CN108984813B - Aluminum electrolysis modeling and optimizing method based on recurrent neural network and angle preference - Google Patents
Aluminum electrolysis modeling and optimizing method based on recurrent neural network and angle preference Download PDFInfo
- Publication number
- CN108984813B CN108984813B CN201810193126.3A CN201810193126A CN108984813B CN 108984813 B CN108984813 B CN 108984813B CN 201810193126 A CN201810193126 A CN 201810193126A CN 108984813 B CN108984813 B CN 108984813B
- Authority
- CN
- China
- Prior art keywords
- preference
- aluminum
- aluminum electrolysis
- production process
- layer
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/004—Artificial life, i.e. computing arrangements simulating life
- G06N3/006—Artificial life, i.e. computing arrangements simulating life based on simulated virtual individual or collective life forms, e.g. social simulations or particle swarm optimisation [PSO]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
Abstract
The invention provides an aluminum electrolysis modeling and optimizing method based on a recurrent neural network and angle preference. Firstly, a recurrent neural network is utilized to model an aluminum electrolysis production process, then a decision maker sets an expected target value, an A-dominance preference domination method is introduced, a multi-objective quantum particle swarm algorithm is combined to optimize a production process model, and an optimal decision variable which best meets the expectation of the decision maker, current efficiency, tank voltage, perfluorinated compound emission and aluminum energy consumption per ton are obtained. The MQPSO algorithm does not need to carry out crossing and mutation operations, and only has the simplest position updating step, so that the coding process is simple, the global search capability is strong, the integrity of preference optimal values in the population evolution process is easy to realize, and the requirements of decision makers are met. The method is used for determining the optimal value of the process parameter in the aluminum electrolysis production process, so that the current efficiency can be effectively improved, the cell voltage is reduced, the greenhouse gas emission is reduced, and the purposes of energy conservation and emission reduction are achieved.
Description
Technical Field
The invention belongs to the field of optimal control, and particularly relates to an aluminum electrolysis modeling and optimizing method based on a recurrent neural network and angle preference.
Background
The environment-friendly aluminum electrolysis production process has been very challenging for a long time, and in the aluminum electrolysis industry, the final aim is to improve the current efficiency, reduce the cell voltage, reduce perfluorinated compounds and reduce the emission of aluminum energy per ton on the basis of the stable operation of an electrolytic cell. However, the aluminum electrolysis cell has more parameters, and the parameters present nonlinearity and strong coupling, which brings great difficulty to the modeling of the aluminum electrolysis production process, and the recurrent neural network has strong nonlinear mapping capability, is suitable for solving the problem of nonlinear system modeling, and provides a new idea for the modeling of the aluminum electrolysis production process. And for four targets, the simultaneous realization is very difficult, and because the targets have a conflict phenomenon, preference information of a decision maker can be introduced, an expected target is set, weights among different targets are flexibly adjusted, and variable optimization is carried out by using a preference A-PMQPSO optimization algorithm. A-PMQPSO is based on MQPSO and introduces an A domination method. The MQPSO is a classical multi-objective optimization algorithm which is simple, high in operation speed and capable of being directly described by an equation in an evolution process, and therefore the MQPSO is widely applied to multiple fields.
Disclosure of Invention
The invention aims to overcome the defects of the prior art, provides an aluminum electrolysis modeling and optimization method based on a recurrent neural network and angle preference, aims to solve the technical problems of huge energy consumption, low efficiency and serious environmental pollution caused by failure in obtaining optimal process parameters in the aluminum electrolysis production process in the prior art, and can introduce preference information of a decision maker to dynamically and flexibly adjust preference weight among targets.
The purpose of the invention is realized by the following steps:
an aluminum electrolysis modeling and optimizing method based on a recurrent neural network and angle preference comprises the following steps:
s1: selecting control parameters having influences on current efficiency, tank voltage and perfluorinated compound emission to form decision variables, wherein the decision variables X = [ X = [) 1 ,x 2 ,···,x M ]M is the number of the selected control parameters;
s2: selecting an aluminum electrolysis industrial field, and collecting N groups of decision variables X 1 ,X 2 ,···,X N And its corresponding current efficiency y 1 ,y 2 ,···,y N Cell voltage z 1 ,z 2 ,···,z N And the amount of perfluoro compound discharged s 1 ,s 2 ,···,s N Energy consumption of aluminum per ton c 1 ,c 2 ,···,c N For data samples, each set of decision variables X i As input, in accordance withCurrent efficiency y i Cell voltage z i And the amount of perfluoro compound discharged s i And ton of aluminum consumption c i As output, training and checking the sample by using a recurrent neural network, and establishing four aluminum cell production process models;
s3: establishing a strict partial order relation based on A domination by using a preference multi-target quantum particle swarm algorithm based on A domination and taking expected values preset by a decision maker as reference points, and optimizing the four production process models obtained in the step S2 to obtain a group of decision variables X which best meet the expectation of the decision maker best And its corresponding current efficiency y best Cell voltage z best And the amount of perfluoro compound discharged s best And ton of aluminum consumption c best ；
S4: according to the optimal decision variable X obtained in the step S3 best The selected aluminum electrolysis industrial site in the step S2 is controlled by the control parameters in the step (2), so that the purposes of energy conservation, emission reduction and consumption reduction are achieved.
