Aluminum electrolysis cell alumina concentration distribution efficient prediction method and device based on process structure knowledge neural network
By constructing a neural network based on process structure knowledge, combined with pre-trained modules for flow field and concentration field and process control equations, the problem of rapid and accurate prediction of alumina concentration distribution in aluminum electrolysis cells was solved, achieving efficient prediction of alumina concentration distribution.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CENT SOUTH UNIV
- Filing Date
- 2024-08-26
- Publication Date
- 2026-07-21
AI Technical Summary
Existing technologies struggle to quickly and accurately predict the alumina concentration distribution within aluminum electrolytic cells, and the existing PINN model exhibits poor convergence during training, failing to effectively adapt to disturbances during electrolytic cell operation.
A neural network based on process structure knowledge is used to fit the relationship between spatial location and flow rate and alumina concentration in aluminum electrolysis cell through flow field pre-training module, concentration field pre-training module and process control equation fusion training module. Combined with fixed time intervals, an alumina concentration distribution prediction model is constructed.
It significantly improves the prediction efficiency and accuracy of alumina concentration distribution, reduces the amount of training data required, expands the applicability of the method, and can quickly adapt to disturbances in the operation of electrolytic cells.
Smart Images
Figure CN119149988B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of concentration prediction in industrial process control, and specifically relates to an efficient method and device for predicting the alumina concentration distribution in aluminum electrolysis cells based on a process structure knowledge neural network. Background Technology
[0002] A favorable alumina concentration distribution is a prerequisite for stable operation and high current efficiency in aluminum electrolysis cells. Uneven alumina concentration distribution can lead to problems such as cathodic corrosion and sedimentation at the bottom of the cell, causing voltage fluctuations and even anodic effects, resulting in increased electrolysis energy consumption. However, due to the strong coupling of multiple physical fields within the aluminum electrolysis cell, and the complex behavior of alumina in the electrolyte involving mass transfer, chemical reactions, and other factors, predicting alumina concentration distribution presents a significant challenge. Furthermore, the high temperature and strong corrosiveness of the electrolyte make related industrial experiments very costly and difficult, often limiting alumina concentration measurements to a few locations, making it difficult to support the construction of accurate alumina concentration field prediction models.
[0003] To accurately predict alumina concentration distribution, current research often employs computational fluid dynamics (CFD) models. While CFD methods yield comprehensive and accurate results, they are prohibitively expensive and inefficient; a complete alumina concentration distribution simulation can take several days or even a week, failing to meet the practical need for rapid alumina concentration distribution prediction. Some researchers have also used machine learning methods for concentration prediction, but current work often focuses only on predicting the overall alumina concentration, failing to effectively predict the concentration distribution.
[0004] Physical Information Neural Networks (PINNs) integrate data and physical laws to train deep neural networks. They are an effective way to obtain field information using internal governing equations and limited data, and have achieved excellent results in many fields. However, most existing work often constructs large-scale, complex deep learning networks with multiple hidden layers to find all possible relationships between inputs and outputs. This results in numerous and chaotic network parameters and poor convergence during training, especially in the context of PINN networks where multiple losses compete. Therefore, some researchers have attempted to use partially connected neural networks in systems with some known phenomena to handle situations where certain inputs only affect certain intermediate states or outputs. Furthermore, measurable perturbations exist during electrolytic cell operation, such as feeding, which have additional effects on alumina concentration distribution. However, most existing PINN models are limited to autonomous partial differential cases in the bounded spatiotemporal domain, often requiring re-collection of data for effective modeling in such cases. Summary of the Invention
[0005] This invention provides a highly efficient method and apparatus for predicting alumina concentration distribution in aluminum electrolytic cells based on a process structure knowledge neural network, which provides highly efficient and accurate predictions.
[0006] To achieve the above technical objectives, the present invention adopts the following technical solution:
[0007] An efficient prediction method for alumina concentration distribution in aluminum electrolytic cells based on process structure knowledge neural networks includes:
[0008] The first neural network is used to fit the relationship between spatial position and flow velocity in the aluminum electrolysis cell to obtain the flow field pre-training module;
[0009] The relationship between spatial location and alumina concentration within the aluminum electrolytic cell is fitted using a second neural network to obtain a concentration field pre-training module.
