values ​​of optimal process parameter combinations for different tank types and tank ages (such as target voltage setting, feeding interval, target range of alumina concentration, etc.). Experience-based adjustment strategies: Standard response models adopted by process experts under typical operating condition deviations; Fuzzy expert rules: Based on the design concept of the aluminum electrolysis fuzzy expert controller, process control rules represented by fuzzy language variables are adopted, such as the pole pitch adjustment strategy corresponding to "high cell voltage" + "low superheat".

[0050] In this embodiment, the process expert experience database stores the optimal process parameter combination experience values ​​for three typical cell types (300kA, 400kA, 500kA) in different cell age ranges (0-365 days, 365-1000 days, and over 1000 days), as well as 12 fuzzy expert decision rules such as "high cell voltage + normal superheat → check alumina concentration first and then adjust electrode spacing".

[0051] (4) Historical Excellent Case Database. This database contains complete case data of electrolyzers that have been verified as "excellent operations" during historical operation. Each excellent case records the following information: The database includes: a tank condition characteristic vector (including temperature, superheat, alumina concentration, tank voltage, anode current distribution, etc.) at the time of the case; a series of control operations taken by the operator (or self-learning system) at the time of the case; evaluation indicators of the tank condition improvement effect after the case is executed; and tags for the applicable tank type, tank age, and tank condition type. This database of excellent historical cases is used to retrieve the best cases that are closest to the current tank condition through similarity matching during the decision-making stage, providing knowledge reference for control decisions.

[0052] In this embodiment, 482 typical cases that were verified as successful operations were selected from the historical operation data of the past 36 months. Each case includes a tank condition feature vector (18 dimensions), a control operation sequence (7 to 12 steps on average), and the tank condition improvement effect after execution (effect score ≥ 85 points).

[0053] S3: Electrolyte Mechanism Model. An electrolyte mechanism model is constructed based on the Hall-Héroult electrolysis reaction principle and the material-heat balance theory of the aluminum electrolysis process. This model is used to perform analytical calculations on the key state variables of the aluminum electrolyzer, providing physical consistency constraints and knowledge injection for the prediction output of the PatchTST network. The electrolyte mechanism model includes the following sub-modules: (1) Material balance mechanism model. A material balance equation is established based on Faraday's law between alumina consumption and the current passing through it. The theoretical consumption rate of alumina is calculated by the following formula: v_Al2O3 = (I × t × M_Al2O3) / (z × F), Where v_Al2O3 is the amount of alumina consumed, I is the series current, t is the energizing time, M_Al2O3 is the molar mass of alumina, z is the number of electrons transferred in the reaction, and F is the Faraday constant.

[0054] The model also considers the dissolution kinetics of alumina in electrolytes, such as... Figure 3 As shown, the relationship between the theoretical dissolution rate and the actual dissolution efficiency of alumina is calculated to estimate the trend of effective alumina concentration change in the electrolyte online, and cross-validation is performed with the alumina concentration drift trend predicted by the PatchTST network.

[0055] (2) Thermal equilibrium mechanism model.

[0056] The heat balance equation for the electrolytic cell is established based on the principle of energy conservation: Q_in = Q_reaction + Q_Joule = Q_out + ΔQ, Where Q_in is the total heat input of the system (the sum of electrochemical reaction heat Q_reaction and Joule heat Q_Joule), Q_out is the heat output of the system (the sum of heat loss from the tank shell Q_shell, heat carried away by the flue gas Q_gas, and heat carried away by the molten aluminum Q_metal), and ΔQ is the change in heat in the tank (reflecting the evolution of the thermal equilibrium state).

[0057] Based on real-time data such as tank voltage, series current, electrolyte temperature, and multi-point temperature of the tank shell, the model calculates the current thermal equilibrium state online and outputs an analytical estimate of the electrolyte superheat ΔT, which serves as the physical constraint benchmark for knowledge injection in S5.

[0058] (3) Joint mechanism model of electrode spacing-voltage-thermal balance.

[0059] Establish the analytical relationship between pole gap and slot voltage: U_cell = U_decomp + U_anode + U_cathode + IR_drop_bath + IR_drop_bubble + d × ρ_electrolyte × I / A, Where U_cell is the cell voltage, U_decomp is the theoretical decomposition voltage of Al2O3, U_anode and U_cathode are the anode and cathode overpotentials, IR_drop_bath is the electrolyte ohmic voltage drop, IR_drop_bubble is the bubble layer resistance voltage drop, d is the electrode distance, ρ_electrolyte is the electrolyte resistivity, I is the current, and A is the anode area.

[0060] Based on this equation, a coupling mapping relationship is established between the pole gap adjustment amount and the changes in slot voltage and thermal balance, which serves as a mechanistic constraint for calculating the pole gap compensation amount.