Preferably, in step S1, the control parameters include series current, blanking times, molecular ratio, aluminum yield, aluminum level, electrolyte level, and bath temperature.
Preferably, in step S2, the current efficiency is used as an output to establish a model of the aluminum electrolysis cell production process, wherein an input layer of the model adopts 10 neuron nodes, a hidden layer adopts 15 neuron nodes, an output layer adopts 1 neuron node, a transfer function between the input layer and the hidden layer is a Tansig function, a function between the hidden layer and the output layer is a Purelin function, and the number of iterations in sample training is 1000.
Preferably, in step S2, the cell voltage is used as an output to establish a model of the aluminum electrolysis cell production process, wherein an input layer of the model adopts 10 neuron nodes, a hidden layer adopts 15 neuron nodes, an output layer adopts 1 neuron node, a transfer function between the input layer and the hidden layer is a Logsig function, a function between the hidden layer and the output layer is a Purelin function, and the number of iterations in sample training is 1000.
Preferably, in step S2, with the perfluoro-compound emission as an output, a model of the aluminum electrolysis cell production process is established, wherein an input layer adopts 10 neuron nodes, a hidden layer adopts 15 neuron nodes, an output layer adopts 1 neuron node, a transfer function from the input layer to the hidden layer is a Logsig function, a function from the hidden layer to the output layer is a Purelin function, and the number of iterations during sample training is 1000.
Preferably, in step S2, a production process model of the aluminum electrolysis cell is established with ton aluminum energy consumption as output, an input layer of the production process model adopts 10 neuron nodes, a hidden layer adopts 15 neuron nodes, an output layer adopts 1 neuron node, a transfer function between the input layer and the hidden layer is a Tansig function, a function between the hidden layer and the output layer is a Purelin function, and the number of iterations in sample training is 1000.
Preferably, step S3 comprises the steps of:
s31: evaluating the fitness of each particle according to the preference relationship governed by A, and replacing the individual optimal value and the global optimal value according to the advantages and disadvantages;
s311: initializing system parameters including population size R, maximum iteration number T, and randomly generating n particles x 1 ,x 2 ,···,x n Making the external archive set Q empty;
s312: the decision maker sets a preference angle alpha and a preference target reference point r (y) p ,z p ,s p ,c p ) The preference target reference points comprise expected values of four targets of current efficiency, tank voltage, perfluoro compound emission and ton aluminum energy consumption;
s313: for each individual x, calculating its fitness and its angle to the reference point datum:
wherein f is j (x) Is the fitness value of the individual x on the jth target;
s314: dividing preference areas on the target space based on the angle information, and if theta (r, x) < alpha, determining that the individual is in the preference areas; otherwise, the cell is in a non-preference area;
s315: judge renMeaning two individuals x i And x k The good and bad relationship between the two cases comprises the following conditions:
when x is i And x k When in the preferred area or the non-preferred area at the same time, if x i Pareto dominate x k Then, consider x i More excellent, if they are not dominated by Pareto, they are considered to be equivalent;
when x is i In a preference area, x k In the non-preference area, if x i Pareto dominate x k Or x i And x k Are not mutually Pareto dominant, then x is considered i Is superior to x k I.e. x i A dominates x k ；
S316: determining individual historical optimal locations pbest for each particle i When the system is initialized, the individual historical optimal position is set as the initial position x of the particle i (ii) a After the next iteration, based on the A dominance relation proposed by S315, the new position x of the particle is i And pbest i Comparing the quality of the product with the quality of the product, and preserving the product as pbest i ；
S317: updating an external archive set Q, adding the archive set Q to the particles which are not dominated by A in the population, and deleting dominated particles;
s318: randomly selecting a particle in an external archive set Q as a global optimal position by utilizing a congestion mechanism and a tabu algorithm;
s32: updating the population:
s321: updating the position of the particle, wherein the updating formula of the position of the particle is as follows:
wherein: i (i =1,2, \8230;, n) represents the ith particle, n is the population size; j (j =1,2, \8230;, M) represents the jth dimension of the particle, M being the search space dimension; t is an evolution algebra;
and u ij (t) are each [0,1]Random numbers uniformly distributed in the interval; x is the number of ij (t),pbest ij (t) and γ ij (t) respectively representing the current position, the individual historical optimal position and the attractor position of the particle i when the evolution algebra is t; gbest j (t) and mbest (t) respectively represent the global optimal position and the average optimal position when the evolutionary algebra is t; α represents an expansion-contraction factor;
s322: and judging whether the current global optimal solution meets the condition or whether the iteration number reaches the maximum iteration number T, if so, outputting the current global optimal solution, otherwise, skipping to the step S321 for repeated calculation until the current global optimal solution meets the condition or the iteration number reaches the maximum iteration number T.