[0010] By combining the third neural network with the flow field pre-training module and the concentration field pre-training module, and by introducing an additional fixed time interval, the process control equation of alumina distribution is incorporated into the loss function to fit the relationship between the spatial location in the aluminum electrolysis cell, the initial alumina concentration and the alumina concentration after the fixed time interval, thus obtaining the alumina concentration distribution prediction model.
[0011] The alumina concentration distribution prediction model is input into the location to be predicted, its initial alumina concentration, and a fixed time interval, and the output is the alumina concentration at the location to be predicted after a fixed time interval.
[0012] Furthermore, the first neural network employs a deep fully connected neural network, and the pre-training module for fitting the flow field is represented as follows:
[0013]
[0014] In the formula, It is a deep fully connected neural network. These are the two-dimensional spatial coordinates within the aluminum electrolysis cell. The weights for the flow field pre-training module; Indicates flow rate, including Flow velocity in the direction and Flow velocity in the direction ,Right now The loss function used in the flow field fitting pre-training module is designed as follows:
[0015]
[0016] In the formula, This represents the loss function used in the pre-training module for fitting the flow field. To determine the number of samples for the pre-training module for fitting the flow field, and These are the predicted output flow velocity of the i-th training sample and the actual flow velocity of the i-th training sample, respectively, provided by the flow field pre-training module.
[0017] Furthermore, the second neural network uses an autoencoder network to reconstruct the spatial location and alumina concentration, mining the relationship between the spatial location and the alumina concentration at the initial time. The pre-training module for fitting the concentration field is represented as follows:
[0018] ,
[0019] In the formula, These are the two-dimensional spatial coordinates within the aluminum electrolysis cell. At the initial moment, the spatial position alumina concentration at the location; For encoder networks, For the weights of the encoder network, Alumina concentration features extracted from the encoder network; For decoder networks, For the weights of the decoder network, Spatial location coordinates and alumina concentration reconstructed by the decoder network;
[0020] The loss function of the autoencoder network is:
[0021]
[0022] In the formula, This represents the loss function used in the pre-training module for fitting the concentration field. To fit the sample size of the concentration field pre-training module, The concentration field pre-training module reconstructs the spatial coordinates and alumina concentration of the j-th training sample. Let be the true spatial coordinates and alumina concentration of the j-th training sample.
[0023] Furthermore, the third neural network employs a deep fully connected neural network, which is combined with the flow field pre-training module and the concentration field pre-training module to fit the alumina concentration distribution prediction model, as follows:
[0024]
[0025] In the formula, This represents a model for predicting alumina concentration distribution. These are the two-dimensional spatial coordinates within the aluminum electrolysis cell. At the initial moment, the spatial position alumina concentration at the location, For fixed time intervals; These are the weights of the third neural network; Predicting the spatial location of the output from the alumina concentration distribution prediction model Alumina concentration after a fixed time interval t;
[0026] The third neural network is combined with the flow field pre-training module and the concentration field pre-training module, specifically as follows:
[0027] First, input the two-dimensional spatial coordinates from the network. The data is fed into the flow field pre-training module, and the spatial position is output. Flow velocity at Additionally, the two-dimensional spatial coordinates input from the network... and the alumina concentration at the initial time of that location The data is fed into the concentration field pre-training module, where the encoder network extracts the spatial location. alumina concentration characteristics at the location ;
[0028] Then, the spatial positions obtained from the flow field pre-training module and the encoder network are respectively... Flow velocity at and alumina concentration characteristics The data, along with a fixed time interval t from the network input, is fed into a third neural network, and the final output yields the spatial location of the aluminum electrolysis cell. alumina concentration after a fixed time interval t .
[0029] Furthermore, the process control equation for alumina distribution is described as follows:
[0030]
[0031] In the formula, and The spatial position of alumina in the aluminum electrolytic cell Along each , Flow velocity in the direction; Spatial location The concentration of alumina; and These represent the volume fraction and density of the electrolyte in the aluminum electrolytic cell, respectively. This represents quality source items, including alumina electrolysis consumption and alumina feed information.