[0061] In this example, material balance, thermal balance, and electrode spacing-voltage joint mechanism models are deployed on the industrial control server. The material balance mechanism model calculates the deviation between the theoretical alumina consumption and the actual feed rate online based on Faraday's law, updating the concentration deviation trend every 10 seconds. The thermal balance mechanism model, based on energy conservation, calculates real-time heat dissipation loss based on the measured temperature distribution of the tank shell and estimates the electrolyte superheat. The electrode spacing-voltage-thermal balance joint mechanism model uses a 38mm electrode spacing as a baseline, and calculates the expected voltage change corresponding to electrode spacing adjustments in real-time based on the theoretical relationship that the theoretical tank voltage increases by approximately 38mV for every 1mm increase in electrode spacing.

[0062] S4: Enhanced PatchTST multi-scale temporal coding network. An enhanced PatchTST multi-scale temporal coding network is constructed and used to perform temporal coding and dynamic prediction of the electrolytic cell operating parameter data, outputting the predicted deviation of alumina concentration drift trend, the predicted deviation of electrolyte superheat, the predicted value of cell resistance, the predicted value of anode effect probability, and the predicted value of electrode distance, such as... Figure 4 As shown.

[0063] The enhanced PatchTST multi-scale temporal coding network includes a temporal data patch partitioning module, a knowledge distillation time-frequency feature extraction module, a channel-independent embedding layer, a multi-scale patch coding layer, a physical information constraint fusion module, and a prediction output layer.

[0064] (1) The temporal data Patch partitioning module divides the electrolytic cell operating parameter data collected by multiple source sensors into multiple time subsequence segments (Patch); each Patch contains a continuous time step, which serves as the input Token for the subsequent Transformer encoder. The Patch partitioning strategy compresses the local information of continuous time points into single-Token features, significantly reducing the computational complexity of Attention.

[0065] (2) Knowledge distillation time-frequency feature extraction module, performs time-domain feature extraction and frequency-domain feature extraction on the multiple time subsequence segments at the same time, and uses the process rules in the aluminum electrolysis process expert knowledge base as teacher knowledge to perform knowledge distillation on the frequency-domain feature extraction process, thereby enhancing the model’s ability to perceive multi-source time-domain information.

[0066] Joint time-frequency feature extraction performs time-domain and frequency-domain feature extraction simultaneously on the original time-series data. In the time domain, local waveform features are extracted using a multi-layer convolutional network; in the frequency domain, a Short-Time Fourier Transform (STFT) is used with a Hamming window (256 sampling points) and a 75% overlap rate to transform the time-domain signal into the frequency domain, extracting the energy distribution characteristics of the slot resistance signal in different frequency bands, including: Low-frequency component, 0–0.5 Hz (thermal equilibrium drift): reflects the long-term drift trend of the thermal and material balance of the electrolyzer.

[0067] Mid-frequency component, 0.5–2 Hz (aluminum molten interface fluctuation): associated with aluminum molten interface fluctuation and uneven anode current distribution.

[0068] High-frequency components, 2–10 Hz (bubbles and local anodic reactions): correspond to short-term dynamic processes such as bubble behavior and precursors of local anodic effects.

[0069] Knowledge distillation feature learning utilizes process rules from a process expert knowledge base to perform knowledge distillation on frequency domain features. Specifically, the process rules in the knowledge base serve as teacher knowledge (teacher model), and the multi-scale Transformer encoder branch of the PatchTST network serves as the student model. A knowledge distillation loss function is introduced to guide the time-frequency feature extraction process. The distillation temperature is T=3, and the knowledge distillation weight coefficient is 0.6. Teacher knowledge includes: reference energy distribution curves for each frequency band under normal tank conditions, typical frequency domain evolution patterns of tank resistance before the anode effect, and characteristic patterns of tank resistance spectra corresponding to different alumina concentration ranges. Knowledge distillation enables the PatchTST network to extract physically meaningful feature components from the tank resistance signal, avoiding the learning of spurious correlation features unrelated to the actual process and fundamentally suppressing "model illusion."

[0070] (3) Channel Independent Embedding Layer: The extracted features are processed using a channel independent mechanism, treating each sensor channel as an independent univariate time series and outputting a multi-channel Patch embedding vector. Each channel shares the same Embedding weight matrix and Transformer encoder parameters to enhance the adaptability to multi-source heterogeneous data.

[0071] (4) Multi-scale Patch Encoding Layer: The Patch embedding vectors of different channels (high-frequency fluctuations, mid-frequency oscillations, and low-frequency drifts) are input into independent Transformer encoder branches. The temporal dependencies within the Patch embedding vectors of each channel are captured through an attention mechanism. Then, a cross-scale feature interaction layer is established between the encoder branches to output a multi-scale joint feature representation. PatchTST parameter configuration: input window length L=7200 (corresponding to 60 minutes), Patch size is set to three scales [2, 4, 8], corresponding to time spans of 20 seconds for the short term, 40 seconds for the medium term, and 80 seconds for the long term. The Transformer encoder has 6 layers, a hidden dimension of 128, and 8 attention heads.

