Molecular ratio optimization control model and prediction method under complex electrolyte system

Abstract

The application relates to the technical field of automation control, and discloses a molecular ratio optimization control model and prediction method under a complex electrolyte system, which comprises the following modules: a data acquisition module, which is used for collecting cell control daily reports and test data in an aluminum electrolysis production process as basic data; a feature selection module, which is used for screening core features; a model training and prediction module, which is internally provided with a prediction model; the prediction model adopts a regression prediction model fusing a self-attention mechanism, the prediction model takes the screened core features as input, and outputs an aluminum fluoride addition amount prediction value through dynamic weighting feature dependency relationship; a control execution module, which is used for converting the addition amount prediction value into an aluminum fluoride addition instruction and issuing the aluminum fluoride addition instruction to a production line control system, so that automatic execution of an addition task is realized; and a man-machine interaction module, which is used for providing a visual interface to realize real-time information display. The application can accurately and efficiently predict the aluminum fluoride addition amount, and realizes intelligent and automatic control of the aluminum fluoride addition process.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of automation control technology, in particular to a complex electrolyte system under a molecular ratio optimization control model and prediction method. BACKGROUND

[0002] In the industrial aluminum electrolysis production process, maintaining a suitable molecular ratio in the electrolyte system is the core element to achieve efficient, stable and energy-saving operation. The accurate control of the molecular ratio directly affects the thermal balance of the electrolytic cell, the dissolution characteristics of alumina, the current efficiency and the energy consumption level. Specifically, the addition amount of aluminum fluoride is the key operating variable to regulate the molecular ratio. Traditionally, the decision of the aluminum fluoride addition amount highly depends on the experience accumulation of the operators and the regular offline electrolyte composition analysis.

[0003] This mode has significant limitations: first, manual experience is limited by individual differences and is difficult to quantify inheritance, which is prone to subjectivity; second, regular detection has serious hysteresis and cannot capture the dynamic changes of electrolyte composition in real time, especially when the production load fluctuates, the raw material characteristics change or the environmental conditions (such as ambient temperature) are disturbed, the hysteresis effect is more prominent; third, manual operation is slow and prone to errors, making it difficult to achieve fine and continuous control. As a result, the electrolyte molecular ratio is often in a non-ideal state with large fluctuations. The excessive fluctuation of the molecular ratio directly damages the stability of the production process, reduces the current efficiency (which means that the production of aluminum per unit power consumption decreases), increases energy consumption (especially invalid ohmic heat loss), and may increase the frequency of abnormal working conditions (such as anode effect) in the electrolytic cell, ultimately affecting the consistency of the primary aluminum product quality and production cost.

[0004] Therefore, with the continuous improvement of the modern aluminum industry's requirements for fine, intelligent, efficient and green production processes, it is urgent to abandon the traditional extensive control mode that relies on manual experience and low-frequency detection, and instead develop a complex electrolyte system molecular ratio optimization control model and prediction method that can be based on real-time data, has high-precision prediction capability and dynamic optimization and adjustment functions. SUMMARY

[0005] The present application aims to provide a complex electrolyte system under a molecular ratio optimization control model and prediction method, which can accurately and efficiently predict the addition amount of aluminum fluoride and realize intelligent and automated control of the aluminum fluoride addition process, helping to improve the overall technical and economic indicators and operational stability of the electrolytic production.

[0006] To achieve the above-mentioned purpose, the present application provides the following basic scheme.

[0007] Scheme One

[0008] A complex electrolyte system under a molecular ratio optimization control model, comprising:

[0009] a data collection module configured to collect cell control daily reports and inspection data in the aluminum electrolysis production process as basic data;

[0010] a feature selection module configured to calculate feature importance scores of the basic data based on a random forest algorithm and select a preset number of core features;

[0011] a model training and prediction module, which is internally provided with a prediction model; the prediction model is a regression prediction model with a fusion self-attention mechanism, and the prediction model takes the selected core features as input and outputs an aluminum fluoride addition amount prediction value through dynamic weighting feature dependency;

[0012] a control execution module configured to convert the addition amount prediction value into an aluminum fluoride addition instruction and issue it to a production line control system to realize automatic execution of the addition task; the aluminum fluoride addition instruction includes an additive amount, a feeding rate and a feeding duration;

[0013] a human-computer interaction module configured to provide a visual interface to display the running state of the electrolytic cell, the addition amount prediction value and the molecular ratio threshold alarm information in real time.