[0032] Furthermore, by incorporating the process control equations governing alumina distribution into the loss function, the overall loss function is obtained. Represented as:
[0033]
[0034]
[0035]
[0036] In the formula, and These represent the PDE loss term and the data loss term, respectively. The weights for the data loss items; and These represent the number of PDE loss term combination points and concentration data points, respectively. For the model to predict the first The concentration of alumina at a spatial location after a fixed time interval t. and The model predicts the first The flow velocity along the x and y directions at each spatial location; and The model prediction and the actual measurement are respectively the first. The concentration of alumina at a spatial location after a fixed time interval t. Quantile level; This indicates that the condition is met. hour, This indicates that the condition is met. hour.
[0037] A highly efficient prediction device for alumina concentration distribution in aluminum electrolytic cells based on a process structure knowledge neural network includes a flow field pre-training module, a concentration field pre-training module, and a process control equation fusion training module.
[0038] The flow field pre-training module uses a first neural network to fit the relationship between spatial position and flow velocity within the aluminum electrolysis cell.
[0039] The concentration field pre-training module uses a second neural network to fit the relationship between the spatial location and the alumina concentration within the aluminum electrolytic cell.
[0040] The process control equation fusion training module uses a third neural network combined with the flow field pre-training module and the concentration field pre-training module, and introduces an additional fixed time interval. By incorporating the process control equation of alumina distribution into the loss function, it fits the relationship between the spatial location in the aluminum electrolysis cell, the initial alumina concentration and the alumina concentration after the fixed time interval, and obtains an alumina concentration distribution prediction model.
[0041] The alumina concentration distribution prediction model takes the location to be predicted, its initial alumina concentration, and a fixed time interval as input, and outputs the alumina concentration at the location to be predicted after a fixed time interval.
[0042] Beneficial effects
[0043] The present invention provides an efficient prediction method for alumina concentration distribution in aluminum electrolytic cells based on a process structure knowledge neural network, which can significantly improve the prediction efficiency and accuracy. Furthermore, it can greatly reduce the amount of data required for network training and expand the applicability of the method. Attached Figure Description
[0044] Figure 1 This is a general framework diagram of the method described in the embodiments of this application.
[0045] Figure 2 The diagram shows the alumina concentration distribution of the embodiments of this application and other comparative methods in an experiment to verify the interpolation capability; wherein, (a) is the initial concentration distribution, (b) is the actual concentration distribution after a fixed time interval t, (c) is the concentration distribution predicted by the method of the present invention after a fixed time interval t, (d) is the concentration distribution predicted by a deep neural network after a fixed time interval t, and (d) is the concentration distribution predicted by a physical information neural network after a fixed time interval t.
[0046] Figure 3 The above diagrams show the alumina concentration distribution of the embodiments of this application compared with other comparative methods in an experiment to verify extrapolation ability; wherein, (a) is the initial concentration distribution, (b) is the actual concentration distribution after a fixed time interval t, (c) is the concentration distribution predicted by the method of the present invention after a fixed time interval t, (d) is the concentration distribution predicted by a deep neural network after a fixed time interval t, and (d) is the concentration distribution predicted by a physical information neural network after a fixed time interval t.
[0047] Figure 4 This is a diagram illustrating the effect of the method described in the embodiments of this application in a time cost experiment. Detailed Implementation
[0048] The embodiments of the present invention will be described in detail below. These embodiments are based on the technical solutions of the present invention and provide detailed implementation methods and specific operation processes to further explain the technical solutions of the present invention.
[0049] This invention proposes an efficient prediction method for alumina concentration distribution in aluminum electrolytic cells based on a process structure knowledge neural network, referencing... Figure 1 As shown, it consists of three organically combined parts: flow field pre-training module, concentration field pre-training module, and process control equation fusion training module.
[0050] 1. Flow field pre-training module: Uses the first neural network to fit the relationship between spatial position and flow velocity within the aluminum electrolysis cell.
