CN104899425A - Variable selection and forecast method of silicon content in molten iron of blast furnace - Google Patents

Abstract

The invention discloses a variable selection and forecast method of the silicon content in molten iron of a blast furnace. According to the method, blast furnace process parameters of a prediction model of the silicon content in molten iron of a blast furnace are used as input variables; after normalization preprocessing is performed on sample data of the input variables, variable selection is performed on the sample data of the input variables by using a multivariate correlation analysis method and a Spearman rank correlation analysis method, to remove the correlation between production process parameters; and a support vector machine algorithm is used to establish the forecast model of the silicon content in molten iron of the blast furnace, and particle swarm optimization algorithm is introduced to optimize model parameters. The variable selection and forecast method of the silicon content in molten iron of a blast furnace has universal applicability in forecasting the silicon content in molten iron during smelting of a blast furnace, and can achieve desirable forecast accuracy and improve the forecast hit rate of the silicon content in molten iron.

Description

A kind of variables choice forecasting procedure of blast furnace molten iron silicon content

Technical field

The present invention relates to the variables choice forecasting procedure of blast furnace molten iron silicon content.

Background technology

Blast fumance is under sealing condition, carry out complicated chemistry, dynamics, thermodynamics change procedure, is the nonlinear system of a complexity, highly coupling.Rational furnace temperature is kept to be one of stable key factor of blast fumance.In smelting process, Control for Kiln Temperature is in normal range, and blast furnace is with regard to direct motion.If Control for Kiln Temperature generation unusual fluctuations, stroke " overheated " or " excessively cold ", then easily bring out working of a furnace fault.The quality of Control for Kiln Temperature directly affects the fluctuation of the working of a furnace, and working of a furnace state decides the control model of furnace temperature.So blast furnace process production run synthetical automatic control technical difficulty is traced it to its cause be to set up accurately rationally blast furnace temperature control mathematical model, due to process complicacy and measure on difficulty, the general temperature variation indirectly reflected by blast furnace molten iron silicon content (being commonly referred to as chemical heat) in stove, judges blast furnace crucibe Warm status.Blast furnace molten iron silicon content becomes the index that in reflection stove, physical-chemical reaction situation, hot situation and iron quality one is very important, and the amplitude of its change and frequency directly reflect the stability of smelting process.Molten iron silicon content is the important indicator of evaluation conditions of blast furnace stability and iron quality, is also one of mark characterizing heat state of blast furnace and change thereof.In order to effectively control conditions of blast furnace stability, obtaining the situation of change of blast furnace internal thermal status, setting up blast furnace molten iron silicon content forecasting procedure extremely important.

For a long time, the chemical reaction that domestic and international researchist occurs according to blast furnace ironmaking inside and transport phenomenon establish multiple mechanism mathematical model, these models play certain positive role for announcement blast furnace internal phenomena and reaction blast furnace ironmaking mechanism in theory, but it is low and calculate the shortcomings such as consuming time also to there is accuracy.Given this, data driven technique is studied widely, to attempt to realize the real time modelling to blast furnace ironmaking process and control, data-driven modeling technique is the main criteria using the feature of data of description as modeling, be under the conviction of self speaking in data, the correlationship between analysis of system variables.At present, the model of the blast furnace molten iron silicon content prediction utilizing the thought of data-driven to set up mainly contains autoregressive model, neural network model, Nonlinear Time Series Analysis model, fuzzy model, Bayesian network model, partial least square model, supporting vector machine model etc., these models are when predicting blast furnace molten iron silicon content, mostly all make use of whole correlated variabless that blast furnace data acquisition arrives as independent variable, but due to the on-the-spot environment of very noisy of blast furnace and the strong correlation of blast furnace data itself, although quote whole variate-values can make full use of the abundant data characteristics of blast furnace as independent variable, also brought very large noise into simultaneously, very large obstruction is provided to the forecasting accuracy of silicone content.

Summary of the invention

In order to overcome the deficiencies in the prior art, the invention provides the variables choice forecasting procedure of a kind of blast furnace molten iron silicon content forecast.

Scheme is as follows:

A variables choice forecasting procedure for blast furnace molten iron silicon content forecast, comprises the following steps:

1) pre-service is carried out, to the normalization of sample data to the correlated variables of the blast furnace molten iron silicon content forecasting model obtained;

2) use the method for multivariate correlation analytical approach and Spearman rank correlation analysis to the variables choice of sample data;

3) determine that the mode input variable that blast furnace molten iron silicon content forecasts, input variable comprise CO, CO in furnace top pressure, top temperature, material speed, stock gas ₂with the silicone content of a upper stove;

4) algorithm of support vector machine is used to set up the model of blast furnace molten iron silicon content forecast;

5) model parameter of particle cluster algorithm to blast furnace silicon content prediction is adopted to be optimized.

The described correlated variables for the blast furnace molten iron silicon content forecasting model obtained carries out pre-service, is: x to the method for normalizing of sample data _i ^*=(x _i-min x _i)/(max x _i-min x _i), wherein x _iand x _i ^*represent the value before and after normalization respectively, max x _iwith min x _irepresent the minimum and maximum value in sample data respectively.