[0014] Further, the cell control daily report includes cell control parameters, and the cell control parameters at least include current intensity, voltage, uniformity difference, temperature, aluminum level, quality level, alumina discharge amount, furnace bottom pressure drop; the inspection data at least includes calcium fluoride content, sodium oxide content, potassium fluoride content, iron content, silicon content, and molecular ratio.

[0015] Further, it further comprises a preprocessing module; the preprocessing module is used for performing standardization processing, outlier elimination and missing value completion operation on the collected basic data.

[0016] Further, when performing outlier elimination, the preprocessing module adopts the interquartile range method to eliminate outliers, including the following operations:

[0017] S1, calculate the 25th percentile Q1 and the 75th percentile Q3 of each type of data to determine the interquartile range IQR = Q3-Q1;

[0018] S2, set the outlier determination interval as [Q1-1.5×IQR, Q3+1.5×IQR], and eliminate the data points outside the interval;

[0019] S3, repeat S1-S2 until there is no outlier in the basic data, and at the same time, the missing data is completed by using the linear interpolation algorithm.

[0020] Further, the core features screened by the feature selection module are the features ranked in the top 10 in terms of feature importance scores, and include calcium fluoride content, aluminum level, quality level, current intensity, voltage, and temperature; and the hyperparameters of the random forest algorithm are optimized through grid search or Bayesian optimization method.

[0021] Further, the regression prediction model is a gradient boosting decision tree model; and the self-attention mechanism of the prediction model is implemented in the following manner:

[0022] receiving the feature vector generated by the gradient boosting decision tree;

[0023] calculating the correlation weight between pairs of features in the feature dimension or time dimension for the feature vector; and generating a dynamically weighted feature representation through weighted fusion according to the correlation weight;

[0024] mapping the weighted feature representation to a predicted value of the aluminum fluoride addition amount.

[0025] Further, the human-computer interaction module further comprises:

[0026] a multi-level permission management unit for setting and distinguishing administrator configuration, operator viewing, and patrol alarm receiving permissions;

[0027] a data interface unit for providing an API for external MES / DCS systems to call the prediction result;

[0028] a historical data tracing unit for supporting molecular ratio fluctuation trend analysis and early warning record query.

[0029] Further, the control execution module is provided with an abnormal response mechanism:

[0030] when the predicted value exceeds the preset threshold for two consecutive rounds, a model re-calculation verification process is triggered, and the prediction model is re-called for verification;

[0031] when the verification result is still abnormal, an alarm information is automatically pushed to a specified terminal device;

[0032] The control execution module also communicates with the production line PLC system in real time, and synchronously starts an audible and visual alarm device when the alarm information is triggered.

[0033] Further, the prediction model supports collaborative prediction and centralized scheduling of multiple electrolytic cells, and the control execution module realizes automatic management of the addition task of the whole production line through a unified control center.

[0034] Scheme two

[0035] A molecular ratio optimization prediction method under a complex electrolyte system, applying a molecular ratio optimization control model under a complex electrolyte system as described in scheme one, to carry out molecular ratio optimization prediction, comprising the following steps:

[0036] Collecting tank control daily reports and test data, generating basic data after pretreatment as a training data set;

[0037] Screening core features using a random forest algorithm, and constructing a regression prediction model integrating a self-attention mechanism;

[0038] Setting a molecular ratio threshold through a visual interface, and outputting a real-time aluminum fluoride addition amount prediction value;

[0039] Converting the prediction value into a control instruction and issuing it to the electrolytic tank control system to dynamically adjust the addition strategy;

[0040] When the prediction value causes the molecular ratio to exceed the preset threshold for two consecutive rounds, an alarm mechanism is started, and the results are closed-loop tracked and processed.

[0041] The working principle and advantages of the present application are:

[0042] The molecular ratio optimization control model and prediction method under a complex electrolyte system of the present application can accurately and efficiently predict the aluminum fluoride addition amount, and realize intelligent and automated control of the aluminum fluoride addition process, which helps to improve the overall technical and economic indicators and operational stability of electrolytic production.

[0043] First, the present scheme constructs a multi-level optimization architecture for complex electrolyte systems, which can maintain high prediction accuracy and efficiency under complex electrolyte systems.

