CN104076686A - Method for controlling aluminum oxide production process dynamic cost - Google Patents

Abstract

The invention discloses a method for controlling aluminum oxide production process dynamic cost. The method includes the following steps of firstly, cost index prediction and secondly, cost driver analysis. An aluminum oxide cost neural network prediction method is adopted in the first step and involves an aluminum oxide cost composition model and an aluminum oxide cost prediction model; each cost index of the aluminum oxide cost composition model further comprises multiple cost factors, and aluminum oxide cost compositions are represented in a matrix mode. According to the method, cost index unit consumption is predicted through a cost index unit consumption prediction model, feedforward control is conducted on production process cost, the difference between a set value and an actual value is analyzed from the production index angle, feedback control over the production process cost is achieved by readjusting the production index set value or rewarding and punishing control cost, and therefore the purposes of consumption reduction and production cost control are achieved.

Description

A kind of aluminum oxide production process Dynamic Cost Control method

Technical field

The invention belongs to automatic control technology field, particularly a kind of aluminum oxide production process Dynamic Cost Control method.

Background technology

Traditional alumina producing enterprise is in the time carrying out the cost control of production phase, only find from aspects such as price differential, poor, the physical variation of amount and labour variances the method reducing costs, cost control is all the waiting cost control of implementing by modes such as forms in accounting for management's category.Because scene lacks the support of infotech, each process cost information that cannot real-time follow-up production run, the feedback of formation information in time, the production cost control of alumina producing enterprise is a kind of control afterwards of hysteresis.Cost control in the past is often confined to collection, the calculating etc. of data, the physical form that does not realize each production element is synchronizeed mobile with the form of value, only pay attention to result and despise process control, cost control and process control are asynchronous, economic target and technical indicator and operating process are not organically combined, make cost control lack operability.The ratio that in the cost of aluminium oxide, the expense of the consumption such as starting material, fuel, power accounts for cost is about 80% left and right.Wherein supplies consumption is mainly reflected in the technical economical index level of each operation.Enterprise is formulating a period when plan target, the not science of Management Technology index that some department provides, and treasurer's department is just difficult to make a practicable performance assessment criteria, makes cost control enforceability poor.Production run dynamic cost and energy consumption, material consumption etc. have multivariate strong coupling, strong nonlinearity, uncertain by force, are difficult to describe by conventional mathematical model, are difficult to set up production run dynamic control model by existing modeling method.

Summary of the invention

For existing aluminum oxide production process Dynamic Cost Control difficult problem, the present invention proposes a kind of aluminum oxide production process Dynamic Cost Control method.

The technical solution adopted in the present invention is:

A kind of aluminum oxide production process Dynamic Cost Control method, comprises the following steps:

Step 1, indicator of costs forecast;

Step 2, Cost Driver analysis.

Described step 1 adopts aluminium oxide cost neural network prediction method;

The method comprises: aluminium oxide cost composition model and aluminium oxide cost forecasting model;

Each indicator of costs of described aluminium oxide cost composition model comprises again many cost factors, taking matrix representation aluminium oxide cost composition model as:

$Y_{m\×l}=\left[\begin{array}{c}\stackrel{\‾}{y_1}\\ \stackrel{\‾}{y_2}\\ .\\ .\\ .\\ \stackrel{\‾}{y_m}\end{array}\right]=\left[\begin{array}{cccc}{y}_{11}& y_{12}& \·\·\·& y_{1l}\\ y_{12}& y_{22}& \·\·\·& y_{2l}\\ .& & & \\ .& & & \\ .& & & \\ y_{m1}& y_{m2}& \·\·\·& y_{\mathrm{ml}}\end{array}\right]$ Formula 1

$l=\max \{\dim \left(\stackrel{\‾}{y_1}\right),\dim \left(\stackrel{\‾}{y_2}\right),...,\dim \left(\stackrel{\‾}{y_m}\right)\}$ Formula 2

Wherein represent to form i indicator of costs vector of alumina product cost; y _irepresent to form i the indicator of costs of cost of products; y _ijrepresent j cost factor of i indicator of costs vector.