By adopting the technical scheme, the method utilizes the recurrent neural network to model the aluminum electrolysis production process, then a decision maker sets an expected target value, an A-dominance preference domination method is introduced, and a multi-objective quantum particle swarm algorithm is combined to optimize the production process model, so that the optimal decision variable which best meets the expectation of the decision maker, the corresponding current efficiency, the cell voltage, the perfluoro compound emission and the ton aluminum energy consumption are obtained. The MQPSO algorithm does not need to carry out cross and variation operations, and only has the simplest position updating step, so that the encoding process is simple, the global search capability is strong, the completeness of preference optimal values in the population evolution process is easy to realize, and the requirements of decision makers are met. The method is used for determining the optimal value of the process parameter in the aluminum electrolysis production process, so that the current efficiency can be effectively improved, the cell voltage is reduced, the greenhouse gas emission is reduced, and the purposes of energy conservation and emission reduction are achieved.
Drawings
FIG. 1 is a flow chart of the method of the present invention;
FIG. 2 is a graph showing the results of the estimation of the emission of CF 4;
FIG. 3 is a diagram of the predicted error of the emission of CF4
FIG. 4 is a graph of current efficiency prediction results;
FIG. 5 is a graph of current efficiency prediction error;
FIG. 6 is a graph showing the results of tank voltage discharge prediction;
FIG. 7 is a graph of tank voltage discharge prediction error;
FIG. 8 is a graph of predicted tonnage aluminum energy consumption;
FIG. 9 is a graph of ton aluminum energy consumption prediction error.
Detailed Description
As shown in fig. 1, an aluminum electrolysis modeling and optimization method based on recurrent neural network and angle preference includes the following steps:
s1: selecting control parameters having influences on current efficiency, tank voltage and perfluorinated compound emission to form decision variables, wherein the decision variables X = [ X = [) 1 ,x 2 ,···,x M ]And M is the number of the selected control parameters.
In the embodiment, original variables which have influences on current efficiency, cell voltage, perfluorinated compound emission and aluminum energy consumption per ton in the aluminum electrolysis production process are counted, and parameters which have large influences on the current efficiency, the cell voltage, the perfluorinated compound emission and the aluminum energy consumption per ton are determined as decision variables X; through statistics of measured parameters in the actual industrial production process, the maximum variables of current efficiency, tank voltage, perfluorinated compound emission and aluminum energy consumption per ton are obtained as follows: series current x 1 Number of times of blanking x 2 Molecular ratio of x 3 Aluminum output x 4 Aluminum level x 5 Electrolyte level x 6 Temperature of the bath x 7 A total of 7 variables;
s2: selecting an aluminum electrolysis industrial field, and collecting N groups of decision variables X 1 ,X 2 ,···,X N And its corresponding current efficiency y 1 ,y 2 ,···,y N Cell voltage z 1 ,z 2 ,···,z N And the amount of perfluoro compound discharged s 1 ,s 2 ,···,s N And ton of aluminum consumption c 1 ,c 2 ,···,c N As a data sample, toEach set of decision variables X i As input, with a corresponding current efficiency y i Cell voltage z i And the amount of perfluoro compound discharged s i Energy consumption of aluminum per ton c i And as output, training and checking the sample by utilizing a recurrent neural network, and establishing four aluminum electrolysis cell production process models, wherein the recurrent neural network comprises an input layer, a hidden layer and an output layer.
The four aluminum electrolytic cell production process models comprise:
for a production process model constructed by aiming at current efficiency, an input layer adopts 10 neuron nodes, a hidden layer adopts 15 neuron nodes, an output layer adopts 1 neuron node, a transfer function from the input layer to the hidden layer is a Tansig function, a function from the hidden layer to the output layer is a Purelin function, and the iteration number during sample training is 1000.
For a production process model constructed by the cell voltage, 10 neuron nodes are adopted in an input layer, 15 neuron nodes are adopted in a hidden layer, 1 neuron node is adopted in an output layer, a transfer function from the input layer to the hidden layer is a Logsig function, a function from the hidden layer to the output layer is a Purelin function, and the iteration number during sample training is 1000.
Aiming at a production process model constructed by the discharge amount of perfluorinated compounds, the production process model of the aluminum electrolytic cell is established, an input layer of the production process model adopts 10 neuron nodes, a hidden layer of the production process model adopts 15 neuron nodes, an output layer of the production process model adopts 1 neuron node, a transfer function from the input layer to the hidden layer is a Loggig function, a function from the hidden layer to the output layer is a Purelin function, and the iteration number of times during sample training is 1000.