[0044] Specifically, the present scheme first realizes feature dimension reduction and importance screening of high-dimensional industrial data through a random forest algorithm, effectively overcoming the overfitting and low computational efficiency problems caused by redundant input parameters in traditional control models. Moreover, the feature selection unit uses random forest instead of the conventional correlation coefficient method, as it can better handle the nonlinear relationships between industrial parameters and match the gradient boosting decision tree with the subsequent attention mechanism. Further, the present scheme innovatively integrates gradient boosting decision trees with self-attention mechanisms, enabling the model to have dual capabilities - gradient boosting decision trees capture local nonlinear relationships through iterative optimization of residuals, while self-attention mechanisms model global feature dependencies through dynamic weight distribution, compensating for the shortcomings of tree models in modeling feature interaction relationships, and significantly enhancing the model's representation ability for the multivariate strong coupling characteristics of the electrolyte system.

[0045] At the control execution layer, the system achieves rapid response from prediction to execution through a real-time data flow closed-loop feedback mechanism (including threshold-triggered recalculation verification and PLC linkage response), overcoming the lag inherent in traditional manual control. The resulting human-machine collaborative framework ensures both the automation of control decisions and the retention of key human intervention nodes through a visual interface, improving accuracy while ensuring operational safety of the industrial system. Furthermore, once deployed on the production line, the model provided by this solution can automatically execute aluminum fluoride addition tasks, significantly reducing manual intervention and improving production automation and management efficiency.

[0046] Second, although random forest, gradient boosting decision tree and self-attention mechanism are all known algorithms, their combination and application in the field of electrolyte molecular ratio control and in solving specific technical problems is highly innovative.

[0047] First, industrial electrolyte systems face unique constraints such as high noise levels and complex variable coupling in high-temperature, highly corrosive environments during data acquisition. This solution creatively embeds a self-attention mechanism into the output layer of the tree model instead of the original input layer: high-order feature representations are extracted first using a gradient-boosting decision tree, and then the attention mechanism redistributes feature dimension weights. This design effectively avoids interference from noise in the original industrial data on the calculation of attention weights. Second, addressing the real-time requirements of aluminum fluoride addition control, this solution designs a dual-threshold mechanism for abnormal responses (verification triggered by two consecutive exceedances). This strategy differs from simple single-threshold alarms and fixed-period verification; instead, it dynamically adjusts the verification intensity based on system confidence, resulting in higher reliability. Furthermore, traditional process control models typically optimize algorithm accuracy or execution efficiency separately. This solution, through the synergy of a feature selection unit and a lightweight attention mechanism, achieves a balance between predictive accuracy and the time constraints of real-time control in industrial settings, thus achieving a dual-efficiency equilibrium. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the model structure of an embodiment of the molecular ratio optimization control model and prediction method for complex electrolyte systems according to the present invention;

[0049] Figure 2 This is a schematic diagram of the method flow of an embodiment of the molecular ratio optimization control model and prediction method for complex electrolyte systems according to the present invention. Detailed Implementation

[0050] The following detailed explanation illustrates the specific implementation methods:

[0051] The basic implementation examples are as follows: Figure 1 As shown: A molecular ratio optimization control model for a complex electrolyte system, including:

[0052] The data acquisition module is used for collecting the cell control daily report and inspection data in the aluminum electrolysis production process as basic data. The cell control daily report contains cell control parameters, and the cell control parameters at least include current intensity, voltage (including cell working voltage, anode effect voltage, polar distance voltage and electrolyte resistance voltage), uniform difference (i.e. single cell anode current uniform variance), temperature (including electrolyte temperature), aluminum level, quality level (i.e. electrolyte level), alumina discharge amount, furnace bottom pressure drop; and the inspection data at least include calcium fluoride content, sodium oxide content, potassium fluoride content, iron content, silicon content and molecular ratio.

[0053] The preprocessing module is used for performing standardization processing, outlier elimination and missing value completion operation on the collected basic data, so as to ensure the accuracy and integrity of the data.

[0054] When the preprocessing module performs outlier elimination, the quartile range method is adopted to eliminate the outliers, including the following operations:

[0055] S1, the 25th percentile Q1 and the 75th percentile Q3 of various data are calculated, and the quartile range IQR = Q3-Q1 is determined;

[0056] S2, the outlier judgment interval is set as [Q1-1.5×IQR, Q3+1.5×IQR], and the data points exceeding the interval are eliminated;

[0057] S3, S1-S2 is repeated until no outlier exists in the basic data, and the linear interpolation algorithm is used to complete the missing data.