In described formula 2, if ? be equal to zero, think that i the indicator of costs is made up of q cost factor;

Aluminium oxide cost can be expressed as:

$y=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{y}_i=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}\underset{j=1}{\overset{l}{\mathrm{\Σ}}}{y}_{\mathrm{ij}}$ Formula 3

Each index aluminium oxide cost can be expressed as:

$y_i=\underset{j=1}{\overset{l}{\mathrm{\Σ}}}{y}_{\mathrm{ij}}$ Formula 4

Described aluminium oxide cost forecasting model is based on BP neural net method, establishes m the aluminium oxide indicator of costs and interacts, and comprehensive history value forecast model and influence factor forecast model, set up many inputs of many index predictions, many output models, and model is as follows:

$\left[\begin{array}{c}{y}_1^{\′}\left(t\right)\\ y_2^{\′}\left(t\right)\\ .\\ .\\ .\\ y_m^{\′}\left(t\right)\end{array}\right]=\left[\begin{array}{c}{f}_1(x_1(t-k),x_1(t-k+1),\·\·\·,x_1(t-1),\·\·\·,x_m(t-k),x_m(t-k+1),\·\·\·,x_m(t-1))\\ f_2(x_1(t-k),x_1(t-k+1),\·\·\·,x_1(t-1),\·\·\·,x_m(t-k),x_m(t-k+1),\·\·\·,x_m(t-1))\\ .\\ .\\ .\\ f_m(x_1(t-k),x_1(t-k+1),\·\·\·,x_1(t-1),\·\·\·,x_m(t-k),x_m(t-k+1),\·\·\·,x_m(t-1))\end{array}\right]$ Formula 5

Wherein x _i, be respectively neural network input, output, x _irepresent the historical unit consumption of the interactive indicator of costs, represent the prediction unit consumption of the indicator of costs; K represents the number (in chronological order) of historical data that each index is got.

Described aluminium oxide cost neural network Cost Analysis Method, X in Cost Analysis Model ₁, X ₂..., X _nfor N the influence factor of cost Y, as neural network input, cost Y exports as BP neural network;

First BP neural network cost analysis algorithm carries out pre-service to the data in sample set, data is carried out to nondimensionalization or normalized, makes it be transformed to the number between [0,1];

Secondly,, by pretreated sample set training BP neural network, set up the model Y=F (X) of relation between reflection independent variable and dependent variable; In the time of training BP neural network, to only have in the time that the weight space data of neural network tend towards stability or be constant, training is just calculated and is finished;

Then,, according to the BP neural network of setting up, obtain the corresponding elasticity number of each input factor; The input feature value that is the sample k in pretreated sample set is h=1,2 ..., n, k=1,2 ..., m, wherein n is the component number (the index factor number of representative for analyzing) of sample input feature value, m is the contained number of samples of sample set; Input feature value using lower column vector as built BP network:

$(x_1^k,x_2^k,\·\·\·,x_h^k,\·\·\·,x_n^k)$

$(x_1^k+\mathrm{\Δx},x_2^k,\·\·\·,x_h^k,\·\·\·,x_n^k)$

$\begin{array}{c}(x_1^k,x_2^k+\mathrm{\Δx},\·\·\·,x_h^k,\·\·\·x_n^k)\\ .\\ .\\ .\end{array}$ Formula 6

$\begin{array}{c}(x_1^k,x_2^k,\·\·\·,x_h^k+\mathrm{\Δx},\·\·\·,x_n^k)\\ .\\ .\\ .\end{array}$

$(x_1^k,x_2^k,\·\·\·,x_h^k,\·\·\·,x_n^k+\mathrm{\Δx})$

BP neural network is inputted according to these, obtains corresponding output valve (according to needs herein, the output characteristic vector is here only containing one-component), is designated as respectively y ^kfor using the input feature value of sample k as input, the output valve that network is obtained; while remaining unchanged for adding all the other component values of Δ x when h component of the input feature value of sample k, the output valve that network is obtained;

Note for the proper vector of sample k the corresponding input factor of component at the flexible absolute value of sample point k:

$E_h^k=\frac{|(y_h^k-y^k)/y^k|}{|\mathrm{\Δx}/x_h^k|}$ Formula 7

Wherein, h=1,2 ..., n, k=1,2 ... m

Ask ${E_h^{*}}^k=\max \{E_1^k,E_2^k,\·\·\·,E_h^k,\·\·\·E_n^k\};$

With remove respectively ?

