Method and device for efficiently predicting aluminum oxide concentration distribution of aluminum electrolysis cell based on process structure knowledge neural network

Abstract

The invention discloses an aluminum electrolysis cell alumina concentration distribution efficient prediction method and device based on a process structure knowledge neural network, and the method comprises the steps: fitting a relation between a spatial position and a flow velocity in an aluminum electrolysis cell through a flow field pre-training module; a concentration field pre-training module is used for fitting the relation between the spatial position in the aluminum electrolysis cell and the aluminum oxide concentration; a fusion training module is combined with a flow field pre-training module and a concentration field pre-training module, a fixed time interval is additionally introduced, a process control equation of aluminum oxide distribution is fused into a loss function, and the relation between the spatial position in an aluminum electrolysis cell and the initial aluminum oxide concentration and the aluminum oxide concentration after the fixed time interval is fitted; the aluminum oxide concentration distribution prediction model is used for predicting the aluminum oxide concentration at any position in the aluminum electrolysis cell after the fixed time interval. According to the method, the aluminum electrolysis cell alumina concentration distribution prediction efficiency and precision can be remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the field of concentration prediction in industrial process control, and in particular relates to a method and device for efficiently predicting the distribution of alumina concentration in an aluminum electrolysis cell based on a process structure knowledge neural network. Background Art

[0002] Good alumina concentration distribution is a prerequisite for stable operation and high current efficiency of aluminum electrolytic cells. If the alumina concentration is unevenly distributed, problems such as cathode corrosion and precipitation formation at the bottom of the cell may occur, which in turn causes voltage fluctuations and even causes an anode effect, resulting in increased energy consumption for electrolysis. However, due to the strong coupling of multiple physical fields inside the aluminum electrolytic cell, the behavior of alumina in the electrolyte is a complex process involving multiple factors such as mass transfer and chemical reactions, which poses a major challenge to the prediction of alumina concentration distribution. At the same time, due to the constraints of high temperature and strong corrosiveness of the electrolyte, the cost and difficulty of related industrial experiments are very high, and the measurement of alumina concentration is often limited to a few locations, which makes it difficult to support the construction of an accurate prediction model for the alumina concentration field.

[0003] In order to achieve accurate prediction of aluminum oxide concentration distribution, current research often adopts the method of establishing a computational fluid dynamics (CFD) model to calculate and solve. Although the CFD method can obtain relatively comprehensive and accurate results, the solution cost is too high and the efficiency is very low. A complete simulation of aluminum oxide concentration distribution often takes several days or even a week, which cannot meet the needs of rapid acquisition of aluminum oxide concentration distribution in actual use. Some scholars also use machine learning methods to predict concentration, but current work often only focuses on the overall concentration prediction of aluminum oxide and cannot effectively achieve the prediction of concentration distribution.

[0004] Physical information neural network (PINN) can integrate data and physical laws to train deep neural networks. It is an effective way to obtain field information by evolving internal control equations and a small amount of data, and has achieved very good results in many fields. However, most existing works often construct large-scale complex deep learning networks with multiple hidden layers to find all possible relationships between inputs and outputs. This leads to many and messy network parameters, and poor convergence during training, especially in an environment where multiple losses compete with each other, such as PINN networks. Therefore, some researchers have tried to use partially connected neural networks in systems with some known phenomena to deal with situations where some inputs can only affect certain intermediate states or outputs. In addition, there are some measurable disturbances in the operation of the electrolytic cell, such as unloading, which will have additional effects on the alumina concentration distribution. However, most of the existing PINN models in the literature are limited to autonomous partial differential situations in bounded time and space domains, and often require re-collection of data modeling to effectively adapt to this situation. Summary of the invention

[0005] The present invention provides a method and device for efficiently predicting the distribution of alumina concentration in an aluminum electrolysis cell based on a process structure knowledge neural network, and the prediction is efficient and accurate.