[0058] Through the quartile range method of multiple iterations to eliminate outliers, the extreme values in the data can be accurately identified and removed step by step, avoiding the possible omission of outliers in single processing; at the same time, the linear interpolation algorithm for completing the missing data not only preserves the original distribution characteristics of the data, but also ensures the continuity of the data, which can provide high-quality, complete and reliable data basis for subsequent model training, thereby effectively improving the stability and prediction accuracy of the model.

[0059] The feature selection module is used for calculating the feature importance score based on the basic data through the random forest algorithm, and screening a preset number of core features.

[0060] The number of core features screened by the feature selection module is the feature whose importance score ranking is in the top 10, and includes calcium fluoride content, aluminum level, quality level, current intensity, voltage and temperature; and the hyperparameters of the random forest algorithm are optimized by grid search or Bayesian optimization method.

[0061] Specifically, the feature selection module takes the basic data as input, calculates the feature importance score by using the random forest algorithm : ;

[0062] in, For the total number of decision trees, Gini impurity before node splitting; Weighted Gini impurity of child nodes after splitting;

[0063] right Perform sliding window normalization and obtain .

[0064] Pick The features are used as candidate features. When the number of candidate features is greater than 10, they are selected according to... The first 10 items are selected in descending order. In this embodiment, The value is 0.8.

[0065] When tuning the hyperparameters of the random forest algorithm, we first search for the optimal tree depth and number of split features through Bayesian optimization; then we select the parameter combination that minimizes the molecular ratio prediction MAE based on time series cross-validation.

[0066] Adjusting the model hyperparameters can further optimize the feature selection effect, thereby providing reliable data support for subsequent prediction of aluminum fluoride addition, and improving the model's generalization ability and prediction accuracy.

[0067] The model training and prediction module includes a prediction model. This prediction model employs a regression prediction model incorporating a self-attention mechanism. It takes selected core features as input and outputs a predicted value for the amount of aluminum fluoride added through dynamically weighted feature dependencies. In this embodiment, the regression prediction model is a gradient boosting decision tree model; specifically, the XGBoost model. This algorithm possesses superior efficiency and flexibility, making it suitable for complex industrial data analysis scenarios.

[0068] Furthermore, the self-attention mechanism of the prediction model is implemented in the following manner:

[0069] Receive the feature vectors generated by the gradient boosting decision tree, which correspond to the feature vectors in the following text. ;

[0070] For the feature vector, the correlation weight between pairs of features is calculated in the feature dimension or time dimension, that is, the correlation weight between different time steps or feature dimensions is calculated, which corresponds to step (3) below; according to the correlation weight, a dynamic weighted feature expression is generated by weighted fusion, which corresponds to step (4) below, thereby improving the model's ability to fit complex nonlinear relationships and further improving the prediction accuracy;

[0071] The weighted feature representation is mapped to the predicted value of aluminum fluoride addition, corresponding to steps (5) to (7) below.

[0072] Specifically, the 10 core features selected Input the XGBoost model and generate A regression decision tree, in this embodiment, is the default. Each tree outputs the index of its leaf nodes. , It is a set of positive integers.

[0073] Reconstruct the feature vector ; Represents the set of real numbers. Let be the dimension of the real matrix, where d is the number of trees, and d is the total number of leaf nodes, which is the feature dimension of the output of each tree (e.g., the length of the vector after encoding the index of a leaf node for each tree).

[0074] .in, This represents vector concatenation. One-hot encoding is performed on the leaf node indexes. This operation can transform the decision logic of the tree model into numerical features while preserving fine-grained decision information.

[0075] Based on feature vectors Perform multi-head attention computation and concatenate the multi-head outputs along the feature dimension to obtain the final output feature vector Z. This process includes the following sub-steps:

[0076] (1) Eigenvectors As input, it is fed into a multi-head attention layer.

[0077] (2) For each attention head, through a learnable weight matrix , , ,Will The linear mapping is defined as query vector Q, key vector K, and value vector V:

[0078] ;

[0079] K ;

[0080] V .

[0081] The purpose of this step is to transform the original feature vector Mapping to different representation spaces prepares for subsequent calculations of the correlation between features.

