Method and device for predicting reference blanking interval of aluminum electrolysis aluminum oxide

Abstract

The invention discloses a method and a device for predicting a reference blanking interval of aluminum electrolysis aluminum oxide, relates to the technical field of aluminum electrolysis, and can solve the technical problem that the reference blanking interval is difficult to accurately determine. The method comprises the following steps: acquiring a multi-source data sample of a preset historical period in the aluminum electrolysis production process, and acquiring a real reference blanking interval of an end point moment of the preset historical period; performing feature extraction on the multi-source data sample, performing feature importance screening according to an extracted feature matrix to obtain a screened feature matrix of screened features, sorting the screened feature matrix into a time sequence sample according to a time sequence, and training an initial prediction model according to the time sequence sample and a real reference blanking interval to obtain a trained prediction model; and target multi-source data of a target historical time period in the aluminum electrolysis production process are obtained, a target time sequence is determined according to the target multi-source data, the target time sequence is predicted according to the prediction model, and the reference blanking interval of the end point moment of the target historical time period is obtained.

Description

Technical Field

[0001] This invention relates to the field of aluminum electrolysis technology, and in particular to a method and apparatus for predicting the baseline feeding interval for aluminum electrolysis of alumina. Background Technology

[0002] Aluminum electrolysis is the process of extracting metallic aluminum from alumina through electrolysis. To ensure stable operation of the electrolytic cell, improve current efficiency, reduce energy consumption costs, and enhance the quality of primary aluminum, it is necessary to control the alumina concentration, temperature, and molecular ratio within their respective optimal ranges to avoid significant fluctuations. Since temperature and molecular ratio cannot be directly controlled, but alumina concentration can be directly controlled, and alumina concentration, temperature, and molecular ratio are interdependent, the alumina concentration can be directly controlled by controlling the baseline feeding interval, while the temperature and molecular ratio can be indirectly controlled.

[0003] However, the baseline feeding interval is difficult to determine precisely. Specifically, aluminum electrolysis is a process with a large time lag. There is a significant time delay between the feeding of alumina and the corresponding response in the electrolyte concentration. Therefore, it is difficult to accurately control the baseline feeding interval based on the real-time alumina concentration in the electrolyte, thus making it impossible to achieve dynamic matching of the feeding frequency with the alumina consumption rate in the cell. Aluminum electrolysis is also a complex, multi-variable, coupled, and nonlinear process. The physicochemical reactions within the electrolytic cell are complex, with alumina concentration, temperature, and molecular ratio influencing each other. Adjusting a single parameter may cause fluctuations in other parameters, further affecting the original parameter. Moreover, there is a nonlinear relationship between alumina concentration, temperature, and molecular ratio, making it impossible to precisely control a single parameter using a fixed proportional relationship.

[0004] Currently, the method of controlling alumina feeding frequency by manually setting a fixed baseline feeding interval based on expert experience suffers from insufficient data utilization. Furthermore, this fixed baseline feeding interval cannot track the electrolytic cell's operating status in real time, limiting its adaptive capabilities. In addition, expert experience can only determine simple correlations between single factors and single results, failing to analyze the coupling relationship between alumina concentration, temperature, and molecular ratio in real time. For example, when alumina concentration decreases, it is difficult to distinguish whether the decrease is due to insufficient feeding or a change in solubility caused by temperature increase. Therefore, it is difficult to effectively address the complex coupling characteristics of multiple variables in aluminum electrolysis production, leading to fluctuations in alumina concentration, temperature, and molecular ratio, affecting the stability and efficiency of the electrolytic cell, and consequently impacting aluminum yield and production efficiency. Summary of the Invention

[0005] In view of this, the present invention provides a method and apparatus for predicting the reference feeding interval of aluminum electrolytic alumina, which can solve the technical problem of difficulty in accurately determining the reference feeding interval.

[0006] According to one aspect of the present invention, a method for predicting the baseline feeding interval for aluminum electrolytic alumina is provided, the method comprising: Obtain multi-source data samples during a preset historical period in the aluminum electrolysis production process, and obtain the actual benchmark feeding interval at the end of the preset historical period; Feature extraction is performed on the multi-source data samples to obtain an extracted feature matrix. Feature importance is then filtered based on the extracted feature matrix to obtain a filtered feature matrix. The filtered feature matrix is ​​then arranged into a time series sample according to the temporal order. An initial prediction model is constructed based on a neural network time series model. The initial prediction model is trained according to the time series sample and the actual baseline feeding interval until the initial prediction model converges, thus obtaining a trained prediction model. Acquire target multi-source data for a target historical period during the aluminum electrolysis production process, determine the target time series based on the target multi-source data, predict the target time series based on the prediction model, and obtain the baseline feeding interval for the end time of the target historical period.

[0007] Preferably, the acquisition of multi-source data samples for a preset historical period during the aluminum electrolysis production process includes: Acquire cell control data samples, production data samples, measurement data samples, and laboratory data samples for preset historical time periods during the aluminum electrolysis production process; The tank control data sample, the production data sample, the measurement data sample, and the laboratory data sample are identified as multi-source data samples.

[0008] Preferably, the step of extracting features from the multi-source data samples to obtain the extracted feature matrix includes: The multi-source data samples are divided into second-level data samples, minute-level data samples, and daily-level data samples. The second-level data samples include: second-level cell voltage, second-level cell current, and second-level feeding trigger signal. The minute-level data samples include: minute-level effective baseline feeding interval, minute-level current adjustment command, minute-level current deviation accumulation, minute-level voltage deviation, minute-level cumulative feeding amount, and minute-level cumulative alumina consumption. The daily-level data samples include: daily cumulative aluminum production, daily cumulative feeding amount, daily average current efficiency, daily average cell voltage, daily cumulative alumina consumption, cell age, daily measured electrolyte temperature, daily measured electrolyte level, daily measured aluminum liquid level, daily molecular ratio analysis, daily measured electrolyte composition analysis, daily measured alumina purity analysis, and daily measured alumina particle size analysis. Statistical calculations are performed on the second-level data samples to determine the first minute-level data sample, and the daily-level data sample is mapped to the second minute-level data sample. The minute-level data sample, the first minute-level data sample, and the second minute-level data sample are determined as the extracted feature matrix.