Finally, by corresponding each input factor add up, and carry out factor analysis accordingly; Note S _hfor input factor, h is corresponding accumulated value:

$S_h=\underset{k=1}{\overset{m}{\mathrm{\Σ}}}{E}_h^{\′k}$ Formula 8

Note F _hfor the degree of affecting of input factor h on output y,

$F_h=\frac{S_h}{\underset{h=1}{\overset{n}{\mathrm{\Σ}}}{S}_h}$ Formula 9

According to F _hsize judge the size of influence degree of input factor h to output factor; F _hgreatly, its corresponding input factor is just large to the influence degree of output factor, and vice versa; Just can be in the hope of each index factor x by above-mentioned algorithm _ion degree of the impact F of output variable y _h.

Advantage of the present invention is:

The present invention utilizes indicator of costs unit consumption forecasting model to forecast indicator of costs unit consumption, production run cost is carried out to feedforward control, and from production target angle analysis setting value and actual value difference before, by readjusting production target setting value or to the rewards and punishments of carrying out of controlling cost, realize the FEEDBACK CONTROL of production run cost, reduce consumption, production control cost object thereby reach.

Brief description of the drawings

Fig. 1 aluminum oxide production process cost control system of the present invention technical architecture plan;

Fig. 2 neural network cost forecast model of the present invention figure;

Fig. 3 three layers of BP neural network diagram of the present invention;

The process flow diagram of Fig. 4 GA of the present invention and BP combination;

Fig. 5 neural network Cost Analysis Model of the present invention figure;

Fig. 6 cost analysis process flow diagram based on BP network of the present invention.

Embodiment

As shown in accompanying drawing Fig. 1-6, a kind of aluminum oxide production process Dynamic Cost Control method, is characterized in that, comprises the following steps: step 1, indicator of costs forecast; Step 2, Cost Driver analysis.

Described step 1 adopts aluminium oxide cost neural network prediction method;

The method comprises: aluminium oxide cost composition model and aluminium oxide cost forecasting model;

Each indicator of costs of described aluminium oxide cost composition model comprises again many cost factors, taking matrix representation aluminium oxide cost composition model as:

$Y_{m\×l}=\left[\begin{array}{c}\stackrel{\‾}{y_1}\\ \stackrel{\‾}{y_2}\\ .\\ .\\ .\\ \stackrel{\‾}{y_m}\end{array}\right]=\left[\begin{array}{cccc}{y}_{11}& y_{12}& \·\·\·& y_{1l}\\ y_{12}& y_{22}& \·\·\·& y_{2l}\\ .& & & \\ .& & & \\ .& & & \\ y_{m1}& y_{m2}& \·\·\·& y_{\mathrm{ml}}\end{array}\right]$ Formula 1

$l=\max \{\dim \left(\stackrel{\‾}{y_1}\right),\dim \left(\stackrel{\‾}{y_2}\right),...,\dim \left(\stackrel{\‾}{y_m}\right)\}$ Formula 2

Wherein represent to form i indicator of costs vector of alumina product cost; y _irepresent to form i the indicator of costs of cost of products; y _ijrepresent j cost factor of i indicator of costs vector;

In described formula 2, if ? be equal to zero, think that i the indicator of costs is made up of q cost factor;

Aluminium oxide cost can be expressed as:

$y=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{y}_i=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}\underset{j=1}{\overset{l}{\mathrm{\Σ}}}{y}_{\mathrm{ij}}$ Formula 3

Each index aluminium oxide cost can be expressed as:

$y_i=\underset{j=1}{\overset{l}{\mathrm{\Σ}}}{y}_{\mathrm{ij}}$ Formula 4

Described aluminium oxide cost forecasting model is based on BP neural net method, establishes m the aluminium oxide indicator of costs and interacts, and comprehensive history value forecast model and influence factor forecast model, set up many inputs of many index predictions, many output models, and model is as follows:

$\left[\begin{array}{c}{y}_1^{\′}\left(t\right)\\ y_2^{\′}\left(t\right)\\ .\\ .\\ .\\ y_m^{\′}\left(t\right)\end{array}\right]=\left[\begin{array}{c}{f}_1(x_1(t-k),x_1(t-k+1),\·\·\·,x_1(t-1),\·\·\·,x_m(t-k),x_m(t-k+1),\·\·\·,x_m(t-1))\\ f_2(x_1(t-k),x_1(t-k+1),\·\·\·,x_1(t-1),\·\·\·,x_m(t-k),x_m(t-k+1),\·\·\·,x_m(t-1))\\ .\\ .\\ .\\ f_m(x_1(t-k),x_1(t-k+1),\·\·\·,x_1(t-1),\·\·\·,x_m(t-k),x_m(t-k+1),\·\·\·,x_m(t-1))\end{array}\right]$ Formula 5

Wherein x _i, be respectively neural network input, output, x _irepresent the historical unit consumption of the interactive indicator of costs, represent the prediction unit consumption of the indicator of costs; K represents the number (in chronological order) of historical data that each index is got.

Described aluminium oxide cost neural network Cost Analysis Method, X in Cost Analysis Model ₁, X ₂..., X _nfor N the influence factor of cost Y, as neural network input, cost Y exports as BP neural network;

First BP neural network cost analysis algorithm carries out pre-service to the data in sample set, data is carried out to nondimensionalization or normalized, makes it be transformed to the number between [0,1];

Secondly,, by pretreated sample set training BP neural network, set up the model Y=F (X) of relation between reflection independent variable and dependent variable; In the time of training BP neural network, to only have in the time that the weight space data of neural network tend towards stability or be constant, training is just calculated and is finished;

Then,, according to the BP neural network of setting up, obtain the corresponding elasticity number of each input factor; The input feature value that is the sample k in pretreated sample set is h=1,2 ..., n, k=1,2 ..., m, wherein n is the component number (the index factor number of representative for analyzing) of sample input feature value, m is the contained number of samples of sample set; Input feature value using lower column vector as built BP network:

$(x_1^k,x_2^k,\·\·\·,x_h^k,\·\·\·,x_n^k)$

$(x_1^k+\mathrm{\Δx},x_2^k,\·\·\·,x_h^k,\·\·\·,x_n^k)$

$\begin{array}{c}(x_1^k,x_2^k+\mathrm{\Δx},\·\·\·,x_h^k,\·\·\·x_n^k)\\ .\\ .\\ .\end{array}$ Formula 6

$\begin{array}{c}(x_1^k,x_2^k,\·\·\·,x_h^k+\mathrm{\Δx},\·\·\·,x_n^k)\\ .\\ .\\ .\end{array}$

$(x_1^k,x_2^k,\·\·\·,x_h^k,\·\·\·,x_n^k+\mathrm{\Δx})$

BP neural network is inputted according to these, obtains corresponding output valve (according to needs herein, the output characteristic vector is here only containing one-component), is designated as respectively y ^kfor using the input feature value of sample k as input, the output valve that network is obtained; while remaining unchanged for adding all the other component values of Δ x when h component of the input feature value of sample k, the output valve that network is obtained;

Note for the proper vector of sample k the corresponding input factor of component at the flexible absolute value of sample point k:

$E_h^k=\frac{|(y_h^k-y^k)/y^k|}{|\mathrm{\Δx}/x_h^k|}$ Formula 7

Wherein, h=1,2 ..., n, k=1,2 ... m

Ask ${E_h^{*}}^k=\max \{E_1^k,E_2^k,\·\·\·,E_h^k,\·\·\·E_n^k\};$

With remove respectively ?