Aiming at a production process model constructed by ton aluminum energy consumption, an aluminum electrolysis cell production process model is established, an input layer adopts 10 neuron nodes, a hidden layer adopts 15 neuron nodes, an output layer adopts 1 neuron node, a transfer function from the input layer to the hidden layer is a Tansig function, a function from the hidden layer to the output layer is a Purelin function, and the iteration number during sample training is 1000.
In order to meet the modeling requirements in this embodiment, the recurrent neural network further includes an association layer.
In the embodiment, the annual production data of the No. 223 cell electrolytic cell 2013 and the 40 days before 2014 in 170KA series electrolytic cells of Chongqing Tiantai aluminum industry Co., ltd are collected, and 405 groups of data are counted, wherein the annual production data of 2013 serves as a modeling training sample, and the 40 groups of data of 2014 serve as a test sample. Data samples are shown in table 1 below.
TABLE 1 data sample
| Sample numbering | 1 | 2 | 3 | 4 | …… |
| x 1 | 1683 | 1682 | 1686 | 1746 | …… |
| x 2 | 624 | 716 | 625 | 743 | …… |
| x 3 | 2.52 | 2.52 | 2.51 | 2.46 | …… |
| x 4 | 1234 | 1230 | 1234 | 1235 | …… |
| x 5 | 18.5 | 16.5 | 17.5 | 20 | …… |
| x 6 | 14 | 14 | 15 | 16 | …… |
| x 7 | 942 | 938 | 946 | 942 | …… |
| y 1 | 94.65 | 94.66 | 94.43 | 93.22 | …… |
| y 2 | 3721 | 3720 | 3725 | 3717 | …… |
| y 3 | 4.25 | 4.84 | 4.01 | 4.15 | …… |
| y 4 | 12354.3 | 12316.4 | 12283.1 | 12747.2 | …… |
In the design of the recurrent neural network, because of the existence of recurrent signals, the network state changes along with the change of time, so that the learning rate also influences the stability and the accuracy of the neural network model besides the number of hidden nodes, which is a serious difficulty in the design of the neural network.
The setting of the number of nodes of the hidden layer is obtained by a trial and error method:
wherein p is the number of nodes of hidden layer neurons, n is the number of neurons in input layer, m is the number of neurons in output layer, and k is a constant between 1 and 10.
The optimal learning rate takes values as:
the setup parameters of the recurrent neural network in this example are shown in table 2 below.
TABLE 2 recurrent neural network setup parameters
| Objective function | Current efficiency | Cell voltage | Total fluoride discharge | Ton aluminium energy consumption |
| Number of iterations | 1000 | 1000 | 1000 | 1000 |
| Implicit layer transfer function | Tansig | Logsig | Logsig | Tansig |
| Output layer transfer function | Purelin | Purelin | Purelin | Purelin |
| Number of hidden layer nodes | 13 | 12 | 12 | 13 |
The training process of the neural network is mainly carried out according to the following steps:
set up X k ＝[x k1 ,x k2 ,···,x kM ](k =1,2, ·, N) is an input vector, N is the number of training samples,
the step S2 of establishing the aluminum electrolysis production process model specifically comprises the following steps:
s21: initializing, setting the initial value of the iteration times g as 0,W MI (0)、W JP (0) All are random values in the interval of (0, 1);
s22: random input sample X k
S23: for input sample X k Forward computing input and output signals for each layer of neurons in a recurrent neural network
S24: output d according to desire k And the actual output Y k (g) Calculating an error E (g);
s25: judging whether the error E (g) meets the requirement, if not, entering a step S26, and if so, entering a step S29;
s26: judging whether the iteration number g +1 is greater than the maximum iteration number, if so, entering a step S29, otherwise, entering a step S27;
s27: for input sample X k Calculating the local gradient of each layer of neurons in a reverse mode;
the network output layer node error is: e (k) = d (k) -y (k), e (k) being the network expected output and y (k) being the network actual output.
The weight change rate of each layer by calculating the node error of the output layer is as follows:
wherein beta is ij (0)＝0；i＝1,2,···,n 1 ；j＝1,2,···,n 0 。
δ i (0)＝0；i＝1,2,···,n 1 。
Wherein
Respectively representing the ith of the hidden layerInput and output of nodes; n is a radical of an alkyl radical 0 、n 1 Respectively the number of nodes of the output layer and the hidden layer;
respectively representing the weight of the associated layer, the output layer and the hidden layer.
S28: the network weight correction calculation formula is as follows:
wherein w (k) can be
In the formula, w (k) may represent a weight of an output layer, a hidden layer or an input layer, η is a learning rate, g = g +1, and step S23 is skipped;
s29: and judging whether all training samples are finished, if so, finishing modeling, and otherwise, continuing to step S22.