[0072] (5) Physical information constraint fusion module, which uses the analytical value of the electrolyte mechanism model as an auxiliary input feature and fuses it with the multi-scale joint feature representation, and uses a loss function with added physical regularization loss term to train the model.

[0073] The analytical calculation results of key state variables from the electrolyte mechanism model are used as physical knowledge injection points and integrated into the loss function and prediction output layer of the PatchTST network. Specific integration methods include: Physical regularization loss term, such as Figure 5 As shown, a physical constraint regularization term is added to the network's training loss function to penalize predictions that violate the principles of material and energy conservation. The physical regularization loss function is defined as follows: L_physics = λ1·||ΔC_Pred - ΔC_Mech||² + λ2·||ΔT_Pred - ΔT_Mech||² + λ3·||U_Pred - U_Mech||², Wherein, ΔC_Pred and ΔC_Mech are the alumina concentration change rate predicted by the PatchTST network and the theoretical change rate calculated by the S31 material balance mechanism model, respectively; ΔT_Pred and ΔT_Mech are the predicted superheat and the superheat reference value calculated by the S32 heat balance mechanism model, respectively; U_Pred and U_Mech are the predicted cell voltage and the reference voltage calculated by the S33 electrode gap-voltage-heat balance mechanism model, respectively; λ_1, λ_2, and λ_3 are weighting coefficients, with weighting coefficients λ1=0.4, λ2=0.4, and λ3=0.2, ensuring a balance between statistical accuracy and physical consistency in the prediction output.

[0074] Feedforward knowledge injection: In the cross-scale feature interaction layer of the PatchTST encoder, the physical state variables calculated by the mechanism model are used as auxiliary input features and concatenated or weighted with the multi-scale encoded features, so that the network output is in a physically feasible state space in the initial stage.

[0075] (6) Output prediction layer, flatten the fused features, and generate the final prediction result through prediction head.

[0076] The above-mentioned cross-scale joint features are extracted by the self-attention mechanism of the PatchTST encoder and mapped by the fully connected prediction head to the following prediction outputs: the predicted sequence of tank resistance in the future time window, the predicted drift trend of alumina concentration in each control area, the predicted evolution curve of electrolyte superheat, the predicted probability of anode effect (output range is 0 to 1), and the physical consistency confidence score: a confidence index for evaluating whether the prediction results conform to the principles of material balance and energy conservation.

[0077] S5: Knowledge graph construction, feature injection, and dynamic identification of slot conditions.

[0078] (1) Knowledge graph feature injection.

[0079] like Figure 6As shown, the process rule base, safety constraint base, and best practice case base in the process expert knowledge base are transformed into a structured knowledge graph. The nodes of the knowledge graph include: tank condition nodes (e.g., "alumina concentration too high," "tank temperature too high," "severe needle vibration"), operation nodes (e.g., "reduce feed," "reduce voltage," "add aluminum fluoride"), and constraint nodes (e.g., "pole distance ≥ 38mm"). Edges between nodes express causal relationships, temporal constraints, or recommendation / prohibition relationships. There are a total of 436 nodes (including 118 condition nodes, 203 operation nodes, and 115 constraint nodes) and 1250 relationship edges.

[0080] In the cross-scale feature interaction layer of the PatchTST network, a knowledge graph feature injection mechanism (GAT, 2 layers, 4 attention heads) is introduced. This mechanism fuses the feature vectors extracted from the multi-source sensor data of the current slot condition with the corresponding state node embedding vectors in the knowledge graph through attention, generating a knowledge-enhanced feature representation. This process enables the network to reason and correct prediction results based on the causal rules stored in the knowledge graph.

[0081] (2) Based on the prediction results, the stability of the tank condition is comprehensively evaluated and the early warning of the anode effect is provided.

[0082] By combining the multi-source prediction output of the PatchTST network and the feature injection results from the knowledge graph, a comprehensive evaluation system for trench stability is established. Tank condition stability level classification: Based on the fluctuation amplitude of the predicted tank resistance sequence, the stability of the predicted alumina concentration trend, and the deviation of the predicted electrolyte superheat curve, combined with the safety boundary constraints in the safety constraint library and the matching results of similar cases in the historical excellent case library, a multi-level fuzzy comprehensive evaluation method is used to classify the predicted tank condition within a set future time period into four levels: stable operation, slight fluctuation, moderate fluctuation, and severe fluctuation.

[0083] Anode effect early warning signal: When the predicted probability of the occurrence of the anode effect based on comprehensive assessment exceeds the first threshold (preferably the threshold is 65% to 75%), an anode effect early warning signal is generated and an anode effect emergency control command is triggered.

[0084] (3) Mechanism explanation output module.

[0085] Along with the prediction results, the module provides an explanation of the underlying mechanism. Based on causal reasoning chains in a knowledge graph, the output format is: "Based on the observed upward trend of cell resistance (feature A) and the slow downward trend of superheat (feature B), combined with the rule recorded in the knowledge graph 'IF cell resistance continues to rise AND superheat decreases THEN alumina concentration is low', the probability of the anodic effect occurring within the next 20 minutes is predicted to be 82%. Confidence score: 88% (good physical consistency)." This mechanism explanation significantly improves the explainability and user trust of the aluminum electrolysis cell control system.