[0051] Alumina is primarily transported to all locations within the electrolytic cell by electromagnetic force and the propulsion of anode bubbles; therefore, the flow field is a key factor influencing the alumina concentration distribution. Here, we consider the case where the forces acting on alumina are stable, i.e., the flow field does not change over time. Our goal is to establish a pre-trained flow field module capable of predicting the alumina flow velocity at a given location within the electrolytic cell. This pre-trained module strengthens the relationship between the flow velocity—a process state variable—and location in the overall model. In this embodiment, a deep fully connected neural network is used to perform this regression task, as shown below:
[0052]
[0053] In the formula, It is a deep fully connected neural network. These are the two-dimensional spatial coordinates within the aluminum electrolysis cell. The weights for the flow field pre-training module; Indicates flow rate, including Flow velocity in the direction and Flow velocity in the direction ,Right now The loss function used in the flow field fitting pre-training module is designed as follows:
[0054]
[0055] In the formula, This represents the loss function used in the pre-training module for fitting the flow field. To determine the number of samples for the pre-training module for fitting the flow field, and These are the predicted output flow velocity of the i-th training sample and the actual flow velocity of the i-th training sample, respectively, provided by the flow field pre-training module.
[0056] The pre-trained network of the flow field module achieves accurate prediction of flow velocity at various spatial locations by fitting the network prediction results to the corresponding target values as closely as possible.
[0057] 2. Concentration Field Pre-training Module: A second neural network is used to fit the relationship between the spatial position within the aluminum electrolysis cell and the alumina concentration to obtain the concentration field pre-training module. In this embodiment, the first neural network adopts a deep fully connected neural network.
[0058] Due to the complexity of the alumina transport mechanism, to ensure the stability of the training process and accelerate the network convergence rate, the network needs to be pre-trained to analyze the relationships between concentration field variables. In the concentration field pre-training module, an autoencoder structure is used to reconstruct spatial coordinates and alumina concentration, exploring the relationship between the spatial coordinates and the initial alumina concentration. This facilitates improved efficiency and accuracy in subsequent predictions of future alumina concentrations. The network representation of the concentration field pre-training module is as follows:
[0059] ,
[0060] In the formula, These are the two-dimensional spatial coordinates within the aluminum electrolysis cell. At the initial moment, the spatial position alumina concentration at the location; For encoder networks, For the weights of the encoder network, Alumina concentration features extracted from the encoder network; For decoder networks, For the weights of the decoder network, Spatial location coordinates and alumina concentration reconstructed by the decoder network;
[0061] The loss function of the autoencoder network is:
[0062]
[0063] In the formula, This represents the loss function used in the pre-training module for fitting the concentration field. To fit the sample size of the concentration field pre-training module, The concentration field pre-training module reconstructs the spatial coordinates and alumina concentration of the j-th training sample. Let be the true spatial coordinates and alumina concentration of the j-th training sample.
[0064] The concentration field pre-training module network achieves accurate prediction of alumina concentration at certain spatial locations at the initial moment by making the network reconstruction results as consistent as possible with the corresponding target values.
[0065] 3. Integrated Training Module
[0066] The third neural network is used as the fusion training module, which is combined with the flow field pre-training module and the concentration field pre-training module. A fixed time interval is also introduced. By incorporating the process control equation of alumina distribution into the loss function, the relationship between the spatial location in the aluminum electrolysis cell, the initial alumina concentration and the alumina concentration after the fixed time interval is fitted to obtain the alumina concentration distribution prediction model.
[0067] Based on the inputs provided by the pre-training modules for flow field and concentration, the fusion training module introduces an additional fixed time interval t, which is used to adjust the overall network parameters (including those from the pre-training modules). and the fusion training module Fine-tuning was performed to enable the network to quickly and accurately predict the alumina concentration after time t. Therefore, here, the spatial coordinates... Alumina concentration at initial time Together with a fixed time interval t, they form the input of the overall network, used to predict the alumina concentration at that spatial location after time t. The network is represented as follows.
[0068]
[0069] In the formula, This represents a model for predicting alumina concentration distribution. These are the two-dimensional spatial coordinates within the aluminum electrolysis cell. At the initial moment, the spatial position alumina concentration at the location, For fixed time intervals; These are the weights of the third neural network; Predicting the spatial location of the output from the alumina concentration distribution prediction model The concentration of alumina after a fixed time interval t.