Through the above loop process, the prediction effect of the recurrent neural network can be obtained as shown in fig. 2, 3, 4, 5, 6, 7, 8 and 9. The establishment of the optimization model is the basis of the optimization of the aluminum electrolysis production process, and the accuracy of the model directly influences the optimization result. By analyzing the graphs in fig. 2, 3, 4, 5, 6, 7, 8 and 9, the maximum prediction error of the current efficiency is 0.41 percent, the maximum prediction error of the cell voltage is 0.08 percent, the prediction error of the carbon tetrafluoride CF4 emission is-1.20 percent, the prediction error of the ton aluminum energy consumption is 0.81 percent, the model prediction precision is high, and the modeling requirement is met through the training of the recurrent neural network.
S3: introducing an A-dominance preference domination method, establishing a strict partial order relation based on A domination according to expected values (reference points) preset by a decision maker by utilizing a preference multi-target quantum particle swarm algorithm based on A domination and combining a multi-target quantum particle swarm algorithm (MQPSO), namely an A-PMQPSO algorithm, optimizing the four production process models obtained in the step S2 to obtain a group of decision variables X which best meet the expectation of the decision maker best And its corresponding current efficiency y best Cell voltage z best And the amount of perfluoro compound discharged s best And ton of aluminum consumption c best 。
On the basis of an aluminum electrolysis production process model, the aluminum electrolysis production process model is optimized in each decision variable range by utilizing an A-PMQPSO algorithm, and the specific variation range of each variable is shown in a table 3.
TABLE 3 value ranges of variables
In step S3, the a-PMQPSO algorithm includes the following steps:
s31: evaluating the fitness of each particle according to the preference relationship governed by A, and replacing the individual optimal value and the global optimal value according to the advantages and the disadvantages;
s311: initializing system parameters including population size R, maximum iteration number T, and randomly generating n particles x 1 ,x 2 ,···,x n Making the external archive set Q empty;
s312: the decision maker sets a preference angle alpha and a preference target reference point r (y) p ,z p ,s p ,c p ) The preference target reference points comprise expected values of four targets of current efficiency, tank voltage, perfluoro compound emission and ton aluminum energy consumption;
s313: for each individual x, calculating the fitness and the angle between the fitness and the reference point datum line:
wherein f is j (x) Is the fitness value of the individual x on the jth target;
s314: dividing preference areas on the target space based on the angle information, if theta (r, x) < alpha, then the individual is in the preference area; otherwise, the mobile terminal is in a non-preference area;
s315: judging any two individuals x i And x k The good and bad relationship between the two cases comprises the following conditions:
when x is i And x k When in the preferred area or the non-preferred area, if x i Pareto dominate x k Then, consider x i More excellent, if they are not dominated by Pareto, they are considered to be equivalent;
when x is i In a preference area, x k In the non-preference area, if x i Pareto dominate x k Or x i And x k Are not mutually dominated by Pareto, then x is considered i Is superior to x k I.e. x i A dominates x k ；
S316: determining individual historical optimal locations for each particle pbest i At the time of system initialization, the individual historical optimal position is set as the initial position x of the particle i (ii) a After the next iteration, based on the A dominance relation proposed by S315, the new position x of the particle is i And pbest i Comparing the quality of the product with the quality of the product, and preserving the product as pbest i ；
S317: updating an external archive set Q, adding the archive set Q to the particles which are not dominated by A in the population, and deleting the dominated particles;
s318: randomly selecting a particle in an external archive set Q as a global optimal position by using a congestion mechanism and a tabu algorithm;
s32: updating the population:
s321: updating the position of the particle, wherein the updating formula of the position of the particle is as follows:
wherein: i (i =1,2, \8230;, n) represents the ith particle, n is the population size; j (j =)1,2, \8230;, M) represents the j-th dimension of the particle, M being the search space dimension; t is an evolution algebra;
and u ij (t) are each [0,1]Random numbers uniformly distributed in the interval; x is a radical of a fluorine atom ij (t),pbest ij (t) and γ ij (t) respectively representing the current position of the particle i, the individual historical optimal position and the attractor position when the evolution algebra is t; gbest j (t) and mbest (t) respectively represent the global optimal position and the average best position when the evolution algebra is t; α represents an expansion-contraction factor;
s322: and judging whether the current global optimal solution meets the condition or whether the iteration number reaches the maximum iteration number T, if so, outputting the current global optimal solution, otherwise, skipping to the step S321 to perform repeated calculation until the current global optimal solution meets the condition or the iteration number reaches the maximum iteration number T.
The aluminum electrolysis production process is optimized through the steps to obtain 100 groups of optimal decision variables and corresponding output values, and the most reasonable 3 groups are selected and listed in the following table 4.