[0086] The system outputs explanatory notes simultaneously with the issuance of control commands: "Based on the observed increase in the uneven distribution of anode current in control area #2, the 18% increase in low-frequency energy of the cell resistance compared to the normal baseline, and the absence of significant abnormalities in high-frequency energy, combined with the rule in the knowledge graph 'IF resistance low-frequency energy ↑ AND uneven regional current distribution THEN local alumina concentration low', the probability of the anode effect occurring within the next 30 minutes is predicted to be 67% (confidence score 86%, good physical consistency). It is recommended to increase the electrode spacing by 1.38mm (approximately 25mV) and implement independent material compensation of 6kg for area #2. This decision is based on the S2 expert knowledge base recommendation of 'prioritizing material replenishment when alumina concentration decreases' and the double confirmation of superheat constraint from the S3 thermal balance mechanism model." S6: Polar Moment Adaptive Adjustment and Tank Condition Co-optimization Decision (1) Prediction-Knowledge-Mechanism Fusion Decision-Making Framework like Figure 7 As shown, the time-domain prediction results output by the PatchTST network, the best practice guidance provided by the process expert knowledge base, and the physical constraint benchmarks provided by the electrolyte mechanism model are all used as inputs to control decisions, constructing a "data-mechanism-knowledge" three-driven fusion calculation module for optimal electrode spacing-voltage setting. The formula for calculating the electrode spacing compensation is: Δd = k1·ΔC_Pred + k2·ΔT_Pred + k3·E_vib + k4·P_AE + k5·K_KB +k6·D_physics, in: ΔC_Pred represents the prediction bias of alumina concentration drift trend based on the output of the PatchTST network in S4; ΔT_Pred is the prediction bias of electrolyte superheat based on the output of the PatchTST network in S4; E_vib is a comprehensive evaluation value of the slot resistance fluctuation sequence based on the output of the PatchTST network in S4; P_AE is the predicted probability of the anode effect output by the PatchTST network in S4; K_KB is the expert knowledge base influence factor: based on the current tank condition, the matching process rules and safety constraints are retrieved from the S2 expert knowledge base, and the comprehensive knowledge representation quantity is generated by injecting the frequency domain features extracted by knowledge distillation and the S51 knowledge graph features. D_physics is the physical deviation correction factor: the deviation correction amount between the reference value of the polar distance compensation mechanism calculated based on the S3 electrolyte mechanism model and the S4 predicted value. It is automatically activated when the predicted output deviates from the mechanism calculation result by more than the preset physical consistency tolerance threshold. k1, k2, k3, k4, k5, and k6 are weight coefficients that are dynamically optimized through online self-learning.

[0087] The core advantage of this polar moment compensation formula compared to existing technologies lies in the fact that it incorporates both the expert knowledge base (K_KB) and the mechanism model (D_physics) into the polar moment compensation calculation, thereby fundamentally improving the physical interpretability of control decisions and their adaptability to operating conditions.

[0088] (2) Dual-track alumina feeding control A dual-track alumina feeding control strategy is implemented: the first track is a fixed-cycle baseline feeding to maintain the basic material supply of the electrolytic cell; the second track is a dynamic feeding adjustment, which performs independent feeding compensation for local areas based on the real-time alumina concentration deviation of each control area collected in S1, the alumina concentration drift trend prediction value output by the PatchTST network in S4, and the optimal feeding strategy for the corresponding concentration range in the process rule library in S2.

[0089] (3) Cooperative control command generation and priority scheduling Taking into account three control dimensions—pole distance adjustment, material feeding adjustment, and thermal balance adjustment—a collaborative control instruction set is generated. The instruction priority setting rule is: emergency response to anode effects > tank condition stability maintenance (combined with safety boundary early warning from the safety constraint library) > fine adjustment of pole distance > material feeding balance optimization. When control instructions are issued, a decision-making basis explanation generated based on the knowledge base is attached for operators to understand and review.

[0090] During operation, changes in tank conditions were encountered: the PatchTST network predicted a 30-minute probability of anode effect P_AE = 0.67 (entering the warning range), the alumina concentration drift trend prediction deviation ΔC_Pred = -0.7wt% (decreasing trend), and the electrolyte superheat prediction deviation ΔT_Pred = +6℃ (slowly increasing). Simultaneously, the S3 thermal balance mechanism model calculated a superheat reference value ΔT_Mech = +4℃, and the matching process rule retrieved from the S2 expert knowledge base was "when the alumina concentration decreases in conjunction with a slight increase in tank temperature, prioritize increasing the feed rate and appropriately reducing the electrode spacing."