[0082] (3) Calculate the attention weight matrix A, where each element... This represents the correlation between the i-th feature and the j-th feature:

[0083] ;

[0084] in, "in "" indicates matrix multiplication. Matrix multiplication is used to calculate the similarity between the query vector Q and the key vector K. Each element in the resulting matrix reflects the degree of association between different features (represented by the dimensions corresponding to the query and key). is the dimension of the key vector; softmax is the normalization function used to normalize the similarity matrix so that each attention weight is between 0 and 1, and the sum of the weights in each row is 1.

[0085] (4) Use the attention weight matrix A to perform a weighted summation on the value vector V to obtain the output of the attention head. The purpose of this step is to aggregate the value vector based on the correlation between features (attention weights).

[0086] (5) Output vectors of all attention heads , ... (H represents the number of attention heads) are concatenated and passed through a final linear layer. Projecting the vectors, we obtain the final feature vector Z of the self-attention mechanism's output:

[0087] .

[0088] in, This represents the concatenation operation. Through concatenation, diverse feature information captured by different attention heads can be integrated. Combined with a linear layer for linear projection, the concatenated high-dimensional vector can be mapped to a suitable dimension, yielding the final feature representation.

[0089] (6) Input the feature vector Z, which is rich in global dependency information, into a fully connected regression layer.

[0090] (7) Fully connected regression layers use learnable weight vectors and bias A linear transformation is performed on the feature vector Z, mapping it to a scalar value, which is the final predicted value of the ammonium fluoride addition amount. :

[0091] .

[0092] In the above steps , , , , and All of these are obtained through iterative optimization during the model training phase by minimizing the loss function (such as mean squared error loss, MSE) between the predicted and true values.

[0093] In this embodiment, the model training phase specifically includes:

[0094] Based on the training dataset, a label indicating the actual amount of aluminum fluoride added is attached to the training dataset, and it is divided into a training set, a validation set, and a test set in a 6:2:2 ratio (in this embodiment, the ratio is set to 6:2:2). The mean squared error function is used as the loss function, and the prediction model is trained iteratively using the training set data. Each iteration includes the following sub-steps: forward propagation, back propagation, and parameter update. The specific steps of forward propagation, back propagation, and parameter update utilize existing technologies and will not be elaborated upon here.

[0095] The training process is monitored based on the validation set loss. When the loss no longer decreases, training is stopped and the model parameters with the best performance on the validation set are saved. The saved model is used to predict the test set to evaluate the final generalization performance of the model. Once the target is met, the model is deployed to the model training and prediction module.

[0096] The criteria for achieving the target are: the mean absolute error (MAE) on the test set is lower than a preset threshold; and the coefficient of determination on the test set is... The value is higher than a preset threshold two. Threshold one and threshold two are set according to the precision requirements of the addition amount in the aluminum electrolysis production process. In this embodiment, it can be set to... , .

[0097] With the above settings, this prediction model can optimize model parameters through gradient boosting decision trees and enhance its ability to capture long-distance dependencies between features using a self-attention mechanism, thereby improving prediction accuracy. Furthermore, the weight parameter matrix is ​​continuously iteratively optimized during the training process to ensure accurate prediction of daily aluminum fluoride addition.

[0098] The control execution module is used to convert the predicted addition amount into aluminum fluoride addition instructions and send them to the production line control system to realize the automatic execution of the addition task.

[0099] Specifically, in this embodiment, the aluminum fluoride addition instruction includes the addition dosage, the addition rate, and the addition duration.

[0100] When converting the predicted addition amount, the control execution module includes the following steps:

[0101] Predicted addition amount The solution is converted into three-dimensional control parameters: dosage D (in kg), feeding rate V (in kg / h), and feeding duration (in h).

[0102] Among them, the added dosage This means directly adopting the quality values ​​predicted by the model.

[0103] The feeding rate V needs to be set according to the type of electrolytic cell used in the actual application and the rated conveying capacity of the feeding actuator (e.g., the maximum conveying capacity corresponding to the maximum speed of the feeding motor).

[0104] The feeding duration ; ; n is the compensation coefficient, which is calibrated based on the actual electrolytic cell structure and feeding actuator, and is used to correct the actual feeding efficiency (such as pipeline resistance and material flowability) to ensure accurate addition.

[0105] After the three-dimensional control parameters are calculated, they are encapsulated into a standard data package and transmitted to the PLC controller of the feeding actuator (such as an aluminum fluoride feeder) in the production line control system to realize the automatic execution of the feeding task.