[0009] Preferably, the step of performing statistical calculations based on the second-level data samples to determine the first minute-level data sample includes: Determine the average, maximum, minimum, and standard deviation of all the second-level slot voltages within each minute, corresponding to the minute-level slot voltage average, minute-level slot voltage maximum, minute-level slot voltage minimum, and minute-level slot voltage standard deviation at the end of that minute. Determine the average, maximum, minimum, and standard deviation of all the second-level slot currents within each minute, corresponding to the minute-level slot current average, maximum, minimum, and standard deviation at the end of that minute. The number of all the second-level material feeding trigger signals within each minute is taken as the minute-level material feeding frequency at the end of that minute. The duration of each two adjacent second-level material feeding trigger signals within each minute is taken as the single material feeding duration. The average of all the single material feeding durations within each minute is calculated as the average minute-level single material feeding duration at the end of that minute. The average value of the minute-level cell voltage, the maximum value of the minute-level cell voltage, the minimum value of the minute-level cell voltage, the standard deviation of the minute-level cell voltage, the average value of the minute-level cell current, the maximum value of the minute-level cell current, the minimum value of the minute-level cell current, the standard deviation of the minute-level cell current, the minute-level feeding frequency, and the average value of the minute-level single feeding time are used as the first data sample at the minute level.

[0010] Preferably, the step of performing feature importance filtering based on the extracted feature matrix to obtain a filtered feature matrix includes: The extracted feature matrix is ​​input into the trained feature importance screening model, and the feature importance score of each extracted feature corresponding to the extracted feature matrix is ​​output. The extracted features whose importance scores are greater than a preset score threshold are used as the filtered features, and the extracted feature matrices corresponding to all the filtered features are used as the filtered feature matrix.

[0011] Preferably, the step of training the initial prediction model based on the time series samples and the actual baseline feeding interval until the initial prediction model converges to obtain the trained prediction model includes: Input the time series samples into the initial prediction model and output the prediction baseline feeding interval; The loss function is calculated based on the actual baseline feeding interval and the predicted baseline feeding interval. The model parameters of the initial prediction model are adjusted until the value of the loss function is less than a preset threshold. The initial prediction model is then determined to have converged, and the trained prediction model is obtained.

[0012] Preferably, the step of acquiring target multi-source data for a target historical period during the aluminum electrolysis production process, and determining the target time series based on the target multi-source data, includes: Acquire target cell control data, target production data, target measurement data, and target test data for the target historical period during the aluminum electrolysis production process; The target tank control data, the target production data, the target measurement data, and the target test data are identified as target multi-source data; Feature extraction is performed on the target multi-source data to obtain the target extracted feature matrix. The target extracted feature matrix corresponding to the filtered features is used as the target filtered feature matrix. The target filtered feature matrix is ​​then arranged into a target time series according to the time sequence.

[0013] According to another aspect of the present invention, a predictive device for the reference feeding interval of aluminum electrolytic alumina is provided, the device comprising: The acquisition module is used to acquire multi-source data samples of a preset historical period during the aluminum electrolysis production process, and to acquire the actual benchmark feeding interval at the end of the preset historical period. The training module is used to extract features from the multi-source data samples to obtain an extracted feature matrix, filter features based on the extracted feature matrix to obtain a filtered feature matrix, organize the filtered feature matrix into time series samples according to the time sequence, construct an initial prediction model based on a neural network time series model, train the initial prediction model according to the time series samples and the actual baseline feeding interval until the initial prediction model converges to obtain a trained prediction model. The prediction module is used to acquire target multi-source data for a target historical period during the aluminum electrolysis production process, determine the target time series based on the target multi-source data, predict the target time series based on the prediction model, and obtain the baseline feeding interval for the end time of the target historical period.

[0014] According to another aspect of the present invention, a storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the above-described method for predicting the reference feeding interval for aluminum electrolytic alumina.

[0015] According to another aspect of the present invention, a computer device is provided, including a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, wherein the processor, when executing the program, implements the above-described method for predicting the reference feeding interval of aluminum electrolytic alumina.

[0016] By employing the above technical solution, the present invention provides a method and apparatus for predicting the baseline feeding interval of aluminum electrolysis alumina. Through the technical solution of the present invention, multi-source data samples in the aluminum electrolysis production process are fully utilized to obtain a trained prediction model. The prediction model is used to make accurate predictions and output the baseline feeding interval in real time and intelligently. The feeding frequency is controlled according to the baseline feeding interval, avoiding reliance on expert experience and errors caused by human intervention, achieving real-time and precise control, and improving the automation of the production process.

[0017] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0018] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of this application. In the drawings: Figure 1 A flowchart illustrating a method for predicting the baseline feeding interval of aluminum electrolytic alumina according to an embodiment of the present invention is shown. Figure 2 A flowchart illustrating another method for predicting the baseline feeding interval of aluminum electrolytic alumina provided by an embodiment of the present invention is shown. Figure 3 A schematic diagram of the structure of a predictive device for the reference feeding interval of aluminum electrolytic alumina provided in an embodiment of the present invention is shown. Figure 4 A schematic diagram of the structure of another predictive device for the reference feeding interval of aluminum electrolytic alumina provided in an embodiment of the present invention is shown; Figure 5 The diagram illustrates a system architecture for predicting the baseline feeding interval for aluminum electrolytic alumina, as provided in an embodiment of the present invention. Detailed Implementation

[0019] The present invention will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0020] This embodiment provides a method for predicting the baseline feeding interval for aluminum electrolytic alumina, such as... Figure 1 As shown, the method includes: 101. Obtain multi-source data samples for a preset historical period during the aluminum electrolysis production process, and obtain the actual benchmark feeding interval at the end of the preset historical period.