[0091] The three-drive fusion input electrode spacing compensation calculation module uses the current weight coefficients (determined through S7 stage self-learning as k1=0.008, k2=0.22, k3=0.12, k4=0.5, k5=0.18, k6=0.10) to calculate an electrode spacing compensation amount Δd≈1.38mm. Among these, the expert knowledge base influence factor K_KB=0.12 (it is recommended to reduce the electrode spacing by 2mm and then observe the cell voltage change before fine-tuning), and the physical deviation correction factor D_physics=0.05mm (given that the predicted superheat is higher than the mechanism calculation, the electrode spacing compensation is slightly reduced to maintain thermal balance). Based on this, the system increases the target voltage by approximately 24mV, coordinates the anode lifting mechanism to perform electrode spacing adjustment, and simultaneously triggers the dynamic adjustment trajectory of the dual-track feeding system, increasing feeding compensation three feeding cycles in advance for the control area with the strongest alumina concentration decrease trend.

[0092] S7: Self-learning module based on knowledge-data collaboration (1) Self-supervised pre-training stage like Figure 8 As shown, during system initialization deployment or new slot migration, the PatchTST network in S4 is pre-trained using offline historical running data, with channel-independent self-supervised training. Pre-training employs a random mask reconstruction task, simultaneously incorporating rule constraints extracted from the S2 expert knowledge base and physical consistency regularization terms extracted from the S3 mechanistic model. This allows the feature representations formed during pre-training to possess both statistical law learning and physical law recognition capabilities.

[0093] (2) Online incremental fine-tuning and parameter update During online system operation, the self-learning module periodically receives tank condition response data after the control system executes commands, and incrementally fine-tunes the decoder part of the PatchTST network. During this fine-tuning process, excellent operational cases stored in the historical best-case library are used as positive samples to guide the network update direction. Simultaneously, the pre-trained mechanism consistency constraint layer is frozen, improving the network's adaptability to the current tank condition while maintaining the physical interpretability of the prediction results.

[0094] The weight coefficients k1 to k6 are dynamically optimized online using the SAC reinforcement learning algorithm, enabling the control strategy to continuously adapt to the slow evolution of the electrolyzer's operating characteristics.

[0095] (3) Incremental updates of the expert knowledge base During online operation, the self-learning module automatically identifies new excellent operation cases (i.e., the complete process in which the tank condition significantly improves after control measures are taken under specific tank conditions). After being reviewed and confirmed by process experts or filtered through automated rules, these cases are stored in the historical excellent case library. When a sufficient number of similar cases are accumulated, the confidence level of relevant rules in the process rule library is automatically updated or new rules are generated.

[0096] The self-learning module performs online incremental fine-tuning daily, updating the PatchTST network decoder parameters. Weight reset and offline retraining are performed every 7 days to maintain the model's adaptability to new operating conditions through rolling updates. New and excellent operational cases discovered during production operations are automatically stored and updated to the historical excellent case library after review and confirmation by process engineers.

[0097] This embodiment was fully deployed on 288 500kA series electrolytic cells. Statistics show that the anode effect coefficient decreased by 48.7% compared to the baseline; the number of physical consistency violation predictions (prediction results deviating from the S3 mechanism model calculation results by more than the tolerance threshold ±5%) decreased by 76.3% compared to the pure PatchTST version; there were no "illusion" outputs violating the first law of thermodynamics; operator trust ratings for the system's predictions and decisions (out of 1-10) increased from 6.2 at the beginning of deployment to 9.1; the knowledge graph-driven mechanism explanation output function received unanimous praise; current efficiency improved by 0.3%; DC power consumption per ton of aluminum decreased by 100 kWh / ton of aluminum; and the expert knowledge base added 126 new, system-verified excellent operating cases to the initial 482 rules within 12 months. The incremental update module of the expert knowledge base enabled the control system to exhibit continuous self-evolutionary characteristics.

[0098] Through the above practices, the aluminum electrolytic cell electrode spacing control method integrating time-series model and mechanism model described in this invention has the following advantages: 1. While maintaining the advantages of PatchTST's multi-scale coding, it deeply integrates two major knowledge-driven forces: the electrolysis mechanism model (material conservation + energy conservation + electrode distance-voltage-thermal balance) and the expert knowledge base (process rules + safety constraints + expert experience + best practices). This fundamentally elevates the control system from "relying on data to guess the tank condition" to a three-dimensional cognitive level of "understanding physical laws + following expert experience guidance + accurate prediction based on data".

[0099] 2. This scheme applies the teacher-student framework of knowledge distillation to the field of aluminum electrolysis cell control. Using an expert knowledge base storing decades of best practices in aluminum electrolysis as the teacher, and the multi-scale encoder of the PatchTST network as the student, "supervised learning" is performed on the energy distribution of the cell resistance signal in three frequency bands: 0–0.5Hz (thermal equilibrium drift), 0.5–2Hz (aluminum melt interface fluctuations), and 2–10Hz (bubbles and local anodic reactions). This ensures that the network's extraction results of cell condition features naturally possess frequency domain distribution attributes with clearly defined physical meaning verified by the process. Essentially, this elevates an "implicit black-box feature extractor" to an "explicit physical feature extractor," ensuring that the model will not produce outputs that violate physical laws under any new operating conditions, thus fundamentally suppressing "model illusion."