[0006] In order to achieve the above technical objectives, the present invention adopts the following technical solutions:

[0007] An efficient prediction method for alumina concentration distribution in aluminum electrolysis cells based on process structure knowledge neural network, comprising:

[0008] Using the first neural network to fit the relationship between the spatial position and flow velocity in the aluminum electrolysis cell, a flow field pre-training module is obtained;

[0009] Using a second neural network to fit the relationship between the spatial position in the aluminum electrolysis cell and the aluminum oxide concentration, a concentration field pre-training module is obtained;

[0010] The third neural network is combined with the flow field pre-training module and the concentration field pre-training module, and a fixed time interval is additionally introduced. By incorporating the process control equation of alumina distribution into the loss function, the relationship between the spatial position in the aluminum electrolytic cell, the initial alumina concentration and the alumina concentration after a fixed time interval is fitted, and the alumina concentration distribution prediction model is obtained.

[0011] The position to be predicted, its initial aluminum oxide concentration and the fixed time interval are input into the aluminum oxide concentration distribution prediction model, and the aluminum oxide concentration of the position to be predicted after the fixed time interval is output.

[0012] Furthermore, the first neural network adopts a deep fully connected neural network, and the fitting flow field pre-training module is expressed as:

[0013] U=F U (x,y;θ1)

[0014] In the formula, F U (·) is a deep fully connected neural network, (x, y) is the two-dimensional spatial position coordinate in the aluminum electrolytic cell, θ1 is the weight of the flow field pre-training module; U represents the flow velocity, including the flow velocity u in the x direction and the flow velocity v in the y direction, that is, U = [u, v]; the loss function used in the fitting flow field pre-training module is designed as:

[0015]

[0016] Where, L U represents the loss function used by the pre-training module for fitting flow field, N U To fit the number of samples of the flow field pre-training module, and U i They are respectively the predicted output flow rate of the flow field pre-training module for the i-th training sample and the actual flow rate of the i-th training sample.

[0017] Furthermore, the second neural network uses an autoencoder network to reconstruct the spatial position and the aluminum oxide concentration, and explores the relationship between the spatial position and the aluminum oxide concentration at the initial moment. The fitting concentration field pre-training module is expressed as:

[0018]

[0019] Where (x, y) is the two-dimensional spatial coordinates of the aluminum electrolysis cell. is the aluminum oxide concentration at the spatial position (x, y) at the initial moment; F E (·) is the encoder network, θ2 is the weight of the encoder network, and Z is the aluminum oxide concentration feature extracted by the encoder network; F D (·) is the decoder network, θ3 is the weight of the decoder network, and X is the spatial position coordinate and aluminum oxide concentration reconstructed by the decoder network;

[0020] The loss function of the autoencoder network is:

[0021]

[0022] Where, L c represents the loss function used by the pre-training module for fitting concentration field, N d is the number of samples of the pre-training module for fitting the concentration field, is the spatial position coordinates and aluminum oxide concentration reconstructed by the concentration field pre-training module for the jth training sample, X i is the real spatial position coordinate and aluminum oxide concentration of the jth training sample.

[0023] Furthermore, the third neural network uses 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 aluminum oxide concentration distribution prediction model, which is expressed as:

[0024]

[0025] In the formula, F C (·) represents the alumina concentration distribution prediction model, (x, y) is the two-dimensional spatial coordinates in the aluminum electrolysis cell, is the aluminum oxide concentration at the spatial position (x, y) at the initial moment, t is a fixed time interval; θ4 is the weight of the third neural network; m t The aluminum oxide concentration distribution prediction model predicts the aluminum oxide concentration at the output spatial position (x, y) 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:

[0027] First, the two-dimensional spatial position coordinates (x, y) of the network input are sent to the flow field pre-training module, and the flow velocity U at the spatial position (x, y) is output; in addition, the two-dimensional spatial position coordinates (x, y) of the network input and the aluminum oxide concentration at the initial moment of the position are The data is sent to the concentration field pre-training module, and the encoder network in the concentration field pre-training module extracts the aluminum oxide concentration feature Z at the spatial position (x, y);

[0028] Then, the flow velocity U and alumina concentration features Z at the spatial position (x, y) obtained by the flow field pre-training module and the encoder network, respectively, are input into the third neural network together with the fixed time interval t input by the network, and the final output is the alumina concentration m at the spatial position (x, y) of the aluminum electrolytic cell after the fixed time interval t. t .