[0021] It should be noted that excessive feeding can lead to problems. In high-concentration areas of the aluminum electrolysis cell, the localized high concentration can prevent the timely dissolution of alumina, resulting in precipitation and even the formation of a hard crust at the furnace bottom. This increases the horizontal current in the molten aluminum and affects the stability of the electrolysis cell. Conversely, insufficient feeding can increase the likelihood of flickering or even a full-cell effect due to locally low alumina concentrations, increasing energy consumption and reducing current efficiency. The periodic fluctuations in feeding frequency within the electrolysis cell can frequently impact electrolyte superheat, local thermal balance, and interface stability, resulting in reduced current efficiency. Therefore, determining the baseline feeding interval is crucial.

[0022] The Normal Break (NB) is the theoretical feeding interval in an aluminum electrolysis prebaked cell that matches the real-time aluminum production efficiency and alumina consumption rate of the electrolysis cell. It is the core benchmark of the feeding system, used to guide the intermittent feeding of alumina under normal operating conditions, and also serves as a reference benchmark for overfeeding and underfeeding.

[0023] In this embodiment, multi-source data is collected in real time during the aluminum electrolysis production process to obtain multi-source data for a preset historical period as a multi-source data sample. The start time of the preset historical period is a, the end time of the preset historical period is b, and the actual benchmark feeding interval at the end time of the preset historical period is obtained based on expert experience.

[0024] 102. Extract features from the multi-source data samples to obtain an extracted feature matrix. Filter the features based on their importance to obtain a filtered feature matrix. Arrange the filtered feature matrix into time series samples according to temporal order. Construct an initial prediction model based on a neural network time series model. Train the initial prediction model according to the time series samples and the actual baseline feeding interval until the initial prediction model converges to obtain a trained prediction model.

[0025] In this embodiment, the initial prediction model is constructed based on a neural network time series model. The neural network time series model includes Long Short Term Memory (LSTM) network and Transformer, etc. In one implementation, one of them is selected to construct the initial prediction model. In another implementation, an initial prediction model is constructed for each of them. The models are trained based on time series samples, and the optimal one is selected as the prediction model after training.

[0026] 103. Obtain target multi-source data for the target historical period during the aluminum electrolysis production process, determine the target time series based on the target multi-source data, predict the target time series based on the prediction model, and obtain the benchmark feeding interval for the end time of the target historical period.

[0027] In this embodiment, after obtaining the reference feeding interval at the end of the target historical period, the reference feeding interval can be input into the tank control system to control the feeding frequency of alumina using the reference feeding interval.

[0028] For example, if the starting time of the target historical period is 10:00 on December 1st and the ending time of the target historical period is 10:10 on December 1st, then the baseline material cutting interval for the ending time of the target historical period is the baseline material cutting interval for 10:10 on December 1st.

[0029] A preset time interval is set. After each preset time interval, the latest target historical period is collected again. Based on the latest target multi-source data for the latest target historical period, the latest target time series is obtained. Based on the latest target time series and the prediction model, the latest baseline feeding interval for the latest target historical period's endpoint is obtained. This latest baseline feeding interval is then input into the tank control system to control the alumina feeding frequency. For example, if the preset time interval is 10 minutes, after determining the baseline feeding interval for 10:10 on December 1st, the latest baseline feeding interval for 10:20 on December 1st needs to be determined. Therefore, using 10:10-10:20 on December 1st as the latest target historical period, the latest target multi-source data for 10:10-10:20 on December 1st is obtained, and finally, the latest baseline feeding interval for 10:20 on December 1st is determined.

[0030] The baseline feeding interval is updated at preset time intervals, which can accurately respond to complex changes in cell conditions. This effectively solves the problem of lag in traditional manual setting of the baseline feeding interval and avoids the negative impact on the stability of the electrolytic cell caused by fluctuations in alumina concentration and decrease in current efficiency due to untimely adjustments.

[0031] This invention provides a method and apparatus for predicting the baseline feeding interval of aluminum electrolysis alumina. Through the technical solution of this invention, multi-source data samples in the aluminum electrolysis production process are fully utilized to obtain a trained prediction model. The prediction model can accurately predict and output the baseline feeding interval in real time and intelligently. The feeding frequency is controlled according to the baseline feeding interval, avoiding reliance on expert experience and errors caused by human intervention, achieving real-time and precise control, and improving the automation of the production process.

[0032] Furthermore, as a refinement and extension of the specific implementation methods of the above embodiments, and to fully illustrate the specific implementation process in this embodiment, another method for predicting the benchmark feeding interval for aluminum electrolytic alumina is provided, such as... Figure 2 As shown, the method includes: 201. Obtain cell control data samples, production data samples, measurement data samples, and test data samples for a preset historical period during the aluminum electrolysis production process. Determine the cell control data samples, production data samples, measurement data samples, and test data samples as multi-source data samples, and obtain the true benchmark feeding interval at the end of the preset historical period.

[0033] Specifically, the process involves real-time acquisition of cell control data, production data, measurement data, and laboratory data during the aluminum electrolysis production process. These data are then identified as multi-source data. Furthermore, pre-defined historical time periods during the aluminum electrolysis production process are used as cell control data samples, production data samples, measurement data samples, and laboratory data samples. These cell control data samples, production data samples, measurement data samples, and laboratory data samples are then used as multi-source data samples.

[0034] 202. Perform feature extraction on the multi-source data samples to obtain the extracted feature matrix.

[0035] The step of extracting features from the multi-source data samples to obtain an extracted feature matrix includes: dividing the multi-source data samples into second-level data samples, minute-level data samples, and day-level data samples. The second-level data samples include: second-level cell voltage, second-level cell current, and second-level material feeding trigger signal. The minute-level data samples include: minute-level effective baseline material feeding interval, minute-level current adjustment command, minute-level current deviation accumulation, minute-level voltage deviation, minute-level cumulative material feeding amount, and minute-level cumulative alumina consumption. The day-level data samples include: daily cumulative aluminum production. The daily cumulative material feeding volume, daily average current efficiency, daily average cell voltage, daily cumulative alumina consumption, cell age, daily measured electrolyte temperature, daily measured electrolyte level, daily measured aluminum liquid level, daily molecular ratio, daily measured electrolyte composition, daily measured alumina purity, and daily measured alumina particle size are statistically calculated based on the second-level data samples to determine the first minute-level data sample, and the daily data samples are mapped to the second minute-level data sample; the minute-level data sample, the first minute-level data sample, and the second minute-level data sample are determined as the extracted feature matrix.

[0036] Specifically, some tank control data samples are divided into second-level data samples, some tank control data samples and some production data samples are divided into minute-level data samples, and some production data samples, all measurement data samples, and all test data samples are divided into day-level data samples.

[0037] Second-level data samples refer to data collected once per second, including second-level cell voltage, second-level cell current, and second-level feeding trigger signal. Second-level cell voltage reflects the stability of the electrolytic reaction; second-level cell current reflects current fluctuations, which directly affect heat release and electrolysis rate; and the second-level feeding trigger signal indicates the start and stop times of feeding, used to statistically analyze feeding frequency. Because second-level data samples contain a large amount of invalid noise, direct use would severely interfere with model training. Therefore, the second-level data samples are integrated into minute-level data, i.e., the first minute-level data sample. Specifically, statistical calculations are performed on the second-level data samples according to the minute-level data, converting them into minute-level data.

[0038] The step of statistically calculating based on the second-level data samples to determine the first minute-level data sample includes: determining the average, maximum, minimum, and standard deviation of all second-level slot voltages within each minute, corresponding to the minute-level slot voltage average, maximum, minimum, and standard deviation at the end of that minute; determining the average, maximum, minimum, and standard deviation of all second-level slot currents within each minute, corresponding to the minute-level slot current average, maximum, minimum, and standard deviation at the end of that minute; and counting the number of times all second-level material feeding trigger signals are generated within each minute. The minute-level material feeding frequency at the end of the minute is calculated by taking the duration of each two adjacent second-level material feeding trigger signals within each minute as the single material feeding duration, and calculating the average of all single material feeding durations within each minute as the average minute-level single material feeding duration at the end of the minute. The minute-level average cell voltage, the minute-level maximum cell voltage, the minute-level minimum cell voltage, the minute-level cell voltage standard deviation, the minute-level average cell current, the minute-level maximum cell current, the minute-level minimum cell current, the minute-level cell current standard deviation, the minute-level material feeding frequency, and the average minute-level single material feeding duration are used as the first minute-level data sample.

[0039] To map the daily-level data samples to minute-level second data samples, specifically, the daily cumulative aluminum production of the day in which the minute is located is determined as the minute-level cumulative aluminum production of that minute; the daily cumulative material feeding of the day in which the minute is located is determined as the minute-level cumulative material feeding of that minute; the daily average current efficiency of the day in which the minute is located is determined as the minute-level average current efficiency of that minute; the daily average cell voltage of the day in which the minute is located is determined as the minute-level average cell voltage of that minute; the daily cumulative alumina consumption of the day in which the minute is located is determined as the minute-level cumulative alumina consumption of that minute; the cell age of the day in which the minute is located is determined as the minute-level cell age of that minute; and the daily cumulative aluminum production of the day in which the minute is located is determined as the minute-level cumulative aluminum production of that minute. The daily measured electrolyte temperature is determined as the minute-level electrolyte temperature for that minute; the daily measured electrolyte level height for that minute is determined as the minute-level electrolyte level height for that minute; the daily measured aluminum liquid height for that minute is determined as the minute-level aluminum liquid height for that minute; the daily molecular ratio for that minute is determined as the minute-level molecular ratio for that minute; the daily measured electrolyte composition for that minute is determined as the minute-level electrolyte composition for that minute; the daily measured alumina purity for that minute is determined as the minute-level alumina purity for that minute; and the daily measured alumina particle size for that minute is determined as the minute-level alumina particle size for that minute.

[0040] The minute-level cumulative aluminum production, minute-level cumulative material feeding, minute-level average current efficiency, minute-level average cell voltage, minute-level cumulative alumina consumption, minute-level cell age, minute-level electrolyte temperature, minute-level electrolyte level, minute-level aluminum liquid height, minute-level molecular ratio, minute-level electrolyte composition, minute-level alumina purity, and minute-level alumina particle size are used as the minute-level second data sample.

[0041] 203. Based on the extracted feature matrix, feature importance is filtered to obtain the filtered feature matrix. The filtered feature matrix is ​​then arranged into time series samples according to the time sequence.

[0042] The step of performing feature importance filtering based on the extracted feature matrix to obtain the filtered feature matrix includes: inputting the extracted feature matrix into the trained feature importance filtering model, and outputting the feature importance score of the extracted feature corresponding to each extracted feature matrix; taking the extracted features with feature importance scores greater than a preset score threshold as filtered features, and taking the extracted feature matrices corresponding to all the filtered features as the filtered feature matrix.

[0043] The initial feature importance selection model is built based on Extreme Gradient Boosting (XGBoost). The initial feature importance selection model is trained to obtain the trained feature importance selection model.

[0044] The extracted feature vectors corresponding to each filtered feature are obtained from the extracted feature matrix. The filtered feature vectors are arranged in chronological order to obtain the chronological filtered feature vectors of the filtered features. The chronological filtered feature vectors of all the filtered features are concatenated to obtain the filtered feature matrix.

[0045] 204. Construct an initial prediction model based on a neural network time series model. Train the initial prediction model according to the time series samples and the actual baseline feeding interval until the initial prediction model converges, and obtain the trained prediction model.

[0046] As one implementation method, the neural network temporal model is a Long Short Term Memory (LSTM) network or a Transformer. An initial prediction model is built based on the LSTM network, trained to obtain a trained prediction model, and then the prediction model is used for the application in step 205 of the embodiment. Alternatively, an initial prediction model is built based on the Transformer, trained to obtain a trained prediction model, and then the prediction model is used for the application in step 205 of the embodiment.

[0047] The step of training the initial prediction model based on the time series samples and the true baseline material feeding interval until the initial prediction model converges to obtain a trained prediction model includes: inputting the time series samples into the initial prediction model and outputting the predicted baseline material feeding interval; calculating the value of the loss function based on the true baseline material feeding interval and the predicted baseline material feeding interval; adjusting the model parameters of the initial prediction model until the value of the loss function is less than a preset threshold; determining that the initial prediction model has converged; and obtaining a trained prediction model.

[0048] As another implementation, the neural network time series model is a Long Short Term Memory (LSTM) network and a Transformer. A first initial prediction model is built based on the LSTM network and trained to obtain a trained first prediction model. A second initial prediction model is built based on the Transformer and trained to obtain a trained second prediction model. For the trained first prediction model, the first root mean square error, the first coefficient of determination, and the first mean absolute error between the first predicted baseline material feeding interval and the actual baseline material feeding interval are calculated after inputting the time series samples of this embodiment. For the trained second prediction model, the second root mean square error, the second coefficient of determination, and the second mean absolute error between the second predicted baseline material feeding interval and the actual baseline material feeding interval are calculated after inputting the same time series samples of this embodiment.

[0049] Compare the first root mean square error (RMSE) with the second RMSE. If the first RMSE is less than the second RMSE, the trained first prediction model is optimal and is used as the trained prediction model. If the RMSE is greater than the second RMSE, the trained second prediction model is optimal and is used as the trained prediction model. If the first RMSE equals the second RMSE, compare the first coefficient of determination (COP) with the second COP. If the first COP is closest to 1, the trained first prediction model is optimal and is used as the trained prediction model. If the second COP is closest to 1, the trained second prediction model is optimal and is used as the trained prediction model. If the first COP equals the second COP, compare the first mean absolute error (MAE) with the second mean absolute error (MAE). If the first MAE is less than the second MAE, the trained first prediction model is optimal and is used as the trained prediction model. If the first MAE is greater than the second MAE, the trained second prediction model is optimal and is used as the trained prediction model.

[0050] 205. Obtain target multi-source data for the target historical period during the aluminum electrolysis production process, determine the target time series based on the target multi-source data, predict the target time series based on the prediction model, and obtain the benchmark feeding interval for the end time of the target historical period.

[0051] In this embodiment, the step of acquiring target multi-source data for a target historical period during the aluminum electrolysis production process and determining the target time series based on the target multi-source data includes: acquiring target cell control data, target production data, target measurement data, and target test data for a target historical period during the aluminum electrolysis production process; determining the target cell control data, the target production data, the target measurement data, and the target test data as target multi-source data; performing feature extraction on the target multi-source data to obtain a target extracted feature matrix; using the target extracted feature matrix corresponding to the selected features as the target selected feature matrix; and organizing the target selected feature matrix into a target time series according to the chronological order.

[0052] Among them, the target cell control data, target production data, target measurement data, and target test data corresponding to the target historical period of the aluminum electrolysis production process are acquired in real time, and the target cell control data, target production data, target measurement data, and target test data are identified as target multi-source data.

[0053] For feature extraction of the target multi-source data to obtain the target extracted feature matrix, the target extracted feature matrix corresponding to the filtered features is used as the target filtered feature matrix. Specifically: the target multi-source data is divided into target second-level data, target minute-level data, and target day-level data. Statistical calculations are performed based on the target second-level data to determine the target minute-level first data (the method is the same as the method for determining the minute-level first data sample based on the second-level data sample, and will not be repeated here). The target day-level data is mapped to the target minute-level second data (the method is the same as the method for mapping the day-level data sample to the minute-level second data sample, and will not be repeated here). The target minute-level data, the target minute-level first data, and the target minute-level second data are determined as the target extracted feature matrix. Since the filtered features are known, the target extracted feature matrix corresponding to the filtered features is used as the target filtered feature matrix.

[0054] The method of organizing the target-filtered feature matrix into a target time series according to the time sequence is the same as the method of organizing the filtered feature matrix into time series samples according to the time sequence, and will not be described again here.

[0055] Specifically, the target time series is predicted based on the prediction model to obtain the baseline material feeding interval for the end time of the target historical period. Specifically, the target time series is input into the prediction model, and the baseline material feeding interval for the end time of the target historical period is output.

[0056] like Figure 5 The diagram shows a system architecture for predicting the baseline feeding interval of aluminum electrolytic alumina. Through the combination of a data layer, a model layer, and an application layer, the system achieves the prediction of the baseline feeding interval for aluminum electrolytic alumina. Specifically: Data layer: (1) Collect multi-source data during the aluminum electrolysis production process, and obtain multi-source data samples for a preset historical period from the multi-source data. The multi-source data samples include cell control data samples, production data samples, measurement data samples, and laboratory data samples. Obtain target multi-source data for the target historical period from the multi-source data. The target multi-source data includes target cell control data, target production data, target measurement data, and target laboratory data.

[0057] (2) In the application service database, the multi-source data samples and target multi-source data are stored through the raw data layer. The multi-source data samples and target multi-source data are cleaned (removing null values, dirty data, data exceeding the range, filling missing values, etc.) through the data detail layer. The multi-source data samples are divided into second-level data samples, minute-level data samples, and day-level data samples through the data dimension layer. Since the second-level data samples contain a lot of invalid noise, direct use will seriously interfere with model training. Therefore, the second-level data samples are integrated into minute-level data, i.e., minute-level first data samples, and the day-level sample data are integrated into minute-level data, i.e., minute-level second data samples. From the minute-level data samples, minute-level first data samples, and minute-level second data samples, a unified time dimension extracted feature matrix is ​​obtained. In the same way, the same time dimension target extracted feature matrix is ​​obtained from the target multi-source data through the data dimension layer. The extracted feature matrix is ​​adapted to the input format of the initial prediction model through the data service layer.

[0058] Model layer: (1) Model library: Extreme Gradient Boosting (XGBoost) is used to rank features by importance, Long Short Term Memory (LSTM) is used to handle long sequence dependencies, effectively modeling the dynamic changes of electrolytic cells, and Transformer is used to capture global dependencies and complex interactions between multiple variables.