[0100] 3. The theoretical values ​​calculated analytically by the S3 electrolyte mechanism model are directly used as constraints for training the PatchTST network. Through a physical regularization loss function and the D_physics physical bias correction factor in the three-drive moment compensation formula, it is ensured that the model's predicted and decision values ​​will not deviate from fundamental natural laws such as material and energy conservation under any changing operating conditions. This design completely changes the traditional model of "predicting first, then making decisions based on the predicted values ​​without understanding the physical meaning of the decisions," with each level of output having a clear physical verification step.

[0101] 4. Expert knowledge in this solution permeates the entire control loop process—S2 knowledge base construction, S4 knowledge distillation guiding feature extraction, S5 knowledge graph feature injection enhancing encoding representation, S6 three-drive fusion decision-making, and S7 incremental knowledge base update. This design enables the control system to truly possess human-like learning capabilities—"utilizing knowledge, accumulating knowledge, and evolving knowledge"—possessing irreplaceable knowledge asset value for aluminum electrolysis enterprises with high personnel turnover and the easy loss of core process experience.

[0102] 5. The mechanism explanation output module proposed in this solution, combined with the causal reasoning chain of the knowledge graph, transforms the prediction result from a confusing "probability number" into clear and easy-to-understand engineering language. Operators no longer see a "confusing 82%", but rather "based on a 18% increase in low-frequency characteristic energy and no abnormalities in high-frequency characteristics, according to knowledge base rule R124 'When the low-frequency energy of the resistor increases and the regional current distribution is uneven, the local alumina concentration is low,' predicting an 82% probability of an anode effect in the next 30 minutes." This significantly improves the reliability of the prediction and user acceptance, giving this patented technology the practical deployment value to directly replace and even surpass traditional manual judgment.

[0103] In another embodiment, an aluminum electrolytic cell electrode spacing control system that integrates timing models and mechanistic models is provided, such as... Figure 9 As shown, the system is used to implement the above-mentioned method for controlling the electrode spacing of aluminum electrolytic cells by integrating the timing model and the mechanism model, including: The multi-source data acquisition unit is used to acquire electrolytic cell operating parameter data collected by multi-source sensors in the aluminum electrolytic cell.

[0104] The expert knowledge base construction unit includes a process rule base module, a safety constraint base module, a process expert experience base module, and a historical best case base module. It is used to store and reason about domain knowledge in the field of aluminum electrolysis cell control, and to provide prior guidance and boundary constraints for model training and control decisions.

[0105] The electrolyte mechanism model unit includes a material balance mechanism module, a thermal balance mechanism module, and a combined electrode distance-voltage-thermal balance mechanism module, which are used to perform analytical calculations on the key state variables of the aluminum electrolysis cell.

[0106] The enhanced PatchTST cell condition dynamic identification unit includes a time-series data patch partitioning module, a knowledge distillation time-frequency feature extraction module, a channel independent embedding module, a multi-scale encoding module, a physical information constraint fusion module, and a prediction output layer module. It performs time-series encoding and dynamic prediction on electrolytic cell operating parameter data, and outputs prediction deviations for alumina concentration drift trend, electrolyte superheat, cell resistance, anode effect probability, and electrode distance, enabling early perception of cell condition evolution.

[0107] The knowledge graph feature injection unit includes a knowledge graph construction module, a feature fusion module, and a mechanism explanation output module. The knowledge graph construction module is used to transform the aluminum electrolysis process expert knowledge base into a structured knowledge graph. The feature fusion module is used to perform attention fusion between the embedding vectors of the corresponding state nodes in the knowledge graph and the multi-scale joint feature representation to obtain the process expert knowledge base influence factor. The mechanism explanation output module is used to provide a mechanism explanation of the prediction results while the enhanced PatchTST multi-scale temporal coding network outputs the prediction results.

[0108] The extreme distance adaptive adjustment unit includes a prediction-knowledge-mechanism fusion decision module and an extreme distance actuator control module. The prediction-knowledge-mechanism fusion decision module takes the prediction results output by the enhanced PatchTST multi-scale time-series coding network, the best practice guidance provided by the process expert knowledge base, and the physical constraint benchmarks provided by the electrolyte mechanism model as multi-source inputs, and generates an extreme distance control strategy through fusion. The extreme distance actuator control module is used to calculate the extreme distance compensation amount and generate a set of cooperative control instructions to be issued to the actuator. The material feeding coordination control unit is used to execute the dual-track alumina feeding control strategy.

[0109] The collaborative self-learning update unit includes a self-supervised pre-training module, an online incremental fine-tuning module, a reinforcement learning weight optimization module, and a knowledge base incremental update module. It is used to continuously learn new data, optimize model parameters, and update rule knowledge during system operation.

[0110] The central control unit connects all units, unifies the management and control timing and command priority, and ensures the coordinated and orderly operation of multiple units.