[0029] Furthermore, the process control equation of alumina distribution is described as follows:

[0030]

[0031] Where u and v are the flow velocities of alumina along the x and y directions at the spatial position (x, y) of the aluminum electrolysis cell, respectively; m is the alumina concentration at the spatial position (x, y); r and ρ represent the volume fraction and density of the electrolyte term in the aluminum electrolysis cell, respectively; S represents the mass source term, including the electrolytic consumption of alumina and the feeding information of alumina.

[0032] Furthermore, the process control equation of alumina distribution is incorporated into the loss function, and the overall loss function L is expressed as:

[0033] L=L p +λ d L d

[0034]

[0035] Where, L p and L d Represent the PDE loss term and data loss term respectively, λ d is the weight of the data loss item; N p and N d are the number of PDE loss term collocation points and concentration data points respectively; The model predicts the aluminum oxide concentration at the kth spatial position after a fixed time interval t, and The model predicts the flow velocity along the x and y directions at the k-th spatial position, respectively; and m jare the alumina concentrations predicted by the model and actually measured at the jth spatial position after a fixed time interval t, respectively, and γ is the quantile level; Indicates that the condition is met hour, Indicates that the condition is met hour.

[0036] An efficient prediction device for aluminum oxide concentration distribution in aluminum electrolysis cells based on process structure knowledge neural network, including a flow field pre-training module, a concentration field pre-training module and a process control equation fusion training module;

[0037] The flow field pre-training module uses a first neural network to fit the relationship between the spatial position and the flow velocity in the aluminum electrolysis cell.

[0038] The concentration field pre-training module uses a second neural network to fit the relationship between the spatial position in the aluminum electrolysis cell and the aluminum oxide concentration;

[0039] The process control equation fusion training module uses a third neural network in combination with a flow field pre-training module and a concentration field pre-training module, and additionally introduces a fixed time interval. By integrating the process control equation of alumina distribution into the loss function, the relationship between the spatial position in the aluminum electrolytic cell, the initial alumina concentration and the alumina concentration after a fixed time interval is fitted to obtain an alumina concentration distribution prediction model;

[0040] The aluminum oxide concentration distribution prediction model takes the position to be predicted and its initial aluminum oxide concentration and a fixed time interval as input, and outputs the aluminum oxide concentration of the position to be predicted after a fixed time interval.

[0041] Beneficial Effects

[0042] The invention provides an efficient prediction method for aluminum oxide concentration distribution in aluminum electrolysis cells based on process structure knowledge neural network, which can significantly improve the prediction efficiency and accuracy. In addition, it can greatly reduce the amount of data required for network training and expand the scope of application of the method. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is an overall framework diagram of the method described in the embodiment of the present application.

[0044] Figure 2 It is a graph of aluminum oxide concentration distribution of the embodiment of the present 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.

[0045] Figure 3 It is a graph of aluminum oxide concentration distribution of the embodiment of the present application and other comparative methods in an experiment to verify the 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.

[0046] Figure 4 This is a diagram showing the effect of the method described in the embodiment of the present application in a time cost experiment. DETAILED DESCRIPTION

[0047] The following is a detailed description of an embodiment of the present invention. This embodiment is based on the technical solution of the present invention, and provides a detailed implementation method and a specific operation process to further explain the technical solution of the present invention.

[0048] The present invention proposes an efficient prediction method for the alumina concentration distribution in an aluminum reduction cell based on a process structure knowledge neural network. Figure 1 As shown, it is composed of an organic combination of three parts: flow field pre-training module, concentration field pre-training module and process control equation fusion training module.

[0049] 1. Flow field pre-training module: Use the first neural network to fit the relationship between the spatial position and flow velocity in the aluminum electrolysis cell.

