A method for predicting global anode effect in large aluminum reduction cells
By using the ultimate gradient lifter and causal expansion convolutional neural network model, the anode effect in the aluminum electrolytic cell is predicted, which solves the problem of low prediction accuracy of global anode effect in the prior art, and achieves higher prediction accuracy and more effective electrolysis process control.
Patent Information
- Application Number
- CN202210608390.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-05-31
- Publication Date
- 2025-05-09
- Estimated Expiration
- 2042-05-31
AI Technical Summary
The prior art is difficult to effectively predict and prevent the global anode effect in aluminum electrolytic cells, resulting in low current efficiency, high energy consumption and increased greenhouse gas emissions.
The extreme gradient lifter is used to sort the importance of the anode effect-related features, and a convolutional neural network model containing causal expansion convolution and long and short-term memory layers is constructed to predict the probability of global anode effect occurrence.
By effectively capturing potential information related to the anode effect, the prediction accuracy of the global anode effect is significantly improved, and the probability of the global anode effect occurring in the aluminum electrolytic cell can be predicted the next day.
Smart Images
Figure CN115101136B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of aluminum electrolysis, and in particular to a method for predicting the global anode effect of a large aluminum electrolysis cell. Background Art
[0002] The anode effect is a phenomenon that occurs during the molten salt electrolysis of halides. It is named because it occurs in the anode part of the electrolytic cell. The anode effect is generally caused by insufficient local alumina concentration, which is determined by the feeding strategy and the flow rate driven by the electromagnetic force of the current.
[0003] In the modern aluminum electrolysis industry, the anode effect is mostly a global anode effect. Because the global anode effect reacts violently, a large number of bubbles are generated on the surface of the aluminum liquid. The occurrence of the global anode effect is generally caused by the spread of the local anode effect. Therefore, to reduce the occurrence of the global anode effect, it is necessary to predict the occurrence of the local anode effect and solve it through distributed material feeding. Before the local anode effect occurs, the corresponding anode guide rod current will fluctuate greatly. If it is not intervened, it will gradually develop into a global anode effect.
[0004] Traditional process monitoring technology based on mechanism models has been difficult to be widely used. It can only be limited to the monitoring of single loops of subsystems where analytical models are easy to establish. It is often powerless for online monitoring of abnormal conditions that may affect normal production, system fault diagnosis and subsequent control compensation strategies. Although a large number of detection instruments have been introduced and the necessary hardware conditions for implementing process monitoring technology are available, for some key process variables that are difficult to measure online or have measurement delays, deviations and data missing problems, real-time and accurate online monitoring is still required through data modeling-based technology. Therefore, the rapid development of data modeling technology will promote its application in modern aluminum electrolysis industry process control.
[0005] The prediction of anode effect and its active extinction are one of the research focuses in the modern aluminum electrolysis industry. Reducing the occurrence of inactive anode effect can greatly improve the efficiency of aluminum electrolysis current, reduce energy consumption, and reduce greenhouse gas emissions. Neural networks and deep learning networks are widely used in the field of industrial fault diagnosis. However, a single neural network structure cannot effectively capture the potential information related to the anode effect, resulting in low prediction accuracy of the global anode effect. Summary of the invention
[0006] The embodiment of the present invention provides a method for predicting the global anode effect of a large aluminum electrolysis cell, which can effectively capture potential information related to the anode effect, thereby improving the prediction accuracy of the global anode effect. The technical solution is as follows, including:
[0007] The extreme gradient boosting machine is used to rank the importance of the relevant features of anode effect prediction, and the features with high importance are selected to construct the data set;
[0008] Construct a global anode effect prediction model consisting of a convolutional neural network based on causal dilated convolution and a long short-term memory layer;
[0009] The constructed data set is used to train a global anode effect prediction model. During the training process, a convolutional neural network is used to extract potential information related to the anode effect from multiple selected features, and a long short-term memory layer is used to learn the obtained potential information to obtain the probability of occurrence of the global anode effect.
[0010] The trained global anode effect prediction model is used to predict the probability of global anode effect occurring in the aluminum electrolytic cell on the next day.
[0011] Furthermore, before using the extreme gradient boosting machine to sort the importance of the relevant features of the anode effect prediction and selecting the features with high importance to construct the data set, the method also includes:
[0012] Obtain sample metadata of multiple aluminum electrolysis cells from the aluminum electrolysis big data platform to construct the original data set for anode effect prediction;
[0013] The missing data in the original dataset of anode effect prediction are supplemented on a daily basis.