Claims

1. A method for controlling the electrode spacing of an aluminum electrolytic cell by integrating a time-series model and a mechanistic model, characterized in that, The method includes: The operating parameter data of the aluminum electrolysis cell were collected by multi-source sensors. Then, an aluminum electrolysis process expert knowledge base was built based on process rules, safety constraints, process expert experience and historical best practices. An electrolyte mechanism model was also built based on the Hall-Herroux electrolysis reaction principle and material-heat balance theory. An enhanced PatchTST multi-scale temporal coding network was constructed and used to perform temporal coding and dynamic prediction on the electrolytic cell operating parameter data, outputting the alumina concentration drift trend prediction deviation, electrolyte superheat prediction deviation, cell resistance prediction value, anode effect probability prediction value, and electrode distance prediction value. The enhanced PatchTST multi-scale temporal coding network includes a temporal data patch partitioning module, a knowledge distillation time-frequency feature extraction module, a channel-independent embedding layer, a multi-scale patch coding layer, a physical information constraint fusion module, and a prediction output layer. The temporal data patch partitioning module divides the electrolytic cell operating parameter data collected by multiple sources of sensors into multiple time-series segments. The knowledge distillation time-frequency feature extraction module simultaneously performs time-domain feature extraction and frequency-domain feature extraction on the multiple time-series segments, and uses process rules from the aluminum electrolysis process expert knowledge base as teacher knowledge to perform knowledge distillation on the frequency-domain feature extraction process. The channel-independent embedding layer extracts... The features are processed using a channel-independent mechanism, outputting multi-channel patch embedding vectors. The multi-scale patch encoding layer inputs the patch embedding vectors of different channels into independent encoder branches to capture the temporal dependencies within each channel's patch embedding vector. Then, a cross-scale feature interaction layer is established between each encoder branch to output a multi-scale joint feature representation. The physical information constraint fusion module uses the analytical values ​​of the electrolyte mechanism model as auxiliary input features and fuses them with the multi-scale joint feature representation. The model is trained using a loss function with an added physical regularization loss term. The output prediction layer flattens the fused features and generates the final prediction result through the prediction head. The aluminum electrolysis process expert knowledge base is transformed into a structured knowledge graph. Then, the embedding vectors of the corresponding state nodes in the knowledge graph are fused with the multi-scale joint feature representation to obtain the process expert knowledge base influence factor. Based on the predicted values ​​of alumina concentration drift trend, electrolyte superheat, cell resistance, anode effect probability, and electrode distance output by the enhanced PatchTST multi-scale time-series coding network, as well as the influence factors from the process expert knowledge base, the electrode distance compensation is calculated and a collaborative control instruction set is generated and sent to the actuator.

2. The method for controlling the electrode spacing of an aluminum electrolytic cell according to claim 1, characterized in that, The formula for calculating the polar distance compensation is as follows: Δd = k1·ΔC_Pred + k2·ΔT_Pred + k3·E_vib + k4·P_AE + k5·K_KB + k6·D_physics, Wherein, Δd is the electrode distance compensation amount, ΔC_Pred is the prediction deviation of alumina concentration drift trend, ΔT_Pred is the prediction deviation of electrolyte superheat, E_vib is the predicted cell resistance value, P_AE is the predicted anode effect probability value, K_KB is the process expert knowledge base influence factor, D_physics is the deviation correction amount between the electrode distance compensation mechanism reference value calculated by the electrolyte mechanism model and the electrode distance prediction value, and k1, k2, k3, k4, k5, and k6 are weighting coefficients.

3. The method for controlling the electrode spacing of an aluminum electrolytic cell according to claim 1, characterized in that, The time-domain feature extraction is performed by extracting local waveform features through a multi-layer convolutional network, and the frequency-domain feature extraction is performed by converting the time-domain signal to the frequency domain using short-time Fourier transform to extract the energy distribution features of the slot resistance signal in different frequency bands.

4. The method for controlling the electrode spacing of an aluminum electrolytic cell according to claim 1, characterized in that, The process rules in the aluminum electrolysis process expert knowledge base are used as teacher knowledge to perform knowledge distillation on the frequency domain feature extraction process. Specifically, this includes: using the process rules in the aluminum electrolysis process expert knowledge base as teacher knowledge, using each independent encoder branch as a student model, and introducing a knowledge distillation loss function to guide the time-frequency feature extraction. The teacher knowledge includes: reference distribution curves of energy in each frequency band under normal tank conditions, typical frequency domain evolution patterns of tank resistance before the occurrence of the anode effect, and characteristic patterns of tank resistance spectrum corresponding to different alumina concentration ranges.