[0050] Alumina is mainly transported to various positions in the entire tank by electromagnetic force and anode bubbles, so the flow field is a key factor affecting the concentration distribution of alumina. Here, we consider the situation where the force acting on alumina is stable, that is, the flow field does not change with time. Our goal is to establish a flow field pre-training module that can predict the flow rate of alumina at that position based on the spatial position information in the electrolytic cell. Through this pre-training module, the relationship between the process state variable of flow rate and position is strengthened in the overall model. In this embodiment, a deep fully connected neural network is used to perform this regression task, which is expressed as follows:

[0051] U=F U (x,y;θ1)

[0052] In the formula, F U (·) is a deep fully connected neural network, (x, y) is the two-dimensional spatial position coordinate in the aluminum electrolytic cell, θ1 is the weight of the flow field pre-training module; U represents the flow velocity, including the flow velocity u in the x direction and the flow velocity v in the y direction, that is, U = [u, v]; the loss function used in the fitting flow field pre-training module is designed as:

[0053]

[0054] Where, L U represents the loss function used by the pre-training module for fitting flow field, N U To fit the number of samples of the flow field pre-training module, and U i They are respectively the predicted output flow rate of the flow field pre-training module for the i-th training sample and the actual flow rate of the i-th training sample.

[0055] The network of the flow field pre-training module achieves accurate prediction of flow velocity at each spatial position by fitting the network prediction results and the corresponding target values ​​as much as possible.

[0056] 2. Concentration field pre-training module: Use the second neural network to fit the relationship between the spatial position in the aluminum electrolysis cell and the aluminum oxide concentration to obtain a concentration field pre-training module. In this embodiment, the first neural network adopts a deep fully connected neural network.

[0057] Due to the complexity of the mechanism in the alumina transport process, in order to ensure the stability of the training process and accelerate the convergence rate of the network, the network needs to be pre-trained on the relationship between concentration field variables. In the concentration field pre-training module, the autoencoder structure is used to reconstruct the spatial coordinates and alumina concentration, and the relationship between the spatial coordinates and the alumina concentration at the initial moment is mined, so as to improve the efficiency and accuracy of the subsequent prediction of alumina concentration at future moments. The network representation of the concentration field pre-training module is as follows:

[0058]

[0059] Where (x, y) is the two-dimensional spatial coordinates of the aluminum electrolysis cell. is the aluminum oxide concentration at the spatial position (x, y) at the initial moment; F E (·) is the encoder network, θ2 is the weight of the encoder network, and Z is the aluminum oxide concentration feature extracted by the encoder network; F D (·) is the decoder network, θ3 is the weight of the decoder network, and X is the spatial position coordinate and aluminum oxide concentration reconstructed by the decoder network;

[0060] The loss function of the autoencoder network is:

[0061]

[0062] Where, L c represents the loss function used by the pre-training module for fitting concentration field, N d is the number of samples of the pre-training module for fitting the concentration field, is the spatial position coordinates and aluminum oxide concentration reconstructed by the concentration field pre-training module for the jth training sample, X jis the real spatial position coordinate and aluminum oxide concentration of the jth training sample.

[0063] The concentration field pre-training module network achieves accurate prediction of the aluminum oxide concentration at some spatial positions at the initial moment by making the network reconstruction results consistent with the corresponding target values ​​as much as possible.

[0064] 3. Fusion training module

[0065] The third neural network is used as a 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 additionally introduced. By incorporating the process control equation of alumina distribution into the loss function, the relationship between the spatial position in the aluminum electrolytic cell, the initial alumina concentration and the alumina concentration after a fixed time interval is fitted to obtain the alumina concentration distribution prediction model.

[0066] Based on the inputs provided by the flow field and concentration pre-training modules, the fusion training module introduces an additional fixed time interval t. By fine-tuning the overall network parameters (including θ1, θ2, θ3 in the pre-training module and θ4 in the fusion training module), the network can quickly and accurately predict the aluminum oxide concentration after time t. and a fixed time interval t together form the input of the overall network to predict the aluminum oxide concentration m at the spatial position after time t t , the network is represented as follows.

