A CRNN-based model for identifying and predicting aluminum electrolysis parameters and its construction method.
By constructing a multi-field coupling model of the electrolytic cell and CRNN, the problem of difficult parameter measurement in the aluminum electrolysis process was solved, and intelligent identification and prediction of key parameters were realized, thereby improving production stability and efficiency.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- KUNMING UNIV OF SCI & TECH
- Filing Date
- 2022-06-10
- Publication Date
- 2026-06-02
Smart Images
Figure HDA0003689035250000011 
Figure HDA0003689035250000021
Abstract
Description
Technical Field
[0001] This invention relates to a CRNN-based aluminum electrolysis parameter identification and prediction model and its construction method, and relates to the field of aluminum electrolysis parameter prediction technology. Background Technology
[0002] The aluminum electrolysis process is an electrochemical reaction that is difficult to analyze in isolation. It is characterized by extremely high spatial complexity and variability, and the process is not easily manipulated. Most known correlations between parameters are based on long-term experience, and parameter control in electrolysis engineering relies heavily on trial and error. Parameter combinations are relatively fixed, and there are no scientific standards for parameter control. Therefore, the scientific and effective adjustment of parameters has always been a key focus in the aluminum electrolysis industry. For example, electrolyte temperature, alumina concentration, and electrode distance within the aluminum electrolysis cell are three crucial process parameters that are difficult to measure online, and they are also key parameters determining the stability of multiphysics models.
[0003] Extensive research has been conducted on the aluminum electrolysis process. Li Jiejia et al. proposed an anomaly prediction method using parallel computation of multiple models to address the detection and prediction of abnormal states in aluminum electrolysis production. Zeng Shuiping et al. used a neural network in a random forest, employing aluminum electrolysis production parameters as input to the algorithm model and the molecular ratio as the output. Finally, the neural network was fitted to a classification feature set, solving the problem of inaccurate prediction of the molecular ratio. Xu Chenhua et al. proposed a cell voltage optimization algorithm that uses a genetic algorithm to find the production parameter with the lowest cell voltage. Yao Lizhong et al. proposed a model based on an unscented Kalman neural network algorithm. This method first establishes an energy-saving model for the aluminum electrolysis process using traditional algorithms. Then, the improved method is used to test the connection weights in the model, proving that the energy-saving model can significantly reduce energy consumption in aluminum electrolysis production. Therefore, utilizing machine learning algorithms to improve industrial capacity and save costs has become a new trend.
[0004] In response, the applicant conducted a physical and chemical mechanism analysis and used software such as ANSYS to construct a multi-field coupling model of the electrolytic cell. Then, by combining measurable data and introducing a convolutional recurrent neural network, a soft measurement model for three parameters, namely temperature, concentration and electrode distance, was established. This model can obtain a predictive model that meets the performance requirements and can be applied to industrial production. Summary of the Invention
[0005] The purpose of this invention is to provide a CRNN-based aluminum electrolysis parameter identification and prediction model and construction method. By constructing a multi-field coupling model of the electrolytic cell, and then combining measurable data to introduce a convolutional recurrent neural network, a soft measurement model of three parameters, namely temperature, concentration and pole distance, is established to obtain a prediction model that meets performance requirements and can be applied to industrial production.
[0006] To achieve the above-mentioned technical objectives and effects, the present invention is implemented through the following technical solution:
[0007] A method for constructing a parameter identification and prediction model for aluminum electrolysis based on CRNN includes the following steps:
[0008] S1: Through physicochemical mechanism analysis, the relationship between the unmeasurable variables electrolyte temperature, alumina concentration, and electrode distance and the measurable variables is obtained;
[0009] A multi-field coupling model of an electrolyzer was constructed using software, and the model was calibrated using measured data. Then, a multi-factor analysis method was used to determine the dependent variables affecting these parameters.
[0010] S2: Using the dependent variables of temperature, concentration and electrode distance obtained in S1 that affect the aluminum electrolysis production process, a recurrent neural network model is constructed, and a massive amount of historical data recorded in the aluminum electrolysis production process is used as training samples to finally construct a key process parameter identification and prediction model.
[0011] The recurrent neural network model consists of an input layer, a one-dimensional convolutional layer, a maximum pooling layer, a long short-term memory (LSTM) layer, a fully connected layer, and an output layer.
[0012] The input layer: Auxiliary variable data of important process parameters such as temperature, concentration and electrode distance in the aluminum electrolysis production process are entered into the network through the input layer;
[0013] The one-dimensional convolutional layer: performs deep feature extraction on the data through convolutional kernels;
[0014] The maximum confluence layer: performs subsequence extraction and maximum value calculation on the output data of the convolutional layer;
[0015] The Long Short-Term Memory (LSTM) layer processes the deep feature sequences after processing.