[0106] The control execution module is equipped with an exception response mechanism:

[0107] When the predicted value causes the molecular ratio to exceed the preset threshold for two consecutive rounds, the model recalculation and verification process is triggered, and the prediction model is re-called for verification. For example, the preset threshold is set to 2.2±0.1 (which can be adjusted according to actual application requirements). When the molecular ratio detection value of a certain electrolyzer exceeds this range for two consecutive rounds (with an interval of 2 hours between each round) (i.e., <2.1 or >2.3), the model recalculation and verification process is automatically triggered.

[0108] Preferably, the molecular ratio detection value for each round must meet the following requirements: the data is chemically confirmed by the Laboratory Information Management System (LIMS), and there are no major adjustments to the tank control parameters during the corresponding period (such as current intensity fluctuation ≤ ±50A, alumina feed deviation ≤ ±2%). Otherwise, it is determined to be an invalid threshold and the process is not triggered.

[0109] If the verification result is still abnormal, an alarm message will be automatically pushed to the designated terminal device, such as via SMS or email, to the terminal device of the on-duty personnel, so as to achieve real-time response and closed-loop management.

[0110] The control execution module also communicates with the production line PLC system in real time and synchronously activates the audible and visual alarm devices installed in the production line when an alarm is triggered. Specifically, in this embodiment, the control execution module is connected to the production line PLC system via industrial Ethernet or Modbus protocol.

[0111] The prediction model supports collaborative prediction and centralized scheduling of multiple electrolyzers. The control execution module achieves automated management of the entire production line's addition tasks through a unified control center, possessing excellent scalability and compatibility, and can be adapted to various types of electrolyzers and industrial control system platforms. Furthermore, in practical applications, the unified control center enables automated distribution and centralized monitoring of aluminum fluoride addition tasks. After applying the prediction model, the fluctuation amplitude of the electrolyte molecular ratio is significantly reduced, reaching 50% compared to manual adjustments, significantly improving current efficiency and achieving energy-saving and consumption-reducing goals.

[0112] The human-machine interface module provides a visual interface to display the electrolyzer's operating status, predicted additive amounts, and molecular ratio threshold alarm information in real time. This real-time visual interface allows operators to intuitively grasp the electrolyzer's status, predicted data, and alarm information, facilitating timely decision-making.

[0113] The human-computer interaction module also includes:

[0114] The multi-level access control unit is used to set and differentiate administrator configuration, operator viewing, and patrolman alarm receiving permissions;

[0115] The data interface unit provides an API for external MES / DCS systems to call and predict results. Integration with external MES / DCS systems via this API enables data exchange between production management systems, supporting higher levels of automation control.

[0116] The historical data traceability unit is used to support the analysis of molecular ratio fluctuation trends and the query of early warning records; it can provide data support for process optimization and fault diagnosis, forming a closed-loop management system from real-time monitoring to historical analysis.

[0117] By integrating visualization and multi-dimensional functional units, the efficiency and safety of electrolytic cell operation and management can be further improved.

[0118] like Figure 2 As shown, this embodiment also provides a method for predicting molecular ratio optimization in complex electrolyte systems. It applies the molecular ratio optimization control model for complex electrolyte systems described above to predict the molecular ratio optimization, including the following steps:

[0119] Step 1: Collect daily control data and inspection data of the tank, and generate basic data after preprocessing, which will serve as the training dataset.

[0120] Step 2: Use the random forest algorithm to select core features and construct a regression prediction model that incorporates a self-attention mechanism.

[0121] Specifically, the preprocessed dataset is input into the random forest feature selection unit to calculate the feature importance score of each parameter to the molecular ratio fluctuation.

[0122] The 10 features with the highest scores (including calcium fluoride content, aluminum level, quality level, mean current intensity, voltage variance, and temperature gradient value) are selected to generate a feature subset.

[0123] The feature subset is input into the regression prediction model that integrates the self-attention mechanism: first, a high-order feature vector representation is generated by the gradient boosting decision tree; then, the correlation matrix on the feature dimension is calculated based on the multi-head self-attention layer to generate weight allocation coefficients; the feature vectors are weighted and fused according to the weight coefficients to output the predicted value of aluminum fluoride addition.

[0124] Step 3: Set the molecular ratio threshold through the visual interface and output the predicted value of aluminum fluoride addition in real time.