[0070] The third neural network is combined with the flow field pre-training module and the concentration field pre-training module, specifically as follows:
[0071] First, input the two-dimensional spatial coordinates from the network. The data is fed into the flow field pre-training module, and the spatial position is output. Flow velocity at Additionally, the two-dimensional spatial coordinates input from the network... and the alumina concentration at the initial time of that location The data is fed into the concentration field pre-training module, where the encoder network extracts the spatial location. alumina concentration characteristics at the location ;
[0072] Then, the spatial positions obtained from the flow field pre-training module and the encoder network are respectively... Flow velocity at and alumina concentration characteristics The data, along with a fixed time interval t from the network input, is fed into a third neural network, and the final output yields the spatial location of the aluminum electrolysis cell. alumina concentration after a fixed time interval t .
[0073] By using the initial concentration and a fixed time interval as one of the network inputs, the network can adapt to prediction tasks with varying initial concentration conditions, thus avoiding the problem that traditional methods need to remodel and predict when encountering changes in initial concentration conditions.
[0074] The variation of alumina concentration can be described by the time-averaged Navier-Stokes equation. For the two-dimensional alumina concentration distribution model, the mathematical expression of its process control equation is described in this embodiment as follows:
[0075]
[0076] In the formula, and The spatial position of alumina in the aluminum electrolytic cell Along each , Flow velocity in the direction; Spatial location The concentration of alumina; and These represent the volume fraction and density of the electrolyte in the aluminum electrolytic cell, respectively. This represents quality source items, including information such as alumina electrolysis consumption and alumina feed rate.
[0077] To ensure that the network prediction results conform to the constraints of the process control equations, the loss function here adds a PDE loss term to the traditional data loss term. Therefore, the overall loss term is expressed as:
[0078]
[0079]
[0080]
[0081] In the formula, and These represent the PDE loss term and the data loss term, respectively. The weights for the data loss terms are used to balance the numerical magnitudes of the PDE loss terms and the data loss terms. and These represent the number of PDE loss term combination points and concentration data points, respectively. For the model to predict the first The concentration of alumina at a spatial location after a fixed time interval t. and The model predicts the first The flow velocity along the x and y directions at each spatial location; and The model prediction and the actual measurement are respectively the first. The concentration of alumina at a spatial location after a fixed time interval t; This represents the quantile level, typically taking values between 0 and 1; This indicates that the condition is met. hour, This indicates that the condition is met. hour.
[0082] The aforementioned PDE loss term is designed based on the analysis process control equations, aiming to minimize the residuals of the process control equations to enhance the alumina transport mechanism during training. The data loss term adopts the quantile loss form because the alumina concentration only shows a brief peak at the moment of feeding, and the concentration value is relatively low for most of the remaining time. Quantile loss has superior performance in handling such tail values.
[0083] The PDE loss term pairing point and concentration data point correspond to the PDE loss term during training. and data loss items Data points. The training requires actual measured alumina concentrations, but since these data often need to be obtained through laboratory analysis, the quantity is very small, and their composition includes spatial coordinates. The alumina concentration at this location at the initial moment. alumina concentration after time interval t .and The actual alumina concentration does not need to be measured. Therefore, the training data is extensive, and its composition only includes spatial location coordinates. and the alumina concentration at that location at the initial moment. Therefore, the number of PDE loss term combination points. Much greater than the number of concentration data points .
[0084] 4. Predicting alumina concentration
[0085] By inputting the location to be predicted, its initial alumina concentration, and a fixed time interval into the alumina concentration distribution prediction model, the alumina concentration at the location to be predicted after a fixed time interval can be output.
[0086] To verify the effectiveness of the method of the present invention, a series of experiments were conducted on an alumina concentration distribution dataset:
[0087] (1) In experiments verifying interpolation ability, such as Figure 2 As shown, the method of the present invention is significantly better than the comparative method in predicting the details of alumina concentration distribution, especially in the circled part.