TABLE 4 optimum production parameters
| y 1 | y 2 | y 3 | y 4 | x 1 | x 2 | x 3 | x 4 | x 5 | x 6 | x 7 |
| 99.14 | 3635 | 3.65 | 10835.15 | 1649 | 628 | 2.54 | 1210 | 16.5 | 14.5 | 942 |
| 98.13 | 3682 | 3.59 | 11527.21 | 1652 | 626 | 2.38 | 1200 | 17.5 | 15 | 925 |
| 95.37 | 3602 | 3.68 | 10478.32 | 1674 | 617 | 2.47 | 1095 | 17.5 | 15.5 | 935 |
By comparing the optimal operation parameters with the average value recorded all year round in 2013, the current efficiency is improved by 3.89%, the cell voltage is reduced by 160mv, the CF4 emission is reduced by 0.39kg, and the energy consumption per ton of aluminum is reduced by 1219.07KWh/t-Al.
S4: according to the optimal decision variable X obtained in the step S3 best The selected aluminum electrolysis industrial site in the step S2 is controlled by the control parameters in the step (2), so that the purposes of energy conservation, emission reduction and consumption reduction are achieved.
In the embodiment of the application, an intelligent control method for energy conservation and emission reduction of aluminum electrolysis based on the A domination relationship is provided, firstly, a recursive neural network is used for modeling an aluminum electrolysis production process, then a decision maker sets an expected target value, an A-dominance preference domination method is introduced, a production process model is optimized by combining a multi-target quantum particle swarm algorithm (MQPSO), and an optimal decision variable which meets the expectation of the decision maker most, and corresponding current efficiency, cell voltage, perfluorinated emission and ton aluminum energy consumption are obtained. The MQPSO algorithm does not need complex operations such as crossing, variation and the like, and only has the simplest position updating step, so the encoding process is simple, and quantum characteristics are introduced, so that particles have strong global search capability, the integrity of preference optimal values in the population evolution process is easily ensured, and the preference requirement of decision makers is met. The method is used for obtaining the optimal value of the process parameters in the aluminum electrolysis production process, can effectively improve the current efficiency, reduce the cell voltage, reduce the greenhouse gas emission and aluminum energy consumption per ton, and achieve the purposes of energy conservation and emission reduction.
Finally, it is noted that the above-mentioned preferred embodiments illustrate rather than limit the invention, and that, although the invention has been described in detail with reference to the above-mentioned preferred embodiments, it will be understood by those skilled in the art that various changes in form and detail may be made therein without departing from the scope of the invention as defined by the appended claims.
Claims (6)
1. An aluminum electrolysis modeling and optimization method based on a recurrent neural network and angle preference is characterized by comprising the following steps:
s1: selecting control parameters having influences on current efficiency, tank voltage and perfluorinated compound emission to form decision variables, wherein the decision variables X = [ X = [) 1 ,x 2 ,···,x M ]M is the number of the selected control parameters;
s2: selecting an aluminum electrolysis industrial field, and collecting N groups of decision variables X 1 ,X 2 ,···,X N And its corresponding current efficiency y 1 ,y 2 ,···,y N Cell voltage z 1 ,z 2 ,···,z N And the amount of perfluoro compound discharged s 1 ,s 2 ,···,s N And ton of aluminum consumption c 1 ,c 2 ,···,c N For data samples, each set of decision variables X i As input, with a corresponding current efficiency y i Cell voltage z i And the amount of perfluoro compound discharged s i And ton of aluminum consumption c i As output, training and checking the sample by using a recurrent neural network, and establishing four aluminum electrolytic cell production process models;
s3: establishing a strict partial order relation based on A domination by using a preference multi-target quantum particle swarm algorithm based on A domination and according to an expected value preset by a decision maker as a reference point, optimizing the four production process models obtained in the step S2 to obtain a group of decision variables X which best meet the expectation of the decision maker best And its corresponding current efficiency y best Cell voltage z best And the amount of perfluoro compound discharged s best Energy consumption of aluminum per ton c best ；
Step S3 includes the following steps:
s31: evaluating the fitness of each particle according to the preference relationship governed by A, and replacing the individual optimal value and the global optimal value according to the advantages and disadvantages;
s311: initializing system parameters including population size R, maximum iteration number T, and randomly generating n particles x 1 ,x 2 ,···,x n Making the external archive set Q empty;
s312: the decision maker sets a preference angle alpha and a preference target reference point r (y) p ,z p ,s p ,c p ) The preference target reference points comprise expected values of four targets of current efficiency, tank voltage, perfluoro compound emission and ton aluminum energy consumption;
s313: for each individual x, calculating its fitness and its angle to the reference point datum:
wherein, f j (x) Is the fitness value of the individual x on the jth target;