[0125] Specifically, the system displays the predicted addition amount and real-time molecular ratio trend curve on a visual interface; it also receives the target molecular ratio threshold range set by the operator. .

[0126] Step 4: Convert the predicted values ​​into control commands and send them to the electrolytic cell control system to dynamically adjust the addition strategy.

[0127] Specifically, the approved predicted addition amount is converted into PLC control instructions; the control instructions include three-dimensional parameters: addition dosage, addition rate, and addition duration.

[0128] Step 5: When the predicted value causes the molecular ratio to exceed the preset threshold for two consecutive rounds, an alarm mechanism is activated, and the processing results are tracked in a closed loop.

[0129] When the molecular ratio corresponding to the predicted addition amount exceeds When the warning area is highlighted, the abnormal response mechanism will be activated.

[0130] Preferably, when the molecules detected in real time deviate from the values ​​of two consecutive periods When this happens, an alarm signal is triggered and the automatic addition program is frozen, switching to manual review mode.

[0131] This embodiment provides a molecular ratio optimization control model and prediction method for complex electrolyte systems. By collecting key process parameters and test data in aluminum electrolysis production, a high-precision prediction model is constructed. Combined with visual scheduling and real-time deployment, it can realize intelligent and automated control of the aluminum fluoride addition process, significantly improve prediction accuracy and response speed, effectively reduce electrolyte molecular ratio fluctuations, and improve current efficiency and energy utilization. It has good application and promotion value and is suitable for the energy-saving and consumption-reducing needs in modern aluminum electrolysis production.

[0132] In addition, to verify the application effect of this model and method, both the traditional method and this scheme were implemented on the same aluminum electrolysis production line for 2 months.

[0133] This solution includes: applying a molecular ratio optimization control model for a complex electrolyte system provided in this embodiment, automatically predicting the predicted value of aluminum fluoride addition by the prediction model, and then converting it into an aluminum fluoride addition instruction containing the addition dosage, addition rate and addition duration by the control execution module, thereby controlling the addition mechanism in the production line.

[0134] Traditional methods include: taking electrolyte samples periodically (e.g., every 4 hours), determining the molecular ratio, and then confirming the dosage based on experience, such as using empirical formulas. Then the operator adds it.

[0135] The obtained implementation data are shown in Table 1:

[0136] Table 1

[0137]

[0138] Wherein, the fluctuation range of the molecular ratio is the standard deviation of the measured values ​​of the molecular ratio within the statistical period, expressed as: The data comes from periodic chemical composition analysis results from the Laboratory Information Management System (LIMS).

[0139] The numerator ratio pass rate is the percentage of samples whose measured numerator ratio falls within the target control range (e.g., 2.2 ± 0.1) within a statistical period (2 months). The calculation formula is: .

[0140] The continuous stable duration is the duration during which the measured molecular ratio remains continuously within the target control range, calculated as the average of the continuous stable durations within the statistical period. The data is sourced from the timestamped historical records of the LIMS system.

[0141] The implementation data shows that after applying this solution, the accuracy of molecular ratio control was significantly improved, abnormal operating conditions were significantly suppressed (increased molecular ratio pass rate), and production stability was enhanced (smaller and more stable fluctuation range). Furthermore, compared to traditional methods, this solution offers higher control efficiency, eliminates the need for manual calculation of addition amounts, and enables automatic closed-loop control.

[0142] The above results fully verify the effectiveness of the design of each module in this scheme. By using random forest feature selection (accurately removing interference parameters) and XGBoost-self-attention fusion model (tree model extracts local features + attention mechanism dynamically weights global dependencies), the problem that traditional manual experience cannot capture strong coupling relationships of multiple variables is solved, which reduces the molecular ratio prediction error and increases the molecular ratio pass rate. In the control layer, the three-dimensional instruction solution (dose-rate-duration) is detailed, and combined with the abnormal response mechanism (continuous over-limit trigger verification), the effective control of molecular ratio fluctuations is achieved.

[0143] The above descriptions are merely embodiments of the present invention. Commonly known structures and characteristics of the solutions are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent.