[0067]

[0068] In the formula, F C (·) represents the alumina concentration distribution prediction model, (x, y) is the two-dimensional spatial coordinates in the aluminum electrolysis cell, is the aluminum oxide concentration at the spatial position (x, y) at the initial moment, t is a fixed time interval; θ4 is the weight of the third neural network; m t The aluminum oxide concentration distribution prediction model predicts the aluminum oxide concentration at the output spatial position (x, y) after a fixed time interval t.

[0069] The third neural network is combined with the flow field pre-training module and the concentration field pre-training module, specifically:

[0070] First, the two-dimensional spatial position coordinates (x, y) of the network input are sent to the flow field pre-training module, and the flow velocity U at the spatial position (x, y) is output; in addition, the two-dimensional spatial position coordinates (x, y) of the network input and the aluminum oxide concentration at the initial moment of the position are The data is sent to the concentration field pre-training module, and the encoder network in the concentration field pre-training module extracts the aluminum oxide concentration feature Z at the spatial position (x, y);

[0071] Then, the flow velocity U and alumina concentration features Z at the spatial position (x, y) obtained by the flow field pre-training module and the encoder network, respectively, are input into the third neural network together with the fixed time interval t input by the network, and the final output is the alumina concentration m at the spatial position (x, y) of the aluminum electrolytic cell after the fixed time interval t. t .

[0072] Taking the initial concentration and the fixed time interval as one of the network inputs can make the network adapt to the prediction task of variable initial concentration conditions, thus getting rid of the problem that traditional methods need to re-model and predict when encountering changes in initial concentration conditions.

[0073] The variation law of aluminum oxide concentration can be described by the time-averaged Navier-Stokes equation. For the two-dimensional aluminum oxide concentration distribution model, the mathematical expression of the process control equation in this embodiment is described as follows:

[0074]

[0075] Where u and v are the flow velocities of alumina along the x and y directions at the spatial position (x, y) of the aluminum electrolysis cell, respectively; m is the alumina concentration at the spatial position (x, y); r and ρ represent the volume fraction and density of the electrolyte term in the aluminum electrolysis cell, respectively; S represents the mass source term, including information such as the electrolytic consumption of alumina and the feeding of alumina.

[0076] In order to make the network prediction results meet the constraints of the process control equation, the loss function here adds a PDE loss term on the basis of the traditional data loss term, so the overall loss term is expressed as:

[0077] L=L p +λ d L d

[0078]

[0079] Where, L p and L d Represent the PDE loss term and data loss term respectively, λ d is the weight of the data loss term, which is used to balance the numerical magnitude of the PDE loss term and the data loss term; N p and N d are the number of PDE loss term collocation points and concentration data points respectively; The model predicts the aluminum oxide concentration at the kth spatial position after a fixed time interval t, and The model predicts the flow velocity along the x and y directions at the k-th spatial position, respectively; and mj are the alumina concentrations predicted by the model and actually measured at the jth spatial position after a fixed time interval t, respectively; γ is the quantile level, which usually takes a value between 0 and 1; Indicates that the condition is met hour, Indicates that the condition is met hour.

[0080] The above PDE loss term is designed based on the analysis of the process control equation, aiming to minimize the residual of the process control equation to strengthen the transport mechanism of aluminum oxide during the training process. The data loss term adopts the form of quantile loss because the concentration of aluminum oxide only has a short peak at the time of material discharge, and the concentration value is relatively low for most of the rest of the time, and quantile loss has better performance in dealing with such tail values.

[0081] The PDE loss term collocation points and concentration data points correspond to the PDE loss term L during training. p and the data loss term L d Data points. d The actual measured aluminum oxide concentration is needed for training, but since these data often need to be obtained by laboratory tests, the number is small. Its composition includes the spatial position coordinates (x, y), the aluminum oxide concentration at the initial time, and the and the aluminum oxide concentration m after time interval t t . And L p The actual aluminum oxide concentration m does not need to be measured. t Therefore, there are a lot of corresponding training data during training, which only includes the spatial position coordinates (x, y) and the aluminum oxide concentration at the initial moment. Therefore, the number of points where the PDE loss term is matched is N p Much larger than the number of concentration data points N d .