[0088] (2) In experiments verifying extrapolation ability, such as Figure 3 As shown, the method of the present invention also accurately predicts the alumina concentration distribution, especially in the circled part, which is very close to the actual alumina concentration distribution.
[0089] (3) In the time cost experiment, such as Figure 4 As shown, the method of this invention has the shortest training and testing time, and compared with traditional computational fluid dynamics (CFD) methods, the efficiency improvement of the proposed method can reach up to 10%. 6 .
[0090] The above embodiments are preferred embodiments of this application. Those skilled in the art can make various changes or improvements based on them. Without departing from the overall concept of this application, these changes or improvements should fall within the scope of protection claimed in this application.
Claims
1. A highly efficient method for predicting alumina concentration distribution in aluminum electrolytic cells based on process structure knowledge neural networks, characterized in that, include: The first neural network is used to fit the relationship between spatial position and flow velocity in the aluminum electrolysis cell to obtain the flow field pre-training module; The relationship between spatial location and alumina concentration within the aluminum electrolytic cell is fitted using a second neural network to obtain a concentration field pre-training module. The third neural network is used as the fusion training module. It is combined with the flow field pre-training module and the concentration field pre-training module. A fixed time interval is introduced. By incorporating the process control equation of alumina distribution into the loss function, the relationship between the spatial location in the aluminum electrolysis cell, the initial alumina concentration and the alumina concentration after the fixed time interval is fitted to obtain the alumina concentration distribution prediction model. The alumina concentration distribution prediction model is input into the location to be predicted, its initial alumina concentration, and a fixed time interval, and the output is the alumina concentration at the location to be predicted after a fixed time interval. The third neural network employs a deep fully connected neural network, which is combined with the flow field pre-training module and the concentration field pre-training module to fit the alumina concentration distribution prediction model, as follows: ； In the formula, This represents a model for predicting alumina concentration distribution. These are the two-dimensional spatial coordinates within the aluminum electrolysis cell. At the initial moment, the spatial position alumina concentration at the location, For fixed time intervals; These are the weights of the third neural network; Predicting the spatial location of the output from the alumina concentration distribution prediction model Alumina concentration after a fixed time interval t; The third neural network is combined with the flow field pre-training module and the concentration field pre-training module, specifically as follows: First, input the two-dimensional spatial coordinates from the network. The data is fed into the flow field pre-training module, and the spatial position is output. Flow velocity at Additionally, the two-dimensional spatial coordinates input from the network... and the alumina concentration at the initial time of that location The data is fed into the concentration field pre-training module, where the encoder network extracts the spatial location. alumina concentration characteristics at the location ; Then, the spatial positions obtained from the flow field pre-training module and the encoder network are respectively... Flow velocity at and alumina concentration characteristics The data, along with a fixed time interval t from the network input, is fed into a third neural network, and the final output yields the spatial location of the aluminum electrolysis cell. alumina concentration after a fixed time interval t .
2. The efficient prediction method for alumina concentration distribution in aluminum electrolytic cells according to claim 1, characterized in that, The first neural network uses a deep fully connected neural network, and the pre-training module for fitting the flow field is represented as follows: ； In the formula, It is a deep fully connected neural network. These are the two-dimensional spatial coordinates within the aluminum electrolysis cell. The weights for the flow field pre-training module; Indicates flow rate, including Flow velocity in the direction and Flow velocity in the direction ,Right now The loss function used in the flow field fitting pre-training module is designed as follows: ； In the formula, This represents the loss function used in the pre-training module for fitting the flow field. To determine the number of samples for the pre-training module for fitting the flow field, and These are the predicted output flow velocity of the i-th training sample and the actual flow velocity of the i-th training sample, respectively, provided by the flow field pre-training module.