Claims

1. A molecular ratio optimization control model for complex electrolyte systems, characterized in that, include: The data acquisition module is used to collect daily cell control reports and inspection data during the aluminum electrolysis production process as basic data. The preprocessing module is used to perform standardization, outlier removal, and missing value completion operations on the collected basic data. When performing outlier removal, the preprocessing module uses the interquartile range method to remove outliers, including the following operations: S1, Calculate the 25th percentile Q1 and 75th percentile Q3 for each type of data, and determine the interquartile range IQR = Q3 - Q1; S2, set the outlier detection range as [Q1-1.5×IQR, Q3+1.5×IQR], and remove data points that exceed this range; S3, repeat S1-S2 until there are no outliers in the basic data. At the same time, use a linear interpolation algorithm to fill in the missing data. The feature selection module is used to calculate the feature importance score based on the basic data using the random forest algorithm, and to select a preset number of core features. The model training and prediction module contains a prediction model. The prediction model adopts a regression prediction model that incorporates a self-attention mechanism. The prediction model takes the selected core features as input and outputs the predicted value of aluminum fluoride addition through dynamic weighted feature dependency relationship. The regression prediction model is a gradient boosting decision tree model; the self-attention mechanism of the prediction model is implemented in the following way: Receive the feature vectors generated by the gradient boosting decision tree; For a feature vector, the correlation weights between pairs of features are calculated in the feature dimension or the time dimension; based on the correlation weights, a dynamic weighted feature representation is generated through weighted fusion. The weighted feature representation is mapped to the predicted value of aluminum fluoride addition amount; The control execution module is used to convert the predicted addition amount into an aluminum fluoride addition command and send it to the production line control system to realize the automatic execution of the addition task; the aluminum fluoride addition command includes the addition dosage, the addition rate, and the addition duration; The human-computer interaction module provides a visual interface to display the electrolyzer's operating status, predicted addition amount, and molecular ratio threshold alarm information in real time. The human-computer interaction module also includes: The multi-level access control unit is used to set and differentiate administrator configuration, operator viewing, and patrolman alarm receiving permissions; The data interface unit is used to provide an API for external MES / DCS systems to call prediction results; The historical data traceability unit is used to support molecular ratio fluctuation trend analysis and early warning record query.

2. The molecular ratio optimization control model for a complex electrolyte system according to claim 1, characterized in that, The daily cell control report includes cell control parameters, which at least include current intensity, voltage, uniformity deviation, temperature, aluminum level, quality level, alumina feed rate, and furnace bottom pressure drop; the inspection data at least include calcium fluoride content, sodium oxide content, potassium fluoride content, iron content, silicon content, and molecular ratio.

3. The molecular ratio optimization control model for a complex electrolyte system according to claim 1, characterized in that, The core features selected by the feature selection module are the top 10 features in terms of feature importance score, including calcium fluoride content, aluminum level, quality level, current intensity, voltage, and temperature; and the hyperparameters of the random forest algorithm are tuned through grid search or Bayesian optimization methods.

4. The molecular ratio optimization control model for a complex electrolyte system according to claim 1, characterized in that, The control execution module is equipped with an exception response mechanism: When the predicted value causes the numerator ratio to exceed the preset threshold for two consecutive rounds, the model recalculation and verification process is triggered, and the prediction model is called again for verification. If the verification result is still abnormal, an alarm message will be automatically pushed to the designated terminal device; The control execution module also communicates with the production line PLC system in real time and simultaneously activates the audible and visual alarm device when an alarm is triggered.

5. The molecular ratio optimization control model for a complex electrolyte system according to claim 1, characterized in that, The prediction model supports collaborative prediction and centralized scheduling of multiple electrolytic cells, and the control execution module realizes automated management of adding tasks to the entire production line through a unified control center.

6. A method for predicting the molecular ratio optimization in complex electrolyte systems, characterized in that, Applying the molecular ratio optimization control model for a complex electrolyte system as described in any one of claims 1-5, the molecular ratio optimization prediction is performed, including the following steps: Collect daily control data and inspection data from the tank, and generate basic data after preprocessing to serve as the training dataset; We used the random forest algorithm to select core features and constructed a regression prediction model that incorporates a self-attention mechanism. The molecular ratio threshold can be set through a visual interface, and the predicted value of aluminum fluoride addition amount can be output in real time. The predicted values ​​are converted into control commands and sent to the electrolytic cell control system to dynamically adjust the addition strategy. When the predicted value causes the molecule ratio to exceed the preset threshold for two consecutive rounds, an alarm mechanism is activated, and the processing results are tracked in a closed loop.

Images