[0082] 4. Predict alumina concentration

[0083] The position to be predicted, its initial aluminum oxide concentration and the fixed time interval are input into the aluminum oxide concentration distribution prediction model, and the aluminum oxide concentration of the position to be predicted after the fixed time interval can be output.

[0084] In order to verify the effectiveness of the method of the present invention, a series of experimental verifications were carried out on an aluminum oxide concentration distribution data set:

[0085] (1) In the experiment to verify the interpolation capability, Figure 2 As shown, it can be seen that the method of the present invention is significantly better than the comparative method in predicting the details of the aluminum oxide concentration distribution, especially in the circled part.

[0086] (2) In experiments to verify the extrapolation capability, e.g. Figure 3 As shown, it can be seen that the method of the present invention also accurately predicts the aluminum oxide concentration distribution, especially in the circled part, which is very close to the actual aluminum oxide concentration distribution.

[0087] (3) In the time cost experiment, Figure 4 As shown in Figure 2, it can be seen that the training and testing time of the proposed method is the shortest. Compared with the traditional computational fluid dynamics (CFD) method, the efficiency of the proposed method can be improved by up to 10 6 .

[0088] The above embodiments are preferred embodiments of the present application. Ordinary technicians in this field can also make various changes or improvements on this basis. Without departing from the overall concept of the present application, these changes or improvements should fall within the scope of protection required by the present application.

Claims

1. An efficient prediction method for alumina concentration distribution in aluminum electrolysis cells based on process structure knowledge neural network, It is characterized in that include: Using the first neural network to fit the relationship between the spatial position and flow velocity in the aluminum electrolysis cell, a flow field pre-training module is obtained; Using a second neural network to fit the relationship between the spatial position in the aluminum electrolysis cell and the aluminum oxide concentration, a concentration field pre-training module is obtained; The third neural network is used as a fusion training module, which is combined with the flow field pre-training module and the concentration field pre-training module, and a fixed time interval is additionally introduced. By incorporating the process control equation of alumina distribution into the loss function, the relationship between the spatial position in the aluminum electrolytic cell, the initial alumina concentration and the alumina concentration after a fixed time interval is fitted, and the alumina concentration distribution prediction model is obtained; The position to be predicted, its initial aluminum oxide concentration and the fixed time interval are input into the aluminum oxide concentration distribution prediction model, and the aluminum oxide concentration of the position to be predicted after the fixed time interval is output.

2. The method for efficiently predicting the alumina concentration distribution in an aluminum electrolysis cell according to claim 1, It is characterized in that The first neural network uses a deep fully connected neural network, and the fitting flow field pre-training module is expressed as: U=F U (x,y;θ 1 ) In the formula, F U (·) is a deep fully connected neural network, (x, y) is the two-dimensional spatial coordinates of the aluminum electrolysis cell, θ 1 is the weight of the flow field pre-training module; U represents the flow velocity, including the flow velocity u in the x direction and the flow velocity v in the y direction, that is, U = [u, v]; the loss function used in the fitting flow field pre-training module is designed as: Where, L U represents the loss function used by the pre-training module for fitting flow field, N U To fit the number of samples of the flow field pre-training module, and U i They are respectively the predicted output flow rate of the flow field pre-training module for the i-th training sample and the actual flow rate of the i-th training sample.

3. The method for efficiently predicting the alumina concentration distribution in an aluminum electrolysis cell according to claim 1, It is characterized in that The second neural network uses an autoencoder network to reconstruct the spatial position and aluminum oxide concentration, and explores the relationship between the spatial position and the aluminum oxide concentration at the initial moment. The fitting concentration field pre-training module is expressed as: Where (x, y) is the two-dimensional spatial coordinates of the aluminum electrolysis cell. is the aluminum oxide concentration at the spatial position (x, y) at the initial moment; F E (·) is the encoder network, θ 2 is the weight of the encoder network, Z is the aluminum oxide concentration feature extracted by the encoder network; F D (·) is the decoder network, θ 3 is the weight of the decoder network, X is the spatial position coordinate and aluminum oxide concentration reconstructed by the decoder network; The loss function of the autoencoder network is: Where, L c represents the loss function used by the pre-training module for fitting concentration field, N d is the number of samples of the pre-training module for fitting the concentration field, is the spatial position coordinates and aluminum oxide concentration reconstructed by the concentration field pre-training module for the jth training sample, X j is the real spatial position coordinate and aluminum oxide concentration of the jth training sample.