Finally, by corresponding each input factor add up, and carry out factor analysis accordingly; Note S _hfor input factor, h is corresponding accumulated value:

$S_h=\underset{k=1}{\overset{m}{\mathrm{\Σ}}}{E}_h^{\′k}$ Formula 8

Note F _hfor the degree of affecting of input factor h on output y,

$F_h=\frac{S_h}{\underset{h=1}{\overset{n}{\mathrm{\Σ}}}{S}_h}$ Formula 9

According to F _hsize judge the size of influence degree of input factor h to output factor; F _hgreatly, its corresponding input factor is just large to the influence degree of output factor, and vice versa; Just can be in the hope of each index factor x by above-mentioned algorithm _ion degree of the impact F of output variable y _h.Obviously, just can be in the hope of each index factor x by above-mentioned algorithm _ion degree of the impact F of output variable y _h.Factor approach based on BP can represent with Fig. 6.

Weight learning method, is easily absorbed in local minimum because BP network exists, restrains the shortcomings such as slow.And genetic algorithm is a kind of global optimization searching method, contribute to search for global optimum's point, can effectively overcome above-mentioned shortcoming., genetic algorithm and neural network are combined for this reason, proposed a kind of combined training method, thereby can improve network convergence speed, avoid network to be absorbed in local minimum.

Coded system

For ensureing e-learning precision, avoid weight step change, adopt real coding herein.The structure of feed-forward type neural network is shown in Fig. 3.Hidden layer transfer function is Sigmoid function.In cataloged procedure, using all weights of neural network and threshold value as chromosomal gene, each genomic constitution chromosome vector V=[v ₁..., v _k..., v _l] formula 10, v _kfor k gene in chromosome.

Fitness function

Adopt square-error to estimate and carry out fitness evaluation, form is as follows:

$f_i=1/E=P/\underset{P}{\mathrm{\Σ}}\underset{k}{\mathrm{\Σ}}{(t_k-y_k)}^2$ Formula 11

In, P is training sample number, p is current learning sample, t _kfor the ideal output of node k, y _kfor the actual output of node k, what adopt here is batch method training sample.

Interlace operation

Interlace operation is the parent chromosome of selecting to participate in intersection by certain crossover probability Pc, adopt and select at random arithmetic intersection or the intersection based on direction, arithmetic intersects can ensure that the offspring who produces is between two parent chromosome, and effectively expanded search space of intersection based on direction:

A) arithmetic intersects:

V ₁'=aV ₁+ (1-α) V ₂formula 12

V ₂'=aV ₂+(1-α)V ₁

B) intersection based on direction:

V ₁'=α (V ₁-V ₂)+V ₁formula 13

V ₂'=α(V ₂-V ₁)+V ₂

V in formula ₁, V ₂for chromosome vector, α is the random number between [0,1].

Mutation operation

The variation of employing self-adaptation.Its objective is self-adaptation adjustment region of search, improve its search capability, improve constringency performance, improve the speed of convergence of genetic algorithm.The design of self-adaptation mutation operator is as follows ^[76]:

V' _m=v _m+ Δ (t, b _k-v _m) or v' _m=v _m-Δ (t, v _m-a _k) formula 14

Wherein, change point v _mthe gene span at place is [a _k, b _k], f (x) is current individual fitness, f _maxit is required problem maximum adaptation value.Conventionally f _maxbe difficult to determine, can replace by the maximum adaptation value in current colony.R is the random number in (0,1) scope, λ ∈ [2,5].

Based on the neural network algorithm of genetic algorithm

As shown in Figure 4, in figure, eg1, eg2 are the target error of setting to its flow process of hybrid algorithm in the method, and sse is neural network error sum of squares.In the hybrid algorithm that method proposes, first utilize the weights of improved self-adapted genetic algorithm to BP network and the distribution of threshold value to carry out the optimum search in global scope, in solution space, orient a good search volume, meet certain condition of convergence.Then utilize weights that genetic algorithm calculates and threshold value as initial weight and the threshold value of BP algorithm, utilize BP algorithm to be good at carrying out the feature of Local Search, in the little space of orienting, carry out optimal solution search.Will converge on quickly like this globally optimal solution.Detailed process is:

A) determine GA and BP algorithm parameter;

B) produce at random the initial value (colony) of N group weights and threshold value, adopt above-mentioned real coding mode to encode to weights and threshold value, and then construct chromosome one by one;

C) error of calculation functional value, thus determine individual fitness function value corresponding to chromosome, error is larger, and fitness value is less;

D) individuality of selection fitness function value maximum, directly entails next generation's (optimized individual reservation);

E) utilize the genetic operation operators such as crossover and mutation to process current generation colony, produce colony of future generation;

F) repeating step 2), 3), 4) and, the distribution of weights and threshold value is constantly evolved, until f _maxtill being less than target error eg1;

G) weights and the Threshold-training neural network of utilizing genetic algorithm to produce, until error sum of squares sse is less than target error eg2.If regulation frequency of training in do not meet the demands, repeating step b)-g), until meet the demands.

Aluminium oxide cost analysis of neural network

1. Cost Analysis Model structure

Cost Analysis Model structure as shown in Figure 5.X in figure ₁, X ₂..., X _nfor N the influence factor of cost Y, as neural network input, cost Y exports as neural network.

2. cost analysis algorithm

First, the data in sample set are carried out to pre-service, data are carried out to nondimensionalization or normalized, make it be transformed to the number between [0,1].

Secondly,, by pretreated sample set BP network, set up the model Y=F (X) of relation between reflection independent variable and dependent variable.In the time of BP network, to only have in the time that the weight space data of network tend towards stability or be constant, training is just calculated and is finished.

Then,, according to the BP network of setting up, obtain the corresponding elasticity of each input factor.The input feature value of remembering the sample k in pretreated sample set is h=1,2 ..., n, k=1,2 ..., m, wherein n is the component number (the index factor number of representative for analyzing) of sample input feature value, m is the contained number of samples of sample set.Input feature value using lower column vector as built BP network:

$(x_1^k,x_2^k,\·\·\·,x_h^k,\·\·\·,x_n^k)$

$(x_1^k+\mathrm{\Δx},x_2^k,\·\·\·,x_h^k,\·\·\·,x_n^k)$

$\begin{array}{c}(x_1^k,x_2^k+\mathrm{\Δx},\·\·\·,x_h^k,\·\·\·x_n^k)\\ .\\ .\\ .\end{array}$ Formula 6

$\begin{array}{c}(x_1^k,x_2^k,\·\·\·,x_h^k+\mathrm{\Δx},\·\·\·,x_n^k)\\ .\\ .\\ .\end{array}$

$(x_1^k,x_2^k,\·\·\·,x_h^k,\·\·\·,x_n^k+\mathrm{\Δx})$

Network is inputted according to these, obtains corresponding output valve (according to needs herein, the output characteristic vector is here only containing one-component), is designated as respectively y ^k, y ^kfor using the input feature value of sample k as input, the output valve that network is obtained. while remaining unchanged for adding all the other component values of Δ x when h component of the input feature value of sample k, the output valve that network is obtained.

Note for the proper vector of sample k the corresponding input factor of component at the flexible absolute value of sample point k:

$E_h^k=\frac{|(y_h^k-y^k)/y^k|}{|\mathrm{\Δx}/x_h^k|}$ Formula 7

Wherein, h=1,2 ..., n, k=1,2 ... m

Ask ${E_h^{*}}^k=\max \{E_1^k,E_2^k,\·\·\·,E_h^k,\·\·\·E_n^k\}.$

With remove respectively ?

Finally, by corresponding each input factor add up, and carry out factor analysis accordingly.Note S _hfor input factor, h is corresponding accumulated value:

$S_h=\underset{k=1}{\overset{m}{\mathrm{\Σ}}}{E}_h^{\′k}$ Formula 8

Note F _hfor the degree of affecting of input factor h on output y,

$F_h=\frac{S_h}{\underset{h=1}{\overset{n}{\mathrm{\Σ}}}{S}_h}$ Formula 9

According to F _hsize judge the size of influence degree of input factor h to output factor.F _hgreatly, its corresponding input factor is just large to the influence degree of output factor, and vice versa.

Obviously, just can be in the hope of each index factor x by above-mentioned algorithm _ion degree of the impact F of output variable y _h.Factor approach based on BP can represent with Fig. 6.

Claims (5)

1. an aluminum oxide production process Dynamic Cost Control method, is characterized in that, comprises the following steps:

Step 1, indicator of costs forecast;

Step 2, Cost Driver analysis.

2. aluminum oxide production process Dynamic Cost Control method described in claim 1, is characterized in that, described step 1 adopts aluminium oxide cost neural network prediction method;

The method comprises: aluminium oxide cost composition model and aluminium oxide cost forecasting model;

Each indicator of costs of described aluminium oxide cost composition model comprises again many cost factors, taking matrix representation aluminium oxide cost composition model as:

formula 1

formula 2

Wherein represent to form i indicator of costs vector of alumina product cost; y _irepresent to form i the indicator of costs of cost of products; y _ijrepresent j cost factor of i indicator of costs vector.

3. aluminum oxide production process Dynamic Cost Control method described in claim 2, is characterized in that, in described formula 2, if ? be equal to zero, think that i the indicator of costs is made up of q cost factor;

Aluminium oxide cost can be expressed as:

formula 3

Each index aluminium oxide cost can be expressed as:

formula 4.

4. aluminum oxide production process Dynamic Cost Control method described in claim 2, it is characterized in that, described aluminium oxide cost forecasting model is based on BP neural net method, if m the aluminium oxide indicator of costs interacts, comprehensive history value forecast model and influence factor forecast model, set up many inputs of many index predictions, many output models, model is as follows:

formula 5

Wherein x _i, be respectively neural network input, output, x _irepresent the historical unit consumption of the interactive indicator of costs, y' _irepresent the prediction unit consumption of the indicator of costs; K represents the number (in chronological order) of historical data that each index is got.

5. aluminum oxide production process Dynamic Cost Control method described in claim 4, is characterized in that, described aluminium oxide cost neural network Cost Analysis Method, X in Cost Analysis Model ₁, X ₂..., X _nfor N the influence factor of cost Y, as neural network input, cost Y exports as BP neural network;

First BP neural network cost analysis algorithm carries out pre-service to the data in sample set, data is carried out to nondimensionalization or normalized, makes it be transformed to the number between [0,1];

Secondly,, by pretreated sample set training BP neural network, set up the model Y=F (X) of relation between reflection independent variable and dependent variable; In the time of training BP neural network, to only have in the time that the weight space data of neural network tend towards stability or be constant, training is just calculated and is finished;

Then,, according to the BP neural network of setting up, obtain the corresponding elasticity number of each input factor; The input feature value that is the sample k in pretreated sample set is h=1,2 ..., n, k=1,2 ..., m, wherein n is the component number (the index factor number of representative for analyzing) of sample input feature value, m is the contained number of samples of sample set; Input feature value using lower column vector as built BP network:

formula 6

BP neural network is inputted according to these, obtains corresponding output valve (according to needs herein, the output characteristic vector is here only containing one-component), is designated as respectively y ^k, y ^kfor using the input feature value of sample k as input, the output valve that network is obtained; while remaining unchanged for adding all the other component values of Δ x when h component of the input feature value of sample k, the output valve that network is obtained;

Note for the proper vector of sample k the corresponding input factor of component at the flexible absolute value of sample point k:

formula 7

Wherein, h=1,2 ..., n, k=1,2 ... m

Ask

With remove respectively ?

Finally, by corresponding each input factor add up, and carry out factor analysis accordingly; Note S _hfor input factor, h is corresponding accumulated value:

formula 8

Note F _hfor the degree of affecting of input factor h on output y,

formula 9

According to F _hsize judge the size of influence degree of input factor h to output factor; F _hgreatly, its corresponding input factor is just large to the influence degree of output factor, and vice versa; Just can be in the hope of each index factor x by above-mentioned algorithm _ion degree of the impact F of output variable y _h.