[0059] (2) Model training: On the one hand, the extracted feature matrix is ​​filtered based on the importance of the features using an extreme gradient boosting tree to obtain a filtered feature matrix of the selected features, so as to remove redundant features, and the target extracted feature matrix corresponding to the filtered features is used as the target filtered feature matrix. On the other hand, the initial prediction model based on the long short-term memory network is trained using the real benchmark feeding interval and the time series samples obtained by sorting the filtered feature matrix, and the initial prediction model based on the converter is trained using the time series samples and the real benchmark feeding interval. The optimal model is selected to obtain the trained prediction model.

[0060] (3) Model Deployment: The trained prediction model is deployed to the edge inference service via the cloud server. During the deployment process, the prediction model is compressed and optimized to adapt to the computing resource limitations of edge devices. The data layer periodically updates multi-source data to the model layer to dynamically update the trained prediction model. The edge inference service applies the trained prediction model to the field.

[0061] Application layer: (1) Model inference: This is used to input the target time series samples obtained by organizing the feature matrix after target screening into the prediction model to obtain the baseline feeding interval.

[0062] (2) Prediction display and evaluation: The prediction results, i.e. the baseline feeding interval, are displayed through the front-end interface, which facilitates real-time monitoring and decision-making by operators.

[0063] (3) Data Reports: Record the baseline feeding interval of the end time of the target historical period in chronological order and construct data reports.

[0064] (4) Display of important features: Display the features after filtering.

[0065] (5) System monitoring: Real-time monitoring of control effect, obtaining feedback data, namely alumina concentration, indirect control temperature and molecular ratio. If all are stable within the corresponding process optimal range, the prediction accuracy of the prediction model is high. If at least one is not stable within the corresponding process optimal range, multi-source data samples for training are added to update the training prediction model.

[0066] (6) Decision-making and control: The reference feeding interval is combined with the slot control system to control the feeding frequency.

[0067] This invention provides a method and apparatus for predicting the baseline feeding interval of aluminum electrolysis alumina. Through the technical solution of this invention, multi-source data samples in the aluminum electrolysis production process are fully utilized to obtain a trained prediction model. The prediction model can accurately predict and output the baseline feeding interval in real time and intelligently. The feeding frequency is controlled according to the baseline feeding interval, avoiding reliance on expert experience and errors caused by human intervention, achieving real-time and precise control, and improving the automation of the production process.

[0068] Furthermore, as Figure 1 and Figure 2 The specific implementation of the method shown in this embodiment of the invention provides a predictive device for the reference feeding interval of aluminum electrolytic alumina, such as... Figure 3 As shown, the device includes: an acquisition module 31, a training module 32, and a prediction module 33; The acquisition module 31 is used to acquire multi-source data samples of a preset historical period during the aluminum electrolysis production process, and to acquire the actual benchmark feeding interval at the end of the preset historical period. Training module 32 is used to extract features from the multi-source data samples to obtain an extracted feature matrix, filter features based on the extracted feature matrix to obtain a filtered feature matrix, organize the filtered feature matrix into time series samples according to the time sequence, construct an initial prediction model based on a neural network time series model, train the initial prediction model according to the time series samples and the actual baseline feeding interval until the initial prediction model converges to obtain a trained prediction model. The prediction module 33 is used to acquire target multi-source data for a target historical period during the aluminum electrolysis production process, determine the target time series based on the target multi-source data, predict the target time series based on the prediction model, and obtain the benchmark feeding interval for the end time of the target historical period.

[0069] Accordingly, in order to obtain multi-source data samples for a preset historical period during the aluminum electrolysis production process, the acquisition module 31 is specifically used to acquire cell control data samples, production data samples, measurement data samples, and laboratory data samples for a preset historical period during the aluminum electrolysis production process; and to determine the cell control data samples, the production data samples, the measurement data samples, and the laboratory data samples as multi-source data samples.

[0070] Accordingly, in order to extract features from the multi-source data samples and obtain the extracted feature matrix, the training module 32 is specifically used to divide the multi-source data samples into second-level data samples, minute-level data samples, and day-level data samples. The second-level data samples include: second-level cell voltage, second-level cell current, and second-level feeding trigger signal. The minute-level data samples include: minute-level effective baseline feeding interval, minute-level current adjustment command, minute-level current deviation accumulation, minute-level voltage deviation, minute-level cumulative feeding amount, and minute-level cumulative alumina consumption. The day-level data samples include... The daily cumulative aluminum production, daily cumulative material feeding, daily average current efficiency, daily average cell voltage, daily cumulative alumina consumption, cell age, daily measured electrolyte temperature, daily measured electrolyte level, daily measured aluminum liquid level, daily molecular weight ratio, daily measured electrolyte composition, daily measured alumina purity, and daily measured alumina particle size are analyzed. Based on the second-level data samples, statistical calculations are performed to determine the first minute-level data sample, and the daily data samples are mapped to the second minute-level data sample. The minute-level data sample, the first minute-level data sample, and the second minute-level data sample are then used to determine the extracted feature matrix.

[0071] Accordingly, in order to perform statistical calculations based on the second-level data samples and determine the first minute-level data sample, the training module 32 is specifically used to determine the average, maximum, minimum, and standard deviation of all the second-level slot voltages within each minute, corresponding to the minute-level slot voltage average, maximum, minimum, and standard deviation at the end of that minute; determine the average, maximum, minimum, and standard deviation of all the second-level slot currents within each minute, corresponding to the minute-level slot current average, maximum, minimum, and standard deviation at the end of that minute; and set all the second-level material feeding contacts within each minute... The number of signal transmissions is taken as the minute-level material feeding frequency at the end of the minute. The duration of each two adjacent second-level material feeding trigger signals within each minute is taken as the single material feeding duration. The average of all single material feeding durations within each minute is calculated as the minute-level average single material feeding duration at the end of the minute. The minute-level average cell voltage, the minute-level maximum cell voltage, the minute-level minimum cell voltage, the minute-level cell voltage standard deviation, the minute-level average cell current, the minute-level maximum cell current, the minute-level minimum cell current, the minute-level cell current standard deviation, the minute-level material feeding frequency, and the minute-level average single material feeding duration are taken as the minute-level first data sample.

[0072] Accordingly, in order to perform feature importance filtering based on the extracted feature matrix and obtain the filtered feature matrix, the training module 32 is specifically used to input the extracted feature matrix into the trained feature importance filtering model, output the feature importance score of the extracted feature corresponding to each extracted feature matrix; take the extracted features with feature importance scores greater than a preset score threshold as the filtered features, and take the extracted feature matrices corresponding to all the filtered features as the filtered feature matrix.

[0073] Accordingly, in order to train the initial prediction model based on the time series samples and the true baseline feeding interval until the initial prediction model converges and a trained prediction model is obtained, the training module 32 is specifically used to input the time series samples into the initial prediction model and output the prediction baseline feeding interval; calculate the value of the loss function based on the true baseline feeding interval and the prediction baseline feeding interval, adjust the model parameters of the initial prediction model until the value of the loss function is less than a preset threshold, determine that the initial prediction model has converged, and obtain a trained prediction model.

[0074] Accordingly, in order to obtain target multi-source data for the target historical period during the aluminum electrolysis production process and determine the target time series based on the target multi-source data, the prediction module 33 includes: an acquisition unit 331 and a processing unit 332. The acquisition unit 331 is specifically used to acquire target cell control data, target production data, target measurement data, and target test data for a target historical period during the aluminum electrolysis production process; and to determine the target cell control data, the target production data, the target measurement data, and the target test data as target multi-source data; The sorting unit 332 is specifically used to extract features from the target multi-source data to obtain a target extracted feature matrix, take the target extracted feature matrix corresponding to the filtered features as the target filtered feature matrix, and sort the target filtered feature matrix into a target time series according to the time sequence.

[0075] It should be noted that for other corresponding descriptions of the functional units involved in the predictive device for the reference feeding interval of aluminum electrolytic alumina provided in this embodiment, please refer to... Figures 1 to 2 The corresponding description will not be repeated here.

[0076] Based on the above, Figures 1 to 2 Accordingly, this embodiment also provides a storage medium, which may be volatile or non-volatile, storing a computer program that, when executed by a processor, implements the above-described method. Figures 1 to 2 The method for predicting the baseline feeding interval for aluminum electrolytic alumina is shown.

[0077] Based on this understanding, the technical solution of the present invention can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, portable hard drive, etc.) and includes several instructions to cause a computer device (such as a personal computer, server, or network device, etc.) to execute the methods of various implementation scenarios of the present invention.

[0078] Based on the above, Figures 1 to 2 The method shown and Figure 3 , Figure 4 To achieve the above objectives, the present application also provides a computer device, specifically a personal computer, server, network device, etc., as shown in the illustrated embodiment. This computer device includes a storage medium and a processor; the storage medium stores a computer program; the processor executes the computer program to achieve the above-described objectives. Figure 1 and Figure 2 The method for predicting the baseline feeding interval for aluminum electrolytic alumina is shown.

[0079] Optionally, the computer device may also include a user interface, a network interface, a camera, radio frequency (RF) circuitry, sensors, audio circuitry, a Wi-Fi module, etc. The user interface may include a display screen, input units such as a keyboard, etc., and optional user interfaces may also include USB interfaces, card reader interfaces, etc. The network interface may optionally include standard wired interfaces, wireless interfaces (such as Wi-Fi interfaces), etc.

[0080] Those skilled in the art will understand that the computer device structure provided in this embodiment does not constitute a limitation on the physical device, and may include more or fewer components, or combine certain components, or have different component arrangements.

[0081] The storage medium may also include an operating system and a network communication module. The operating system is a program that manages the hardware and software resources of the aforementioned computer device, supporting the operation of information processing programs and other software and / or programs. The network communication module is used to enable communication between the various components within the non-volatile storage medium, as well as communication with other hardware and software in the information processing entity device.

[0082] Through the above description of the embodiments, those skilled in the art can clearly understand that the present invention can be implemented by means of software plus necessary general-purpose hardware platform, or it can be implemented by hardware.

[0083] This invention provides a method and apparatus for predicting the baseline feeding interval of aluminum electrolysis alumina. Through the technical solution of this invention, multi-source data samples in the aluminum electrolysis production process are fully utilized to obtain a trained prediction model. The prediction model can accurately predict and output the baseline feeding interval in real time and intelligently. The feeding frequency is controlled according to the baseline feeding interval, avoiding reliance on expert experience and errors caused by human intervention, achieving real-time and precise control, and improving the automation of the production process.

[0084] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing the present invention. Those skilled in the art will understand that the modules in the apparatus of the embodiment can be distributed within the apparatus of the embodiment as described, or they can be located in one or more apparatuses different from this embodiment, with corresponding changes. The modules of the above-described embodiment can be combined into one module, or further divided into multiple sub-modules.

[0085] The serial numbers used above are for descriptive purposes only and do not represent the superiority or inferiority of the implementation scenarios. The above disclosures are merely a few specific implementation scenarios of the present invention; however, the present invention is not limited thereto, and any variations conceived by those skilled in the art should fall within the protection scope of the present invention.

Claims

1. A method for predicting the baseline feeding interval in aluminum electrolytic alumina production, characterized in that, The method includes: Obtain multi-source data samples during a preset historical period in the aluminum electrolysis production process, and obtain the actual benchmark feeding interval at the end of the preset historical period; Feature extraction is performed on the multi-source data samples to obtain an extracted feature matrix. Feature importance is then filtered based on the extracted feature matrix to obtain a filtered feature matrix. The filtered feature matrix is ​​then arranged into a time series sample according to the temporal order. An initial prediction model is constructed based on a neural network time series model. The initial prediction model is trained according to the time series sample and the actual baseline feeding interval until the initial prediction model converges, thus obtaining a trained prediction model. Acquire target multi-source data for a target historical period during the aluminum electrolysis production process, determine the target time series based on the target multi-source data, predict the target time series based on the prediction model, and obtain the baseline feeding interval for the end time of the target historical period.