[0014] Furthermore, the metadata includes: slot number, average working voltage, aluminum output, slot temperature, molecular ratio, aluminum level, iron content, silicon content, multi-point aluminum level, furnace bottom pressure drop, anode effect average voltage, average voltage 15 minutes before anode effect occurs, set voltage and whether anode effect occurs;
[0015] Among them, on the day when no anode effect occurs, the average voltage of 15 minutes before multiple hours is used to replace the average voltage of 15 minutes before the occurrence of anode effect.
[0016] Furthermore, the extreme gradient boosting machine selects feature importance by finding the best feature segmentation point through a greedy algorithm. The number of times a feature is segmented determines the gain of the feature for the training model. The greater the average gain of the feature, the higher the importance of the feature.
[0017] Furthermore, the global anode effect prediction model includes: a plurality of convolutional neural networks; each convolutional neural network includes: a first causal dilated convolution, a first maximum pooling layer, a second causal dilated convolution, a second maximum pooling layer and a flattening process; wherein the first causal dilated convolution is used to extract local features; and the second causal dilated convolution is used to extract global features;
[0018] The method of extracting potential information related to the anode effect from the selected multiple features using a convolutional neural network includes:
[0019] A separate convolutional neural network is used to process the time series of each input feature. In the convolutional neural network, the input feature time series is processed sequentially through the first causal dilated convolution, the first maximum pooling layer, the second causal dilated convolution, the second maximum pooling layer and flattening processing to obtain a flattened potential information matrix related to the anode effect.
[0020] Furthermore, each causal dilated convolution includes: an input layer, a plurality of hidden layers, and an output layer; wherein the dilated convolution calculation is performed on the hidden layer;
[0021] Each causal dilated convolution takes the current input value and the input value at the previous moment as the hidden layer input, and its convolution calculation is only related to the current input and historical input.
[0022] Further, the long short-term memory layer includes: a first long short-term memory neural network, a first neural network unit discard layer, a second long short-term memory neural network, a second neural network unit discard layer and a fully connected layer;
[0023] Among them, the first long short-term memory neural network, the first neural network unit discard layer, the second long short-term memory neural network, the second neural network unit discard layer and the fully connected layer are connected in sequence.
[0024] The beneficial effects brought about by the technical solution provided by the embodiment of the present invention include at least:
[0025] In an embodiment of the present invention, an extreme gradient boosting machine is used to sort the importance of relevant features of anode effect prediction, and features with high importance are selected to construct a data set; a global anode effect prediction model including a convolutional neural network based on causal dilated convolution and a long short-term memory layer is constructed; the global anode effect prediction model is trained using the constructed data set, wherein, during the training process, the convolutional neural network is used to extract potential information related to the anode effect from the selected multiple features, and the long short-term memory layer is used to learn the obtained potential information to obtain the probability of occurrence of the global anode effect; the trained global anode effect prediction model is used to predict the probability of occurrence of the global anode effect in the aluminum electrolytic cell on the next day; in this way, the potential information related to the anode effect can be effectively captured, thereby improving the prediction accuracy of the global anode effect. BRIEF DESCRIPTION OF THE DRAWINGS
[0026] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.
[0027] Figure 1 A schematic diagram of a flow chart of a method for predicting the global anode effect of a large aluminum electrolysis cell provided by an embodiment of the present invention;
[0028] Figure 2 A detailed schematic diagram of a method for predicting the global anode effect of a large aluminum electrolysis cell provided by an embodiment of the present invention;
[0029] Figure 3 A schematic diagram of causal dilated convolution provided for an embodiment of the present invention. DETAILED DESCRIPTION
[0030] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.
[0031] like Figure 1 and Figure 2 As shown, an embodiment of the present invention provides a method for predicting the global anode effect of a large aluminum electrolysis cell, comprising:
[0032] S101, using an extreme gradient boosting machine (XGBoost) to sort the importance of the features related to the anode effect prediction, and select the features with high importance to construct a data set (specifically: a full anode effect prediction data set); wherein the data set includes: a training set and a test set;
[0033] In view of the complex and hysteresis-prone data related to the anode effect in the modern aluminum electrolysis industry, the features selected by expert experience fail to discover the potential information of the anode effect. At the same time, there will be redundant data that affects the prediction effect. Reducing redundant data and extracting effective data will be beneficial to the prediction of the anode effect. In order to solve the problem of anode effect sample selection and reduce the impact of feature drift that often occurs in the prediction of aluminum electrolysis cell conditions based on data, in this embodiment, the XGBoost algorithm is used to sort the feature vectors for the importance of the occurrence of the anode effect, and the features with lower importance are eliminated, and the features with higher importance are used as the initial input of the global anode effect prediction model. The principle of the XGBoost algorithm for feature importance selection is: to achieve it by finding the best feature segmentation point through a greedy algorithm, the number of times the feature is segmented determines the gain of the feature for the training model, and the average gain of the feature affects its importance ranking. The larger the average gain, the higher the importance.
[0034] In this embodiment, the XGBoost algorithm improves the traditional GBDT objective function by adding a regular term to the original objective function, which reduces the possibility of overfitting and speeds up the convergence speed. The objective loss function defined by the XGBoost algorithm is:
[0035]
[0036]
[0037] Among them, Obj (i) is the target loss function value, n is the number of trees, is the true value y i With the predicted value The squared error loss function; Ω(f) is the regularization parameter; γ is the difficulty coefficient of the segmentation tree, which is used to control the generation of the segmentation tree; T represents the number of leaf nodes; λ represents the L2 regularization parameter. The addition of regularization reduces the overfitting phenomenon and speeds up the convergence speed; w represents the weight of the leaf node.
[0038] In this embodiment, the gain Gain is expressed as:
[0039]
[0040] Among them, g L , g R They represent the sum of the first-order partial derivatives of the sample sets contained in the left leaf node and the right leaf node during the split, respectively. L 、h R They represent the sum of the second-order partial derivatives of the sample sets contained in the left leaf node and the right leaf node at the time of splitting;
[0041] In this embodiment, after multiple simulation verifications, with gain Gain as the criterion, features with feature importance values above the threshold d=30 are selected as metadata features for global anode effect prediction model training. For example, features with feature importance scores above 30 (d=30) can be selected to participate in model training.
[0042] In this embodiment, before using the extreme gradient boosting machine to sort the importance of the relevant features of the anode effect prediction and selecting the features with high importance to construct the data set (S101), the method further includes:
[0043] A1, obtain sample metadata of multiple aluminum electrolysis cells from the aluminum electrolysis big data platform and construct the original data set for anode effect prediction;
[0044] In this embodiment, the metadata (referring to features) include: slot number, average working voltage, aluminum output, slot temperature, molecular ratio, aluminum level, iron content, silicon content, multi-point aluminum level, furnace bottom pressure drop, anode effect average voltage, average voltage 15 minutes before anode effect occurs (collected every 30 seconds), set voltage and whether anode effect occurs;
[0045] Among them, on the day when no anode effect occurs, the average voltage of the 15 minutes before multiple hours (for example, 8:00, 12:00, 16:00) is used to replace the average voltage of the 15 minutes before the anode effect occurs.
[0046] A2, using days as the unit, completes the missing data in the original dataset for anode effect prediction.
[0047] In this embodiment, the original data set in step A1 is preprocessed, and different missing data are classified based on days, and different filling methods are selected according to the missing data ratios of different feature values, where:
[0048]
[0049] For features with missing values greater than 60% (e.g., furnace bottom pressure drop, molecular ratio), the mean value of adjacent data of missing data is selected for filling;
[0050] For features with less than n% missing (for example, multi-point aluminum levels), the KNN (K-NearestNeighbor, K nearest neighbor classification algorithm) algorithm is used to fill in the missing values, and the RMSE (Root Mean Square Error) value is used to verify the filled values. The number of K nearest neighbors whose RMSE value between the predicted value and the actual value is less than the threshold value p=4.9 is selected to ensure the validity of the filled data.
[0051] S102, constructing a global anode effect prediction model including a convolutional neural network (CNN) based on causal dilated convolution and a long short-term memory layer; wherein,
[0052] The convolutional neural network includes: a one-dimensional first causal dilated convolution, a first maximum pooling layer, a second causal dilated convolution, a second maximum pooling layer and a flattening process; wherein the first causal dilated convolution is used to extract local features; and the second causal dilated convolution is used to extract global features;
[0053] The long short-term memory layer includes: a first long short-term memory neural network (LSTM), a first neural network unit discard layer, a second long short-term memory neural network, a second neural network unit discard layer and a fully connected layer; wherein, the first long short-term memory neural network, the first neural network unit discard layer, the second long short-term memory neural network, the second neural network unit discard layer and the fully connected layer are connected in sequence.
[0054] The calculation of traditional convolution in convolutional neural network is not only related to the current input and previous historical data, but also takes the state of future data as input and performs convolution calculation, which will cause some future data to leak in the hidden layer. For the prediction of anode effect, the convolutional neural network constructed using traditional convolution will use future production data as convolution input during training, which will be used as the basis for training and affect the accuracy of anode effect prediction. In order to solve the problem of leakage of future information, the traditional convolution is improved and causal dilated convolution is introduced.
[0055] like Figure 3 As shown in Figure 2, each causal dilated convolution consists of an input layer, multiple hidden layers, and an output layer. Figure 3 It can be seen that the causal dilated convolution uses the current input value and the input value at the previous moment as the hidden layer input, and its convolution calculation is only related to the current input and the historical input, which ensures the causal relationship of the time series and solves the problem of future information leakage. However, the memory of causal convolution requires multiple hidden layers as the initial setting. The more layers there are, the larger the memory capacity and the longer the time span of the memory. This will cause the parameters of the neural network model to grow exponentially, making the network structure complex and inconvenient to calculate. To solve this problem, the hidden layer is expanded and convoluted. In the case of no loss of the original input data, the number of parameters for each convolution calculation is reduced. Causal dilated convolution is a form of causal convolution combined with dilated convolution.
[0056] S103, using the constructed data set (specifically the training set) to train a global anode effect prediction model, wherein during the training process, a convolutional neural network is used to extract potential information related to the anode effect from the selected multiple features, and a long short-term memory layer is used to learn the obtained potential information to obtain a global anode effect occurrence probability;
[0057] In this embodiment, after the original data set is sorted, its potential feature information is still implicit in the relevant feature variables. Therefore, the constructed data set is used as the input of the convolutional neural network, and the input feature time series is processed in sequence by the first causal dilated convolution, the first maximum pooling layer, the second causal dilated convolution, the second maximum pooling layer and the flattening process to obtain a flattened potential information matrix related to the anode effect. In this way, the potential information related to the anode effect is extracted using the improved two causal dilated convolutions, which can reduce the leakage of future information in the hidden layer while improving the prediction credibility. In this embodiment, adding a maximum pooling layer after the causal dilated convolution can reduce the size of the feature output and reduce the amount of calculation, thereby greatly reducing the training time of the convolutional neural network.
[0058] In this embodiment, the global anode effect prediction model includes: multiple convolutional neural networks, and each input feature time series is processed using a separate convolutional neural network. In the convolutional neural network, the input feature time series is processed in sequence through the first causal dilated convolution, the first maximum pooling layer, the second causal dilated convolution, the second maximum pooling layer and flattening processing to obtain a flattened potential information matrix related to the anode effect.
[0059] In this embodiment, the flattened potential information matrix output by the convolutional neural network is used as the input of the long short-term memory layer. After two long short-term memory neural network LSTM and neural network unit dropout layer learning, the obtained single-dimensional data is input into the fully connected layer, and finally the probability of occurrence of the global anode effect is output.
[0060] In this embodiment, the global anode effect prediction model obtained after multiple rounds of training is saved. The global anode effect prediction model is a hybrid network model used to predict the probability of occurrence of the anode effect. It is tested on a test set, and the prediction accuracy can reach more than 96.65%. After testing, the prediction effect is better than that of a single LSTM network, which can meet the needs of early prediction of the global anode effect, has good applicability and robustness, is conducive to stabilizing the normal operation of the electrolytic cell, and has fast training speed and strong transferability.
[0061] S104, using the trained global anode effect prediction model to predict the probability of global anode effect occurring in the aluminum electrolytic cell on the next day.
[0062] When the global anode effect prediction method for large-scale aluminum electrolytic cells provided in this embodiment is applied to actual production, the daily production data of multiple days (for example, 500 days) can be processed through steps A1, A2, and S101, and then input into the global anode effect prediction model for training to obtain a trained global anode effect prediction model. The voltage values in daily production are extracted online, and the average voltage value 15 minutes before every half hour is selected as the short-term data update feature and input into the trained global anode effect prediction model to replace the average voltage 15 minutes before the occurrence of the anode effect in the original data set for training once, and other feature values are updated with the daily values of the new day. If the actual production time of the aluminum electrolysis plant is: 8:00-23:00, then the average voltage value of 15 minutes before each time point such as 8:30, 9:00, 9:30, etc. can be selected for training to obtain the probability of global anode effect under data at multiple different time points. The probability of global anode effect under the data at the multiple different time points is added together to obtain the average value, and the probability of global anode effect occurring in the corresponding aluminum electrolysis cell on the next day is obtained, so as to provide assistance for subsequent targeted solutions to the occurrence of the anode effect.
[0063] In summary, in an embodiment of the present invention, an extreme gradient boosting machine is used to rank the importance of relevant features for anode effect prediction, and features with high importance are selected to construct a data set; a global anode effect prediction model including a convolutional neural network based on causal dilated convolution and a long short-term memory layer is constructed; the global anode effect prediction model is trained using the constructed data set, wherein, during the training process, a convolutional neural network is used to extract potential information related to the anode effect from multiple selected features, and the long short-term memory layer is used to learn the obtained potential information to obtain the probability of occurrence of the global anode effect; the trained global anode effect prediction model is used to predict the probability of occurrence of the global anode effect in the aluminum electrolytic cell on the next day; in this way, the potential information related to the anode effect can be effectively captured, thereby improving the prediction accuracy of the global anode effect.
[0064] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims
1. A method for predicting the global anode effect of a large aluminum electrolysis cell, characterized in that: include: The extreme gradient boosting machine is used to rank the importance of the relevant features of anode effect prediction, and the features with high importance are selected to construct the data set; Construct a global anode effect prediction model consisting of a convolutional neural network based on causal dilated convolution and a long short-term memory layer; The constructed data set is used to train a global anode effect prediction model. During the training process, a convolutional neural network is used to extract potential information related to the anode effect from multiple selected features, and a long short-term memory layer is used to learn the obtained potential information to obtain the probability of occurrence of the global anode effect. The trained global anode effect prediction model is used to predict the probability of global anode effect occurring in the aluminum electrolytic cell on the next day; Among them, the extreme gradient boosting machine selects feature importance by finding the best feature segmentation point through a greedy algorithm. The number of times a feature is segmented determines the gain of the feature for the training model. The greater the average gain of the feature, the higher the importance of the feature. The global anode effect prediction model includes: a plurality of convolutional neural networks; each convolutional neural network includes: a first causal dilated convolution, a first maximum pooling layer, a second causal dilated convolution, a second maximum pooling layer and a flattening process; wherein the first causal dilated convolution is used to extract local features; the second causal dilated convolution is used to extract global features; The method of extracting potential information related to the anode effect from the selected multiple features using a convolutional neural network includes: A separate convolutional neural network is used to process the time series of each input feature. In the convolutional neural network, the input feature time series is processed in sequence through the first causal dilated convolution, the first maximum pooling layer, the second causal dilated convolution, the second maximum pooling layer and the flattening process to obtain a flattened potential information matrix related to the anode effect. Each causal dilated convolution includes: an input layer, a plurality of hidden layers and an output layer; wherein the dilated convolution calculation is performed on the hidden layer; Each causal dilation convolution takes the current input value and the input value at the previous moment as the hidden layer input, and its convolution calculation is only related to the current input and historical input; Wherein, the long short-term memory layer includes: a first long short-term memory neural network, a first neural network unit discard layer, a second long short-term memory neural network, a second neural network unit discard layer and a fully connected layer; Wherein, the first long short-term memory neural network, the first neural network unit discard layer, the second long short-term memory neural network, the second neural network unit discard layer and the fully connected layer are connected in sequence; The method of using the long short-term memory layer to learn the obtained potential information to obtain the probability of occurrence of the global anode effect includes: The flattened potential information matrix output by the convolutional neural network is used as the input of the long short-term memory layer. After two long short-term memory neural network and neural network unit discard layer learning, the obtained single-dimensional data is input into the fully connected layer to output the probability of the global anode effect.
2. The method for predicting the global anode effect of a large aluminum electrolysis cell according to claim 1, characterized in that: Before using the extreme gradient boosting machine to sort the importance of the features related to the anode effect prediction and selecting the features with high importance to construct the data set, the method further includes: Obtain sample metadata of multiple aluminum electrolysis cells from the aluminum electrolysis big data platform to construct the original data set for anode effect prediction; The missing data in the original data set for anode effect prediction are supplemented on a daily basis.
3. The method for predicting the global anode effect of a large aluminum electrolysis cell according to claim 2, characterized in that: The metadata includes: slot number, average working voltage, aluminum output, slot temperature, molecular ratio, aluminum level, iron content, silicon content, multi-point aluminum level, furnace bottom pressure drop, anode effect average voltage, average voltage 15 minutes before anode effect occurs, set voltage and whether anode effect occurs; Among them, on the day when no anode effect occurs, the average voltage of 15 minutes before multiple hours is used to replace the average voltage of 15 minutes before the occurrence of anode effect.