s314: dividing preference areas on the target space based on the angle information, and if theta (r, x) < alpha, determining that the individual is in the preference areas; otherwise, the cell is in a non-preference area;
s315: judging any two individuals x i And x k The good and bad relation between the two comprises the following conditions:
when x is i And x k When in the preferred area or the non-preferred area, if x i Pareto dominate x k Then x is considered to be i More excellent, if they are not dominated by Pareto, they are considered to be equivalent;
when x is i In a preference area, x k In the non-preference area, if x i Pareto dominate x k Or x i And x k Are not mutually Pareto dominant, then x is considered i Is superior to x k I.e. x i A dominating x k ；
S316: determining individual historical optimal locations for each particle pbest i At the time of system initialization, the individual historical optimal position is set as the initial position x of the particle i (ii) a After the next iteration, based on the A dominance relation proposed by S315, the new position x of the particle is i And pbest i Comparing the quality of the product with the quality of the product, and preserving the product as pbest i ；
S317: updating an external archive set Q, adding the archive set Q to the particles which are not dominated by A in the population, and deleting dominated particles;
s318: randomly selecting a particle in an external archive set Q as a global optimal position by using a congestion mechanism and a tabu algorithm;
s32: updating the population:
s321: updating the position of the particle, wherein the updating formula of the position of the particle is as follows:
wherein: i (i =1,2, \8230;, n) represents the ith particle, and n is the population size; j (j =1,2, \8230;, M) represents the jth dimension of the particle, M being the search space dimension; t is an evolution algebra;
and u ij (t) are each [0,1 ]]Random numbers uniformly distributed in the interval; x is a radical of a fluorine atom ij (t),pbest ij (t) and γ ij (t) respectively representing the current position, the individual historical optimal position and the attractor position of the particle i when the evolution algebra is t; gbest j (t) and mbest (t) denote evolution, respectivelyWhen the algebra is t, the global optimal position and the average best position; α represents an expansion-contraction factor;
s322: judging whether the current global optimal solution meets the condition or whether the iteration number reaches the maximum iteration number T, if so, outputting the current global optimal solution, otherwise, skipping to the step S321 for repeated calculation until the current global optimal solution meets the condition or the iteration number reaches the maximum iteration number T;
s4: according to the optimal decision variable X obtained in the step S3 best The selected aluminum electrolysis industrial site in the step S2 is controlled by the control parameters in the step (2), so that the purposes of energy conservation, emission reduction and consumption reduction are achieved.
2. The aluminum electrolysis modeling and optimization method based on the recurrent neural network and the angle preference as claimed in claim 1, wherein in step S1, the control parameters include series current, blanking times, molecular ratio, aluminum yield, aluminum level, electrolyte level, and bath temperature.
3. The aluminum electrolysis modeling and optimizing method based on the recurrent neural network and the angle preference as claimed in claim 1, wherein in step S2, a current efficiency is used as an output to establish an aluminum electrolysis cell production process model, an input layer of the model adopts 10 neuron nodes, a hidden layer of the model adopts 15 neuron nodes, an output layer of the model adopts 1 neuron node, a transfer function from the input layer to the hidden layer is a Tansig function, a function from the hidden layer to the output layer is a Purelin function, and the number of iterations in sample training is 1000.
4. The aluminum electrolysis modeling and optimizing method based on the recurrent neural network and the angle preference as claimed in claim 1, wherein in step S2, a cell voltage is used as an output to establish an aluminum electrolysis cell production process model, an input layer of the aluminum electrolysis cell production process model adopts 10 neuron nodes, a hidden layer of the aluminum electrolysis cell production process model adopts 15 neuron nodes, an output layer of the aluminum electrolysis cell production process model adopts 1 neuron node, a transfer function from the input layer to the hidden layer is a Logsig function, a function from the hidden layer to the output layer is a Purelin function, and the number of iterations in sample training is 1000.
5. The aluminum electrolysis modeling and optimizing method based on the recurrent neural network and the angle preference as claimed in claim 1, wherein in step S2, the model of the aluminum electrolysis cell production process is established by taking the amount of perfluoro compounds as output, the input layer adopts 10 neuron nodes, the hidden layer adopts 15 neuron nodes, the output layer adopts 1 neuron node, the transfer function from the input layer to the hidden layer is Logsig function, the function from the hidden layer to the output layer is Purelin function, and the number of iterations in sample training is 1000.