2. The method according to claim 1, characterized in that, The acquisition of multi-source data samples for a preset historical period during the aluminum electrolysis production process includes: Acquire cell control data samples, production data samples, measurement data samples, and laboratory data samples for preset historical time periods during the aluminum electrolysis production process; The tank control data sample, the production data sample, the measurement data sample, and the laboratory data sample are identified as multi-source data samples.

3. The method according to claim 1, characterized in that, The step of extracting features from the multi-source data samples to obtain the extracted feature matrix includes: The multi-source data samples are divided into second-level data samples, minute-level data samples, and daily-level data samples. The second-level data samples include: second-level cell voltage, second-level cell current, and second-level feeding trigger signal. The minute-level data samples include: minute-level effective baseline feeding interval, minute-level current adjustment command, minute-level current deviation accumulation, minute-level voltage deviation, minute-level cumulative feeding amount, and minute-level cumulative alumina consumption. The daily-level data samples include: daily cumulative aluminum production, daily cumulative feeding amount, daily average current efficiency, daily average cell voltage, daily cumulative alumina consumption, cell age, daily measured electrolyte temperature, daily measured electrolyte level, daily measured aluminum liquid level, daily molecular ratio analysis, daily measured electrolyte composition analysis, daily measured alumina purity analysis, and daily measured alumina particle size analysis. Statistical calculations are performed on the second-level data samples to determine the first minute-level data sample, and the daily-level data sample is mapped to the second minute-level data sample. The minute-level data sample, the first minute-level data sample, and the second minute-level data sample are determined as the extracted feature matrix.

4. The method according to claim 3, characterized in that, The step of performing statistical calculations based on the second-level data samples to determine the first minute-level data sample includes: Determine the average, maximum, minimum, and standard deviation of all the second-level slot voltages within each minute, corresponding to the minute-level slot voltage average, minute-level slot voltage maximum, minute-level slot voltage minimum, and minute-level slot voltage standard deviation at the end of that minute. Determine the average, maximum, minimum, and standard deviation of all the second-level slot currents within each minute, corresponding to the minute-level slot current average, maximum, minimum, and standard deviation at the end of that minute. The number of all the second-level material feeding trigger signals within each minute is taken as the minute-level material feeding frequency at the end of that minute. The duration of each two adjacent second-level material feeding trigger signals within each minute is taken as the single material feeding duration. The average of all the single material feeding durations within each minute is calculated as the average minute-level single material feeding duration at the end of that minute. The average value of the minute-level cell voltage, the maximum value of the minute-level cell voltage, the minimum value of the minute-level cell voltage, the standard deviation of the minute-level cell voltage, the average value of the minute-level cell current, the maximum value of the minute-level cell current, the minimum value of the minute-level cell current, the standard deviation of the minute-level cell current, the minute-level feeding frequency, and the average value of the minute-level single feeding time are used as the first data sample at the minute level.

5. The method according to claim 1, characterized in that, The step of performing feature importance filtering based on the extracted feature matrix to obtain the filtered feature matrix includes: The extracted feature matrix is ​​input into the trained feature importance screening model, and the feature importance score of each extracted feature corresponding to the extracted feature matrix is ​​output. The extracted features whose importance scores are greater than a preset score threshold are used as the filtered features, and the extracted feature matrices corresponding to all the filtered features are used as the filtered feature matrix.

6. The method according to claim 1, characterized in that, The step of training the initial prediction model based on the time series samples and the actual baseline feeding interval until the initial prediction model converges, to obtain the trained prediction model, includes: Input the time series samples into the initial prediction model and output the prediction baseline feeding interval; The loss function is calculated based on the actual baseline feeding interval and the predicted baseline feeding interval. The model parameters of the initial prediction model are adjusted until the value of the loss function is less than a preset threshold. The initial prediction model is then determined to have converged, and the trained prediction model is obtained.

7. The method according to claim 1, characterized in that, The acquisition of target multi-source data for a target historical period during the aluminum electrolysis production process, and the determination of the target time series based on the target multi-source data, includes: Acquire target cell control data, target production data, target measurement data, and target test data for the target historical period during the aluminum electrolysis production process; The target tank control data, the target production data, the target measurement data, and the target test data are identified as target multi-source data; Feature extraction is performed on the target multi-source data to obtain the target extracted feature matrix. The target extracted feature matrix corresponding to the filtered features is used as the target filtered feature matrix. The target filtered feature matrix is ​​then arranged into a target time series according to the time sequence.

8. A device for predicting the baseline feeding interval of aluminum electrolytic alumina, characterized in that, The device includes: The acquisition module is used to acquire multi-source data samples of a preset historical period during the aluminum electrolysis production process, and to acquire the actual benchmark feeding interval at the end of the preset historical period. The training module is used to extract features from the multi-source data samples to obtain an extracted feature matrix, filter features based on the extracted feature matrix to obtain a filtered feature matrix, organize the filtered feature matrix into time series samples according to the time sequence, construct an initial prediction model based on a neural network time series model, train the initial prediction model according to the time series samples and the actual baseline feeding interval until the initial prediction model converges to obtain a trained prediction model. The prediction module is used to acquire target multi-source data for a target historical period during the aluminum electrolysis production process, determine the target time series based on the target multi-source data, predict the target time series based on the prediction model, and obtain the benchmark feeding interval for the end time of the target historical period.

9. A storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method for predicting the reference feeding interval of aluminum electrolytic alumina as described in any one of claims 1 to 7.

10. A computer device comprising a memory, a processor, and a computer program stored on a storage medium and executable on the processor, characterized in that, When the processor executes the program, it implements the method for predicting the baseline feeding interval of aluminum electrolytic alumina as described in any one of claims 1 to 7.

Images