5. The method for controlling the electrode spacing of an aluminum electrolytic cell according to claim 1, characterized in that, The physical regularization loss function is specifically as follows: L_physics = λ1·||ΔC_Pred - ΔC_Mech||² + λ2·||ΔT_Pred - ΔT_Mech||² +λ3·||U_Pred - U_Mech||², Wherein, ΔC_Pred is the prediction deviation of alumina concentration drift trend, ΔC_Mech is the deviation of theoretical change in alumina concentration calculated by the electrolyte mechanism model, ΔT_Pred is the prediction deviation of electrolyte superheat, ΔT_Mech is the reference value of electrolyte superheat calculated by the electrolyte mechanism model, U_Pred is the predicted value of cell voltage, U_Mech is the reference value of cell voltage calculated by the electrolyte mechanism model, and λ1, λ2, and λ3 are weighting coefficients.

6. The method for controlling the electrode spacing of an aluminum electrolytic cell according to claim 1, characterized in that, The actuator adopts a dual-track alumina feeding control strategy, including: the first track is a fixed-cycle baseline feeding to maintain the basic material supply of the electrolytic cell; the second track is a dynamic adjustment feeding, which performs independent feeding compensation for local areas based on the real-time alumina concentration deviation of each control area collected by the sensor, the alumina concentration drift trend prediction value output by the enhanced PatchTST multi-scale time-series coding network, and the optimal feeding strategy for the corresponding concentration range in the process rules.

7. The method for controlling the electrode spacing of an aluminum electrolytic cell according to claim 1, characterized in that, While the enhanced PatchTST multi-scale temporal coding network outputs the prediction results, it also provides an explanation of the mechanism behind the prediction results.

8. The method for controlling the electrode spacing of an aluminum electrolytic cell according to claim 1, characterized in that, It also includes receiving the tank condition response data after execution, updating the prediction parameters of the enhanced PatchTST multi-scale temporal coding network and the aluminum electrolysis process expert knowledge base through a knowledge-data collaborative self-learning module.

9. The method for controlling the electrode spacing of an aluminum electrolytic cell according to claim 1, characterized in that, The operating parameter data includes: electrical parameter data, thermodynamic parameter data, and tank condition characteristic parameter data.

10. A control system for the electrode spacing of an aluminum electrolytic cell that integrates a time-series model and a mechanistic model, characterized in that, The system is used to implement the aluminum electrolytic cell electrode spacing control method as described in any one of claims 1-9, comprising: The multi-source data acquisition unit is used to acquire electrolytic cell operating parameter data collected by multi-source sensors in the aluminum electrolytic cell; The expert knowledge base construction unit includes a process rule base module, a safety constraint base module, a process expert experience base module, and a historical excellent case base module. It is used to store and reason about domain knowledge in the field of aluminum electrolysis cell control, and to provide prior guidance and boundary constraints for model training and control decisions. The electrolyte mechanism model unit includes a material balance mechanism module, a thermal balance mechanism module, and an electrode distance-voltage-thermal balance combined mechanism module, which are used for analytical calculation of key state variables of aluminum electrolysis cells; The enhanced PatchTST cell condition dynamic identification unit includes a time-series data patch partitioning module, a knowledge distillation time-frequency feature extraction module, a channel independent embedding module, a multi-scale encoding module, a physical information constraint fusion module, and a prediction output layer module. It performs time-series encoding and dynamic prediction on electrolytic cell operating parameter data, and outputs the alumina concentration drift trend prediction deviation, electrolyte superheat prediction deviation, cell resistance prediction value, anode effect probability prediction value, and electrode distance prediction value to achieve early perception of cell condition evolution. The knowledge graph feature injection unit includes a knowledge graph construction module, a feature fusion module, and a mechanism explanation output module. The knowledge graph construction module is used to transform the aluminum electrolysis process expert knowledge base into a structured knowledge graph. The feature fusion module is used to perform attention fusion between the embedding vector of the corresponding state node in the knowledge graph and the multi-scale joint feature representation to obtain the process expert knowledge base influence factor. The mechanism explanation output module is used to provide a mechanism explanation of the prediction results while the enhanced PatchTST multi-scale temporal coding network outputs the prediction results. The extreme distance adaptive adjustment unit includes a prediction-knowledge-mechanism fusion decision module and an extreme distance actuator control module. The prediction-knowledge-mechanism fusion decision module takes the prediction results output by the enhanced PatchTST multi-scale time-series coding network, the best practice guidance provided by the process expert knowledge base, and the physical constraint benchmark provided by the electrolyte mechanism model as multi-source inputs, and generates an extreme distance control strategy through fusion. The extreme distance actuator control module is used to calculate the extreme distance compensation amount and generate a set of cooperative control instructions to be issued to the actuator. The material feeding coordination control unit is used to execute the dual-track alumina feeding control strategy; The collaborative self-learning update unit includes a self-supervised pre-training module, an online incremental fine-tuning module, a reinforcement learning weight optimization module, and a knowledge base incremental update module. It is used to continuously learn new data, optimize model parameters, and refresh rule knowledge during system operation. The central control unit connects all units, unifies the management and control timing and command priority, and ensures the coordinated and orderly operation of multiple units.

Citation Information

Patent Citations

  • Data- and Knowledge-Based Methods and Equipment for Obtaining Process Control Parameters in Aluminum Electrolysis Cells

    CN108360021B