[0016] The fully connected layer: outputs prediction results of important process parameters such as temperature, concentration and electrode distance in the aluminum electrolysis production process through the LSTM layer;
[0017] The output layer outputs predicted values of temperature, concentration, and electrode spacing during the aluminum electrolysis production process.
[0018] S3: Train the recurrent neural network model using actual operating data from the aluminum electrolysis plant to obtain a predictive model that meets performance requirements.
[0019] Furthermore, the electrolyte temperature, in the auxiliary variables selected in the aluminum electrolysis cell temperature sensing model, is: fluoride salt content, aluminum level, electrolyte level, cell voltage, molecular ratio, and aluminum output; the output is the electrolysis temperature; the determined variables are to prepare for subsequent recurrent neural network data fusion.
[0020] Furthermore, regarding the alumina concentration, a mapping model is first established between the instantaneous cell voltage value, alumina concentration, and the feeding interval NB value. The influence factors of processes such as aluminum removal and anode replacement on the alumina concentration are introduced. Then, the model is calibrated using measurable data, and soft measurement of the alumina concentration is then implemented.
[0021] Furthermore, the alumina concentration is addressed by employing multi-factor coupling to resolve the measurement error problem caused by the nonlinear relationship between alumina concentration and cell resistance in existing studies.
[0022] Furthermore, the electrode spacing considers the relationship between the remaining electrode spacings in terms of measurable cell voltage and anode resistance; the cell voltage, guide rod equidistant voltage drop, guide rod temperature, and anode resistance, which can be directly measured online, are used as auxiliary variables (the guide rod temperature and anode resistance are determined by the number of days the anode is in operation, so the relationship between the number of days the anode is replaced and the anode guide rod temperature and anode resistance can be found through experiments).
[0023] Furthermore, the data collection for the prediction model is subject to interference from various factors, and the collection time scale and data structure are different, resulting in multi-source heterogeneous data. Therefore, there are problems such as outliers, default values, random noise, abnormal values, missing values, incompleteness, and inconsistency. In the modeling process, intelligent data preprocessing is required, and normalization, interpolation, and noise reduction are performed to remove outliers, default values, and random noise.
[0024] The beneficial effects of this invention are:
[0025] The present invention provides a method for constructing a CRNN-based aluminum electrolysis parameter identification and prediction model. By constructing a multi-field coupling model of the electrolytic cell, and then combining measurable data to introduce a convolutional recurrent neural network, a soft measurement model of three parameters, namely temperature, concentration and electrode distance, is established to obtain a prediction model that meets performance requirements and can be applied to industrial production.
[0026] The present invention describes a method for constructing a CRNN-based aluminum electrolysis parameter identification and prediction model. First, through physicochemical mechanism analysis, the relationship between unmeasurable variables (electrolyte temperature, alumina concentration, and electrode distance) and measurable variables is derived and analyzed. Second, a recurrent neural network model is constructed by establishing the dependent variables affecting temperature, concentration, and electrode distance during aluminum electrolysis production. Massive historical data from the aluminum electrolysis production process is used as training samples to ultimately construct a key process parameter identification and prediction model. The model combines collected operational data to model important influencing mechanisms and determine model parameters. Finally, based on the derived model and big data analysis, intelligent identification of unmeasurable variables is achieved.
[0027] The present invention provides a method for constructing a CRNN-based aluminum electrolysis parameter identification and prediction model. After outputting predicted values of temperature, concentration, and electrode distance during the aluminum electrolysis production process, the CRNN model is trained using actual operating data from the aluminum electrolysis plant to obtain a prediction model that meets performance requirements.
[0028] This invention introduces machine learning algorithms to improve industrial capacity and save costs, making it suitable for industrial applications.
[0029] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description
[0030] Figure 1 This is a flowchart illustrating the coupling mechanism and determining factors among the various process parameters described in the embodiments of the present invention;
[0031] Figure 2 This is a flowchart of the aluminum electrolysis parameter identification and prediction model based on CRNN as described in an embodiment of the present invention; Detailed Implementation
[0032] To more clearly illustrate the technical solutions of the embodiments of the present invention, the present invention will be described in detail below with reference to the accompanying drawings.
[0033] A method for constructing a parameter identification and prediction model for aluminum electrolysis based on CRNN includes the following steps:
[0034] S1: Through physicochemical mechanism analysis, the relationship between the unmeasurable variables electrolyte temperature, alumina concentration, and electrode distance and the measurable variables is obtained;
[0035] A multi-field coupling model of an electrolyzer was constructed using software, and the model was calibrated using measured data. Then, a multi-factor analysis method was used to determine the dependent variables affecting these parameters.
[0036] S2: Using the dependent variables of temperature, concentration and electrode distance obtained in S1 that affect the aluminum electrolysis production process, a recurrent neural network model is constructed, and a massive amount of historical data recorded in the aluminum electrolysis production process is used as training samples to finally construct a key process parameter identification and prediction model.
[0037] The recurrent neural network model consists of an input layer, a one-dimensional convolutional layer, a maximum pooling layer, a long short-term memory (LSTM) layer, a fully connected layer, and an output layer.
[0038] The input layer: Auxiliary variable data of important process parameters such as temperature, concentration and electrode distance in the aluminum electrolysis production process are entered into the network through the input layer;
[0039] The one-dimensional convolutional layer: performs deep feature extraction on the data through convolutional kernels;
[0040] The maximum confluence layer: performs subsequence extraction and maximum value calculation on the output data of the convolutional layer;
[0041] The Long Short-Term Memory (LSTM) layer processes the deep feature sequences after processing.
[0042] The fully connected layer: outputs prediction results of important process parameters such as temperature, concentration and electrode distance in the aluminum electrolysis production process through the LSTM layer;
[0043] The output layer outputs predicted values of temperature, concentration, and electrode spacing during the aluminum electrolysis production process.
[0044] S3: Train the recurrent neural network model using actual operating data from the aluminum electrolysis plant to obtain a predictive model that meets performance requirements.
[0045] The present invention will now be described in conjunction with specific embodiments:
[0046] Example 1
[0047] Coupled analysis of key process parameters such as electrolyte temperature, alumina concentration, and electrode spacing
[0048] First, based on the physicochemical mechanism of aluminum electrolysis, the relationships between key characteristics and process parameters related to the production state of aluminum electrolysis are analyzed. The relationships between key unmeasurable variables such as electrolyte temperature, alumina concentration, and electrode distance, and measurable variables are explored. An electrolytic cell model is constructed using ANSYS software, and the model is calibrated using measured data. Then, a multi-factor analysis method is used to determine the dependent variables affecting these parameters (such as…). Figure 1 (As shown). For electrolysis temperature identification, the auxiliary variables to be initially selected in the aluminum electrolysis cell temperature sensing model are: fluoride salt content, aluminum level, electrolyte level, cell voltage, molecular ratio, and aluminum output; the output is the electrolysis temperature. Determining these variables prepares for subsequent recurrent neural network data fusion.
[0049] To address the issue of the inability to continuously measure alumina concentration online, this paper considers that some characteristics of alumina concentration can be reflected by the instantaneous value of the cell voltage, which is affected by the cell feeding interval (NB value). The proposed approach is to first establish a mapping model between the instantaneous cell voltage, alumina concentration, and feeding interval (NB value), and to introduce the influence factors of processes such as aluminum tapping and anode replacement on alumina concentration. Then, the model will be calibrated using measurable data, leading to soft measurement of alumina concentration. A multi-factor coupling approach is proposed to address the measurement error caused by the nonlinear relationship between alumina concentration and cell resistance in existing research.
[0050] To overcome the difficulties of high temperature and strong corrosion in direct electrode spacing measurement, it is proposed to consider the relationship between other electrode spacings from the perspectives of measurable cell voltage and anode resistance. The cell voltage, equidistant voltage drop of the guide rod, guide rod temperature, and anode resistance, which can be directly measured online, will be considered as auxiliary variables (the guide rod temperature and anode resistance are determined by the number of days the anode operates; therefore, the relationship between the number of days the anode is replaced and the anode guide rod temperature and anode resistance can be found through experiments).
[0051] Example 2
[0052] Intelligent sensing and identification of key process parameters in aluminum electrolysis
[0053] Based on the identification of key process parameters influencing factors according to the mechanism, the important influencing mechanisms will be modeled and model parameters determined by combining the collected operational data. Then, based on the derived model and big data analysis, intelligent identification of unmeasurable variables will be achieved. The work to be carried out here will be as follows: on the one hand, based on the aforementioned mechanism model and using operational data to calibrate the model, and on the other hand, parameter identification and recognition methods will be studied based on the model; on the other hand, recurrent neural networks will be introduced to effectively compensate for model errors, ultimately realizing data-driven identification of key process parameters.
[0054] In this study, a recurrent neural network (CRNN) was constructed using auxiliary variables that influence important process parameters such as temperature, concentration, and electrode distance during aluminum electrolysis production. A massive amount of historical data recording the aluminum electrolysis production process was used as training samples to ultimately build a key process parameter identification and prediction model. This CRNN model consists of an input layer, a one-dimensional convolutional layer, a maximum pooling layer, a long short-term memory (LSTM) layer, a fully connected layer, and an output layer. The structure of the CRNN-based soft measurement prediction model for aluminum electrolysis parameters is as follows: Figure 2 As shown.
[0055] The functions of each layer are as follows: Input layer - Auxiliary variable data of important process parameters such as temperature, concentration, and electrode distance in aluminum electrolysis production are input into the network through the input layer; One-dimensional convolutional layer - Deep feature extraction is performed on the data through convolutional kernels; Maximum merge layer - Subsequence extraction and maximum value calculation are performed on the output data of the convolutional layer; LSTM layer - The processed deep feature sequences are processed; Connection layer - The prediction results of important process parameters such as temperature, concentration, and electrode distance in aluminum electrolysis production are calculated through the output of the LSTM layer; Output layer - The predicted values of temperature, concentration, and electrode distance in aluminum electrolysis production are output. Then, the CRNN model is trained using actual operating data of the aluminum electrolysis plant to obtain a prediction model that meets the performance requirements.
[0056] In the above modeling process, a special consideration is needed: the data acquisition of the aluminum electrolysis industrial process is affected by various factors, and the acquisition time scale and data structure differ, resulting in multi-source heterogeneous data. Therefore, problems exist such as outliers, default values, random noise, anomalies, missing values, incompleteness, and inconsistencies. Intelligent data preprocessing is required in the modeling process, including normalization, interpolation, and noise reduction to remove outliers, default values, and random noise.
[0057] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.
[0058] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.
Claims
1. A method for constructing a parameter identification and prediction model for aluminum electrolysis based on CRNN, characterized in that: Includes the following steps: S1: Through physicochemical mechanism analysis, the relationship between the unmeasurable variables electrolyte temperature, alumina concentration, and electrode distance and the measurable variables is obtained; A multi-field coupling model of an electrolyzer was constructed using software, and the model was calibrated using measured data. Then, a multi-factor analysis method was used to determine the dependent variables affecting these parameters. S2: Using the dependent variables of temperature, concentration and electrode distance obtained in S1 that affect the aluminum electrolysis production process, a recurrent neural network model is constructed, and a massive amount of historical data recorded in the aluminum electrolysis production process is used as training samples to finally construct a key process parameter identification and prediction model. The recurrent neural network model consists of an input layer, a one-dimensional convolutional layer, a maximum pooling layer, a long short-term memory (LSTM) layer, a fully connected layer, and an output layer. The input layer: Auxiliary variable data of important process parameters such as temperature, concentration and electrode distance in the aluminum electrolysis production process are entered into the network through the input layer; The one-dimensional convolutional layer: performs deep feature extraction on the data through convolutional kernels; The maximum confluence layer: performs subsequence extraction and maximum value calculation on the output data of the convolutional layer; The Long Short-Term Memory (LSTM) layer processes the deep feature sequences after processing. The fully connected layer: outputs prediction results of important process parameters such as temperature, concentration and electrode distance in the aluminum electrolysis production process through the LSTM layer; The output layer outputs predicted values of temperature, concentration, and electrode spacing during the aluminum electrolysis production process. S3: Train the recurrent neural network model using actual operating data from the aluminum electrolysis plant to obtain a predictive model that meets performance requirements; The alumina concentration is determined by first establishing a mapping model between the instantaneous cell voltage, alumina concentration, and feeding interval NB value, and then introducing the influence factors of aluminum removal and anode replacement processes on the alumina concentration. The model is then calibrated using measurable data, and soft measurement of the alumina concentration is then implemented. The alumina concentration is addressed by employing multi-factor coupling to resolve the measurement error caused by the nonlinear relationship between alumina concentration and cell resistance in existing studies. The electrode spacing is considered in terms of the measurable tank voltage and anode resistance, taking into account the relationship between the other electrode spacings; the tank voltage, equidistant voltage drop of the guide rod, guide rod temperature and anode resistance, which can be directly measured online, are used as auxiliary variables.
2. The method for constructing a CRNN-based aluminum electrolysis parameter identification and prediction model as described in claim 1, characterized in that: The electrolyte temperature, in the aluminum electrolysis cell temperature sensing model, is selected as the auxiliary variables: fluoride salt content, aluminum level, electrolyte level, cell voltage, molecular ratio, and aluminum output; the output is the electrolysis temperature; the determined variables are to prepare for subsequent recurrent neural network data fusion.
3. The method for constructing a CRNN-based aluminum electrolysis parameter identification and prediction model as described in claim 1, characterized in that: The data collected for the prediction model is subject to interference from various factors, and the data collection time scale and data structure are different, which are multi-source heterogeneous data. Therefore, there are outliers, default values, random noise, abnormal values, missing values, incompleteness and inconsistency. In the modeling process, intelligent data preprocessing is required, and normalization, interpolation and noise reduction are carried out to remove outliers, default values and random noise.