4. The method for efficiently predicting the alumina concentration distribution in an aluminum electrolysis cell according to claim 1, It is characterized in that The third neural network uses 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 aluminum oxide concentration distribution prediction model, which is expressed as: In the formula, F C (·) represents the alumina concentration distribution prediction model, (x, y) is the two-dimensional spatial coordinates in the aluminum electrolysis cell, is the aluminum oxide concentration at the spatial position (x, y) at the initial moment, t is a fixed time interval; θ 4 is the weight of the third neural network; m t The aluminum oxide concentration distribution prediction model predicts the aluminum oxide concentration at the output spatial position (x, y) 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: First, the two-dimensional spatial position coordinates (x, y) of the network input are sent to the flow field pre-training module, and the flow velocity U at the spatial position (x, y) is output; in addition, the two-dimensional spatial position coordinates (x, y) of the network input and the aluminum oxide concentration at the initial moment of the position are The data is sent to the concentration field pre-training module, and the encoder network in the concentration field pre-training module extracts the aluminum oxide concentration feature Z at the spatial position (x, y); Then, the flow velocity U and alumina concentration features Z at the spatial position (x, y) obtained by the flow field pre-training module and the encoder network, respectively, are input into the third neural network together with the fixed time interval t input by the network, and the final output is the alumina concentration m at the spatial position (x, y) of the aluminum electrolytic cell after the fixed time interval t. t .

5. The method for efficiently predicting the alumina concentration distribution in an aluminum electrolysis cell according to claim 1, It is characterized in that The process control equation for alumina distribution is described as follows: Where u and v are the flow velocities of alumina along the x and y directions at the spatial position (x, y) of the aluminum electrolysis cell, respectively; m is the alumina concentration at the spatial position (x, y); r and ρ represent the volume fraction and density of the electrolyte term in the aluminum electrolysis cell, respectively; S represents the mass source term, including the electrolytic consumption of alumina and the feeding information of alumina.

6. The method for efficiently predicting the aluminum oxide concentration distribution in an aluminum electrolysis cell according to claim 5, It is characterized in that Incorporating the process control equation of alumina distribution into the loss function, the overall loss function L is expressed as: L=L p +λ d L d Where, L p and L d Represent the PDE loss term and data loss term respectively, λ d is the weight of the data loss item; N p and N d are the number of PDE loss term collocation points and concentration data points respectively; The model predicts the aluminum oxide concentration at the kth spatial position after a fixed time interval t, and The model predicts the flow velocity along the x and y directions at the k-th spatial position, respectively; and m j are the alumina concentrations predicted by the model and actually measured at the jth spatial position after a fixed time interval t, respectively, and γ is the quantile level; Indicates that the condition is met hour, Indicates that the condition is met hour.

7. An efficient prediction device for alumina concentration distribution in aluminum electrolysis cells based on process structure knowledge neural network, It is characterized in that Including flow field pre-training module, concentration field pre-training module and process control equation fusion training module; The flow field pre-training module uses a first neural network to fit the relationship between the spatial position and the flow velocity in the aluminum electrolysis cell. The concentration field pre-training module uses a second neural network to fit the relationship between the spatial position in the aluminum electrolysis cell and the aluminum oxide concentration; The process control equation fusion training module uses a third neural network in combination with a flow field pre-training module and a concentration field pre-training module, and additionally introduces a fixed time interval. By integrating the process control equation of alumina distribution into the loss function, the relationship between the spatial position in the aluminum electrolytic cell, the initial alumina concentration and the alumina concentration after a fixed time interval is fitted to obtain an alumina concentration distribution prediction model; The aluminum oxide concentration distribution prediction model takes the position to be predicted and its initial aluminum oxide concentration and a fixed time interval as input, and outputs the aluminum oxide concentration of the position to be predicted after a fixed time interval.

Images