3. The efficient prediction method for alumina concentration distribution in aluminum electrolytic cells according to claim 1, characterized in that, The second neural network uses an autoencoder network to reconstruct spatial location and alumina concentration, exploring the relationship between spatial location and alumina concentration at the initial time. The pre-training module for fitting the concentration field is represented as follows: ， ； In the formula, These are the two-dimensional spatial coordinates within the aluminum electrolysis cell. At the initial moment, the spatial position Alumina concentration at the location; For encoder networks, For the weights of the encoder network, Alumina concentration features extracted from the encoder network; For decoder networks, For the weights of the decoder network, Spatial location coordinates and alumina concentration reconstructed by the decoder network; The loss function of the autoencoder network is: ； In the formula, This represents the loss function used in the pre-training module for fitting the concentration field. To fit the sample size of the concentration field pre-training module, The concentration field pre-training module reconstructs the spatial coordinates and alumina concentration of the j-th training sample. Let be the true spatial coordinates and alumina concentration of the j-th training sample.
4. The efficient prediction method for alumina concentration distribution in aluminum electrolytic cells according to claim 1, characterized in that, The process control equation for alumina distribution is described as follows: ； In the formula, and The spatial position of alumina in the aluminum electrolytic cell Along each , Flow velocity in the direction; Spatial location The concentration of alumina; and These represent the volume fraction and density of the electrolyte in the aluminum electrolytic cell, respectively. This represents quality source items, including alumina electrolysis consumption and alumina feed information.
5. The efficient prediction method for alumina concentration distribution in aluminum electrolytic cells according to claim 4, characterized in that, By incorporating the process control equations governing alumina distribution into the loss function, the overall loss function is obtained. Represented as: ； ； ； In the formula, and These represent the PDE loss term and the data loss term, respectively. The weights for the data loss items; and These represent the number of PDE loss term combination points and concentration data points, respectively. For the model to predict the first The concentration of alumina at a spatial location after a fixed time interval t. and The model predicts the first The flow velocity along the x and y directions at each spatial location; and The model prediction and the actual measurement are respectively the first. The concentration of alumina at a spatial location after a fixed time interval t. Quantile level; This indicates that the condition is met. hour, This indicates that the condition is met. hour.
6. A highly efficient prediction device for alumina concentration distribution in aluminum electrolytic cells based on process structure knowledge neural networks, characterized in that, It includes a flow field pre-training module, a concentration field pre-training module, and a process control equation fusion training module; The flow field pre-training module uses a first neural network to fit the relationship between spatial position and flow velocity within the aluminum electrolysis cell. The concentration field pre-training module uses a second neural network to fit the relationship between the spatial location and the alumina concentration within the aluminum electrolytic cell. The process control equation fusion training module uses a third neural network combined with the flow field pre-training module and the concentration field pre-training module, and introduces an additional fixed time interval. By incorporating the process control equation of alumina distribution into the loss function, it fits the relationship between the spatial location in the aluminum electrolysis cell, the initial alumina concentration and the alumina concentration after the fixed time interval, and obtains an alumina concentration distribution prediction model. The alumina concentration distribution prediction model takes the location to be predicted, its initial alumina concentration, and a fixed time interval as input, and outputs the alumina concentration at the location to be predicted after a fixed time interval. The third neural network employs a deep fully connected neural network, which is combined with the flow field pre-training module and the concentration field pre-training module to fit the alumina concentration distribution prediction model, as follows: ； In the formula, This represents a model for predicting alumina concentration distribution. These are the two-dimensional spatial coordinates within the aluminum electrolysis cell. At the initial moment, the spatial position alumina concentration at the location, For fixed time intervals; These are the weights of the third neural network; Predicting the spatial location of the output from the alumina concentration distribution prediction model Alumina concentration after a fixed time interval t; The third neural network is combined with the flow field pre-training module and the concentration field pre-training module, specifically as follows: First, input the two-dimensional spatial coordinates from the network. The data is fed into the flow field pre-training module, and the spatial position is output. Flow velocity at Additionally, the two-dimensional spatial coordinates input from the network... and the alumina concentration at the initial time of that location The data is fed into the concentration field pre-training module, where the encoder network extracts the spatial location. alumina concentration characteristics at the location ; Then, the spatial positions obtained from the flow field pre-training module and the encoder network are respectively... Flow velocity at and alumina concentration characteristics The data, along with a fixed time interval t from the network input, is fed into a third neural network, and the final output yields the spatial location of the aluminum electrolysis cell. alumina concentration after a fixed time interval t .