6. The aluminum electrolysis modeling and optimizing method based on the recurrent neural network and the angle preference as claimed in claim 1, wherein in step S2, an aluminum electrolysis cell production process model is established with ton aluminum energy consumption as output, an input layer of the model adopts 10 neuron nodes, a hidden layer of the model adopts 15 neuron nodes, an output layer of the model adopts 1 neuron node, a transfer function from the input layer to the hidden layer is a Tansig function, a function from the hidden layer to the output layer is a Purelin function, and the number of iterations in sample training is 1000.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810193126.3A CN108984813B (en) | 2018-03-09 | 2018-03-09 | Aluminum electrolysis modeling and optimizing method based on recurrent neural network and angle preference |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810193126.3A CN108984813B (en) | 2018-03-09 | 2018-03-09 | Aluminum electrolysis modeling and optimizing method based on recurrent neural network and angle preference |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108984813A CN108984813A (en) | 2018-12-11 |
| CN108984813B true CN108984813B (en) | 2022-12-13 |
Family
ID=64541753
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810193126.3A Active CN108984813B (en) | 2018-03-09 | 2018-03-09 | Aluminum electrolysis modeling and optimizing method based on recurrent neural network and angle preference |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108984813B (en) |
Family Cites Families (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP4557397B2 (en) * | 2000-09-05 | 2010-10-06 | 本田技研工業株式会社 | Blade shape design method and information medium |
| CN103903072A (en) * | 2014-04-17 | 2014-07-02 | 中国矿业大学 | High-dimensional multi-target set evolutionary optimization method based on preference of decision maker |
| CN105207573A (en) * | 2015-09-04 | 2015-12-30 | 浙江大学 | Quantitative optimal configuration method of wind-solar hybrid power system based on discrete probability model |
| CN105447567B (en) * | 2015-11-06 | 2017-12-05 | 重庆科技学院 | Aluminium electroloysis energy-saving and emission-reduction control method based on BP neural network Yu MPSO algorithms |
| CN105404926B (en) * | 2015-11-06 | 2017-12-05 | 重庆科技学院 | Aluminum electrolysis production technique optimization method based on BP neural network Yu MBFO algorithms |
| CN106886656B (en) * | 2017-03-15 | 2020-12-25 | 南京航空航天大学 | Three-dimensional array antenna directional pattern sidelobe suppression method |
-
2018
- 2018-03-09 CN CN201810193126.3A patent/CN108984813B/en active Active
Also Published As
| Publication number | Publication date |
|---|---|
| CN108984813A (en) | 2018-12-11 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109085752B (en) | Aluminum electrolysis preference multi-objective optimization algorithm based on angle domination relationship | |
| CN105447567B (en) | Aluminium electroloysis energy-saving and emission-reduction control method based on BP neural network Yu MPSO algorithms | |
| CN108445756B (en) | Aluminum electrolysis energy-saving emission-reduction intelligent control method based on AR domination relationship | |
| CN105404926B (en) | Aluminum electrolysis production technique optimization method based on BP neural network Yu MBFO algorithms | |
| CN112564098B (en) | High-proportion photovoltaic power distribution network voltage prediction method based on time convolution neural network | |
| CN110782658B (en) | Traffic prediction method based on LightGBM algorithm | |
| CN105302973A (en) | MOEA/D algorithm based aluminum electrolysis production optimization method | |
| CN108846526A (en) | A kind of CO2 emissions prediction technique | |
| CN105321000A (en) | Aluminum electrolytic process parameter optimization method based on BP neural network and MOBFOA algorithm | |
| CN107992645B (en) | Sewage treatment process soft measurement modeling method based on chaos-firework hybrid algorithm | |
| CN105404142B (en) | Aluminium electroloysis multi parameters control method based on BP neural network Yu MBFO algorithms | |
| CN109086469B (en) | Aluminum electrolysis modeling and optimizing method based on recurrent neural network and preference information | |
| CN114169251A (en) | Ultra-short-term wind power prediction method | |
| CN111832817A (en) | Small world echo state network time sequence prediction method based on MCP penalty function | |
| CN105426959B (en) | Aluminium electroloysis energy-saving and emission-reduction method based on BP neural network Yu adaptive M BFO algorithms | |
| CN109100995B (en) | Aluminum electrolysis energy-saving emission-reduction optimization method based on preference information of decision maker | |
| CN106021698A (en) | Iterative updating-based UKFNN aluminum electrolysis power consumption model construction method | |
| CN115689070A (en) | Energy prediction method for optimizing BP neural network model based on imperial butterfly algorithm | |
| CN108363303B (en) | Differential evolution aluminum electrolysis multi-objective optimization method based on AR preference information | |
| CN105302976A (en) | Aluminum electrolysis production optimization method based on SPEA2 algorithm | |
| CN108984813B (en) | Aluminum electrolysis modeling and optimizing method based on recurrent neural network and angle preference | |
| CN105334824A (en) | Aluminum electrolysis production optimization method based on NSGA-II algorithm | |
| CN112434888A (en) | PM2.5 prediction method of bidirectional long and short term memory network based on deep learning | |
| CN105420760B (en) | Aluminium electroloysis multi-parameters optimization method based on adaptive step bacterial foraging algorithm | |
| CN105426960B (en) | Aluminium electroloysis energy-saving and emission-reduction control method based on BP neural network Yu MBFO algorithms |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |