A migration optimization method for an aluminum electrolysis early return prediction model

Abstract

The application discloses a kind of migration optimization methods of aluminum electrolysis early return prediction model, belong to aluminum electrolysis technical field, including extracting existing aluminum electrolysis cell return prediction model from model library, record the state vector of newly started aluminum electrolysis cell and input into the existing aluminum electrolysis cell return prediction model extracted, record predicted value, according to predicted value and the latest timestamp corresponding response value prediction error value of newly started aluminum electrolysis cell, and obtain the minimum value of predicted error value etc..The application is based on existing return prediction model to construct the return prediction model of newly started aluminum electrolysis cell, avoids the risk caused by using only a certain model, makes full use of the similar information of small amount of data of newly started aluminum electrolysis cell and large amount of data of existing aluminum electrolysis cell;Optimization algorithm is used to realize rapid optimization, and the best weight value combination can be effectively found out, to provide efficient return prediction model construction support for newly started aluminum electrolysis cell.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of aluminum electrolysis, in particular to a migration optimization method of an early aluminum electrolysis regression prediction model. BACKGROUND

[0002] The aluminum electrolysis cell is the core equipment of the aluminum electrolysis industry, and its production process is designed under complex conditions such as high temperature and high magnetism. It is a complex nonlinear process with large time delay, large time variation, and multiple inputs and outputs. The parameters that can be obtained in real time during the production process are limited to current and voltage, and other key parameters are difficult to obtain in real time. This limitation makes it difficult for traditional operation decision-making methods to accurately reflect the state of the aluminum electrolysis cell and respond to the complex changes in the aluminum electrolysis process.

[0003] In addition, with the development of the aluminum electrolysis industry in a large scale and group, the industry concentration is becoming higher and higher, and a large amount of data will be generated every day. Under such a background, many effective data-driven soft measurement technologies have emerged, such as regression prediction models. However, in actual production, the characteristics of different aluminum electrolysis cells are different, which requires that a regression prediction model must be established for each aluminum electrolysis cell independently. However, the modeling has high requirements for data, especially the amount of data. Too little amount of data may result in the inability of the model to represent the complete characteristics of the aluminum electrolysis cell, which leads to the fact that for a newly started aluminum electrolysis cell, an effective regression prediction model cannot be established before the data accumulation is sufficient.

[0004] In addition, a large number of established regression prediction models are one-to-one bound with aluminum electrolysis cells, which means that the life cycle of these regression prediction models will end with the shutdown of the aluminum electrolysis cell, resulting in the waste of many data features. SUMMARY

[0005] The technical problem to be solved by the present application is to solve the above-mentioned deficiencies of the prior art, and to provide a migration optimization method of an early aluminum electrolysis regression prediction model, which can migrate the key features of the existing aluminum electrolysis cell regression prediction model to construct a regression prediction model for a newly started aluminum electrolysis cell, thereby solving the problem of insufficient data accumulation for training a new regression prediction model in the early stage and increasing the possibility of early application of data-driven regression prediction models.

[0006] In order to achieve the above-mentioned purpose, the main technical scheme adopted by the present application is as follows:

[0007] A migration optimization method of an early aluminum electrolysis regression prediction model, comprising the following steps:

[0008] S1, extracting an existing aluminum electrolysis cell regression prediction model from a model library ;

[0009] ;

[0010] in Number the model. This represents the total number of extracted regression prediction models for aluminum electrolytic cells. It is a positive integer greater than 1; For the membership center matrix, For the membership width matrix, Consequence parameter matrix;

[0011] S2. Record the state vector of the newly started aluminum electrolysis cell. , and input into In step S1, the existing aluminum electrolysis cell regression prediction model is extracted, and the record is... A model Predicted value The mathematical model is: ;

[0012] in, This is the state vector corresponding to the latest timestamp in the recorded data for the newly started aluminum electrolysis cell; The first stage of aluminum electrolysis cell production Each state value State value In the state vector Index in The total number of state values ​​in the state vector. S3, according to the first... A model Predicted value Response value corresponding to the latest timestamp of the newly started aluminum electrolytic cell Calculate the first The prediction error value of each model And obtain the minimum value of the prediction error, that is, the minimum error value. The prediction error value is compared with the existing aluminum electrolytic cell regression prediction model extracted in step S1. One-to-one correspondence;

[0013] S4, Preset Threshold traversal The extracted aluminum electrolytic cell regression prediction model ,like Prediction error value Less than the threshold Then mark A promising model The mathematical model is as follows:

[0014] , A promising model In all the potential models, is a positive integer greater than 1;

[0015] S5, according to the information of the parameter matrix of the potential model Initialize the weight matrix ;

[0016] S6, taking the minimization of the prediction error As the optimization goal, use the differential evolution algorithm to optimize the weight matrix Obtain the parameter matrix of the newly started aluminum electrolysis cell regression prediction model, and according to the obtained parameter matrix, the newly started aluminum electrolysis cell regression prediction model can be obtained.

[0017] The model library in step S1 is the model library of the existing aluminum electrolysis cell regression prediction model, and all the regression prediction models in the model library are based on fuzzy neural network.

[0018] In step S1, each model contains three parameter matrices, namely membership center matrix , membership width matrix and consequent parameter matrix , wherein , and are all two-dimensional floating point matrices, and the dimension of each matrix is a fixed positive integer value.

[0019] In step S2, the production state of the aluminum electrolysis cell includes but is not limited to aluminum electrolysis cell temperature, aluminum electrolysis cell current, aluminum liquid level, molecular ratio, electrolyte level, furnace bottom pressure drop, aluminum electrolysis cell average voltage, aluminum electrolysis cell working voltage, effect frequency, and aluminum electrolysis cell age.

[0020] The threshold value in step S4 is a value preset by the user according to experience, The greater the value, the more potential models are screened.

[0021] In step S5 , the parameter matrix information includes:

[0022] The number of all parameter matrix types: that is, three parameter matrices;

[0023] The number of the same type of parameter matrix: that is, the number of potential models ;

[0024] Initialize the weight matrix , representing a real number, the whole is represented as a a two-dimensional matrix with 3 rows and 3 columns, each element of the matrix is a real number;

[0025] weight matrix in the middle in the row the element in the column is denoted as wherein is a row index, is a column index, and takes a positive integer; is a real number, and , that is, the sum of the elements in each column of the weight matrix is 1.

[0026] In step S6, the weight matrix and the parameter matrix of the potential model are summed by weighting to obtain a new parameter matrix , and a new regression prediction model for the aluminum electrolytic cell is obtained ;

[0027] The optimization objective is determined to be the minimization of the prediction error , and the mathematical model is:

[0028] ;

[0029] wherein is the prediction error, and is a positive real number greater than or equal to 0, is the new regression prediction model for the aluminum electrolytic cell obtained after inputting the state vector to obtain a predicted value.

[0030] The present application has the following beneficial effects and advantages:

[0031] 1. The present application is based on the existing regression prediction model to construct a regression prediction model for a newly started aluminum electrolytic cell, avoids the risk brought by using only a certain model, and fully utilizes the similar information of the small amount of data of the newly started aluminum electrolytic cell and the large amount of data of the existing aluminum electrolytic cell; the optimization algorithm is used to realize fast optimization, and the best weight value combination can be effectively found out, thereby providing efficient regression prediction model construction support for the newly started aluminum electrolytic cell.

[0032] 2. The present application solves the challenge that the early data of the newly started aluminum electrolytic cell is insufficient to train the regression prediction model by using transfer learning; the existing regression prediction model can be fully utilized, and the life cycle of the data features is prolonged. DETAILED DESCRIPTION

[0033] The following section further elaborates on the migration optimization method for an early regression prediction model of aluminum electrolysis according to the present invention.

[0034] Step S1: Extract existing aluminum electrolytic cell regression prediction models from the model library, and use symbolic representation. To refer to the extracted first A regression prediction model for aluminum electrolytic cells, in which Number the model. This represents the total number of extracted regression prediction models for aluminum electrolytic cells. It is a positive integer greater than 1. Furthermore, this Each aluminum electrolytic cell regression prediction model corresponds to the source domain in transfer learning.

[0035] Aluminum electrolysis regression prediction model in the described embodiments All are built on fuzzy neural networks, so each model Both contain three parameter matrices: membership center matrix and membership matrix. Membership width matrix and consequent parameter matrix ,in , and All are two-dimensional floating-point matrices, with each matrix having a fixed dimension of positive integers. Since the parameter matrix is ​​the core of the aluminum electrolysis regression prediction model, determining the parameter matrix determines the aluminum electrolysis regression prediction model. Therefore, the first... A regression prediction model for aluminum electrolytic cells Recorded as .

[0036] Step S2: Record the state vector of the newly started aluminum electrolytic cell. Input to From the extracted existing aluminum electrolysis regression prediction models, the following formula can be used to calculate the first... A model Predicted value : ;

[0037] in, This is the state vector corresponding to the latest timestamp in the recorded data for the newly started aluminum electrolysis cell. The first of the production statuses of aluminum electrolysis cells (including aluminum electrolysis cell temperature, aluminum electrolysis cell current, aluminum liquid level, etc.) Each state value State value In the state vector Index in The total number of state values ​​in the state vector. It is a positive integer greater than 1. Furthermore, This is the mapping function for the regression prediction model of aluminum electrolytic cells. This function can be based on the input model. parameter matrix and state vector Calculate the model Predicted value ,in For the first A model The corresponding predicted value is a floating-point number.

[0038] Step S3: Calculate these predicted values Response values ​​of newly started aluminum electrolytic cells (as recorded in the latest data) The absolute value of the difference, i.e. ,in, The response value corresponding to the latest timestamp in the record for the newly started aluminum electrolysis cell is a floating-point number. For the first The prediction error values ​​of each model are taken as floating-point numbers, and the minimum prediction error value among these models is recorded as the minimum error value. , take a floating point number.

[0039] Step S4: Traverse One model. Only one model. Prediction error value Less than a preset threshold ,Model These were then marked as promising models, among which, The threshold is preset by the user based on experience. ,and The higher the value, the more potential models are selected, and vice versa.

[0040] Potential models can be labeled as , A promising model In all An index in a promising model, Take a positive integer greater than 1, following the model. The description of the three parameter matrices, the model The parameter matrix can be represented as , and , recorded as .

[0041] Step S5: Based on the information obtained in step S4 The parameter matrix information of a potential model, that is, the number of types of parameter matrices (here it equals 3, that is, 3 types of parameter matrices: , and ) and the number of the same parameter matrix (here equal to i.e. potential models), the weight matrix , is initialized as a two-dimensional matrix with 3 rows and 3 columns, the weight matrix The element in the i-th row and the j-th column of the weight matrix may be denoted as where i is the row index, is the column index, and i and j are positive integers. is a real number, and , i.e. the sum of the elements of each column of the weight matrix is 1.

[0042] Step S6: The weight matrix and the parameter matrices of the potential models are summed by weighting to obtain a new parameter matrix , and , and their calculation formulas are as follows:

[0043] ;

[0044] ;

[0045] ;

[0046] According to the new parameter matrix , and , a new aluminum electrolytic cell regression prediction model can be obtained.

[0047] Before using the differential evolution algorithm to optimize the weight matrix , the optimization objective, decision variable and constraint condition need to be determined, wherein the optimization objective is to minimize the prediction error , as shown in the following formula:

[0048] ;

[0049] wherein is the prediction error, which is a positive real number greater than or equal to 0, is the new aluminum electrolytic cell regression prediction model obtained after inputting the state vector .

[0050] The decision variable is the weight matrix​​ , the constraint condition is , , that is, the sum of each column element of the weight matrix is 1.

[0051] Next, the weight matrix is optimized using the differential evolution algorithm.

[0052] First, set the population size to be a positive integer greater than , less than , the scaling factor is generally a floating-point number greater than 0.5 and less than 1, the crossover probability is a floating-point number greater than 0.1 and less than 0.9, the convergence threshold is a floating-point number greater than 0 and less than , it is worth noting that the above four setting values can be determined according to the algorithm running situation, generate the initial population, each weight matrix in the population is a two-dimensional matrix with the same dimension as the weight matrix , and satisfies the constraint condition, denoted as , is the index of the weight matrix in the population.

[0053] Then, start the algorithm iteration, traverse the index from 1 to : randomly select three different weight matrices , and from the population, perform mutation operation on the weight matrix to generate a mutated weight matrix , the calculation formula is as follows:

[0054] .

[0055] Next, traverse the index from 1 to , traverse all elements of the th weight matrix: first generate a random number between 0 and 1, if the random number is less than the crossover probability , the element of the th row th column of the crossover weight matrix is selected from the element of the th row th column of the mutated weight matrix , otherwise, it is selected from the element of the th row th column of the weight matrix in the population.

[0056] After that, the index is traversed from 1 to , the corresponding prediction errors and are obtained according to the formula mentioned at the beginning of step S6 using the weight matrix and the weight matrix respectively. If , the weight matrix is replaced by the weight matrix , otherwise, the weight matrix is kept unchanged.

[0057] Finally, if the smallest prediction error in the population is less than the convergence threshold , that is , the differential evolution algorithm terminates the optimization of the weight matrix , and outputs the weight matrix corresponding to the smallest prediction error, and calculates the parameter matrix , and of the regression prediction model of the newly started aluminum reduction cell, and the regression prediction model of the newly started aluminum reduction cell can be obtained according to the obtained parameter matrix , in addition, the regression prediction model of the newly started aluminum reduction cell obtained here corresponds to the target domain in transfer learning.

[0058] The above only discloses the preferred embodiments of the present application, of course, cannot limit the scope of the right of the present application, therefore, the equivalent changes made according to the claims of the present application, still belongs to the scope covered by the present application.

Claims

1. A migration optimization method of an aluminum electrolysis early return prediction model, characterized by, The method comprises the following steps: S1, extracting an existing aluminum electrolysis cell regression prediction model from a model library ; ； wherein is a model number, is the total number of extracted aluminum reduction cell regression prediction models, is a positive integer greater than 1; is a membership degree center matrix, is a membership degree width matrix, is a consequent parameter matrix; S2, record the state vector of the newly started aluminum reduction cell , and input to the regression prediction model of the existing aluminum reduction cell extracted in step S1, record the first prediction value of the model The mathematical model is:​ ； wherein, is a state vector corresponding to the latest time stamp in the recorded data corresponding to the newly started aluminum electrolysis cell; is the state value in the aluminum electrolysis cell production state, is the state value in the state vector at the index, is the total number of state values in the state vector, is a positive integer greater than 1;​ S3, according to the first A model Predicted value Response value corresponding to the latest timestamp of the newly started aluminum electrolytic cell Calculate the first The prediction error value of each model And obtain the minimum value of the prediction error, that is, the minimum error value. The prediction error value is compared with the existing aluminum electrolytic cell regression prediction model extracted in step S1. One-to-one correspondence; S4, preset threshold value , traverse extracted aluminum electrolysis cell regression prediction model , if the prediction error value is less than the threshold value , mark as a potential model ; mathematical model: , is a potential model in all of the potential models, is a positive integer greater than 1; S5. initializing the weight matrix with information from the parameter matrix of the potential model ;​ S6, to minimize the prediction error For the optimization target, a differential evolution algorithm is used to optimize the weight matrix The parameter matrix of the regression prediction model of the newly started aluminum electrolysis cell is obtained by optimizing the weight matrix, and the regression prediction model of the newly started aluminum electrolysis cell can be obtained according to the obtained parameter matrix.

2. The migration optimization method of an aluminum electrolysis early return prediction model according to claim 1, characterized in that, The model library in the step S1 is a model library of existing regression prediction models of the aluminum electrolysis cell, and all the regression prediction models in the model library are built based on a fuzzy neural network.

3. The migration optimization method of an aluminum electrolysis early return prediction model according to claim 1, characterized in that, Each model in step S1 contains three parameter matrices, namely a membership center matrix , a membership width matrix and a consequent parameter matrix , wherein , and are two-dimensional floating-point matrices, each of which has a fixed positive integer value for its dimensions.

4. The migration optimization method of an aluminum electrolysis early return prediction model according to claim 1, characterized in that, In the step S2, the production state of the aluminum electrolysis cell includes but is not limited to the aluminum electrolysis cell temperature, the aluminum electrolysis cell current, the aluminum liquid level, the molecular ratio, the electrolyte level, the furnace bottom pressure drop, the aluminum electrolysis cell average voltage, the aluminum electrolysis cell working voltage, the effect frequency and the aluminum electrolysis cell age.

5. The migration optimization method of an aluminum electrolysis early return prediction model according to claim 1, characterized in that, the threshold value in the step S4 a value pre-set by the user according to experience, The larger the value, the more potential models are screened.

6. The migration optimization method of an aluminum electrolysis early return prediction model according to claim 1, characterized in that, The step S5 The parameter matrix information includes: The number of types of all parameter matrices: namely three parameter matrices; Number of parameter matrices of the same type: i.e. number of potential models ; initializing a weight matrix , representing a real number, is represented as a three-by-three two-dimensional matrix, each element of the matrix being a real number; weight matrix in the middle of the row of the column is denoted by where is the row index, is the column index, taking positive integer values; is a real number, and i.e. the sum of the elements of each column of the weight matrix is 1.

7. The migration optimization method of an aluminum electrolysis early return prediction model according to claim 1, characterized in that, The weight matrix in the step S6 With The parameter matrix of the potential model is obtained by weighted summation to obtain a new parameter matrix , And Obtain a new aluminum electrolytic cell regression prediction model ; determining an optimization goal to be minimization of prediction error , the mathematical model is: ； wherein, is a positive real number greater than or equal to 0, is a new aluminum electrolytic cell regression prediction model at an input state vector a prediction value is obtained.

Images