Described for use multivariate correlation analytical approach and Spearman rank correlation analytical approach to the variables choice of sample data, the method for wherein multivariate correlation analysis is:

$r(x,y)=\frac{\mathrm{cov}(x,y)}{{\mathrm{\σ}}_x{\mathrm{\σ}}_y}$

Wherein x and y represents any Two Variables respectively, and cov (x, y) represents the covariance matrix of variable x and y.As x=y, r (x, y)=1 represents to have complete linear dependence, and r (x, y)=0 represents to have nonfull dependency, and r (x, y)=-1 represents to have complete negative linear correlation.The method of multivariate correlation coefficient analysis is used to can be good at weighing the linear dependence between blast furnace variable.Spearman rank correlation analytical approach is:

$\mathrm{\ρ}(x,y)=1-\frac{6\underset{i=1}{\overset{N}{\mathrm{\Σ}}}d{(x_i,y_i)}^2}{N(N^2-1)}$

Wherein N is the number of variable sample, d (x _i, y _i) represent the difference of the grade of Two Variables every pair of sample.

Described for variables choice and forecasting procedure, in conjunction with multivariate correlation analytical approach and Spearman rank correlation property coefficient analytical approach, select multiple variable average correlation coefficient to be less than the variable of 0.15 as input variable.

Described use algorithm of support vector machine sets up the model of blast furnace molten iron silicon content forecast: note mode input variable data sample is (x _i, y _i), i=1,2 ... N, x _i∈ R ⁿ, wherein x _ifor input variable, y _ibe corresponding output variable, N is the number of variable sample. be Nonlinear Mapping input variable x being mapped to high-dimensional feature space, carry out at feature space the non-linear regression that linear regression just corresponds to the low-dimensional input space.A linear regression function is built in feature space here ω is the vector of feature space.If all training datas by linear function fit under precision ε, can be considered the situation allowing matching, introduce relaxation factor ξ, ξ ^*>=0, according to structural risk minimization, we can obtain:

$\min \frac{1}{2}{\left|\right|\mathrm{\ω}\left|\right|}^2+C(\mathrm{\ν\ϵ}+\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}({\mathrm{\ξ}}_i+{\mathrm{\ξ}}_i^{*}))$

ε≥0，ξ _i,ξ _i ^*≥0，i＝1,2,…N

Wherein, parameter ν ∈ [0,1] is used for controlling the number of support vector or the number of error sample point, and C >=0 represents penalty factor, is used for punishing the point exceeding fit Plane.The dual problem that we can obtain above-mentioned optimization problem is:

$s.t.\underset{i=1}{\overset{N}{\mathrm{\Σ}}}({\mathrm{\α}}_i-{\mathrm{\α}}_i^{*})=0$

$0\≤{\mathrm{\α}}_i\≤C/N,0\≤{\mathrm{\α}}_i^{*}\≤C/N,i=\mathrm{1,2},...N$

Wherein α _iwith represent Lagrange multiplier.In the process addressed this problem, the training sample that the Lagrange multiplier of non-zero is corresponding is called support vector, and the super dawn of matching has support vector to determine completely:

$f\left(x\right)=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}({\mathrm{\α}}_i^{*}-{\mathrm{\α}}_i)k(x_i,x)+b$

Wherein be defined as kernel function, the function meeting not this condition can be called as kernel function, here, uses gaussian kernel function, is defined as follows:

$k(x_i,x)=\exp (-\frac{{\left|\right|x_i-x\left|\right|}^2}{2{\mathrm{\σ}}^2})$

The described model parameter of employing particle cluster algorithm to support vector machine is optimized: the parameter ν in model, parameter σ in penalty factor and kernel function has a great impact the precision of prediction of model, adopts the particle cluster algorithm with whole search capability to select this three parameters.With the root mean square index of the checking sample set of regression model for fitness function, wherein y _ifor the actual value of checking collection sample, the predicted value that the model that checking sample evidence trains obtains, N is the number of checking sample.

Beneficial effect of the present invention:

The present invention with the blast furnace technology parameter of blast furnace molten iron silicon content forecasting model for input variable, after pre-service is normalized to the sample data of input variable, multivariate correlation analysis and Spearman rank correlation property coefficient analytical approach is adopted to carry out correlation analysis to input variable, eliminate the correlativity between processing parameter, according to the result determination input variable analyzed, use algorithm of support vector machine to set up the model of blast furnace molten iron silicon content forecast, introduce particle cluster algorithm with Optimized model parameter.To the molten iron silicon content forecast of blast furnace ironmaking process, there is general versatility, good forecast precision can be obtained, improve the forecast hit rate of blast furnace molten iron silicon content.

Accompanying drawing explanation

Fig. 1 is the comparison diagram of molten iron silicon content predicted value and actual value;

Fig. 2 is molten iron silicon content prediction error figure.

Embodiment

The present invention is directed to the nonlinearity of blast furnace ironmaking process, between each blast fumance parameter, the feature such as strong coupling, is applied to blast furnace ironmaking production run, sets up the blast furnace molten iron silicon content forecasting model based on support vector machine by algorithm of support vector machine.The molten iron silicon content forecast of the present invention to blast furnace ironmaking process has general versatility, improves blast furnace molten iron silicon content accuracy of the forecast and hit rate, provides technical guarantee to the steady reliability service of blast furnace.The method of multivariate correlation analysis is only the correlativity weighed from linear case between data variable, and blast furnace data not only have Linearity, also have very strong non-linear, so only use multivariate correlation analysis mode to be very unilateral, and the method for Spearman rank correlation analysis is the method according to correlationship between ranked data research Two Variables, it carries out calculating according to two differences arranging into each reciprocity progression of In Grade, do not have Coefficient of production-moment correlation strict to the requirement of data qualification, transform by continuous variable observational data the ranked data obtained, no matter the population distribution form of Two Variables, the size of sample size how, can study with Spearman rank correlation.So be applicable to very much the analysis of blast furnace data, simultaneously in conjunction with the method that multivariate correlation is analyzed, from linear angles, also accurately can not only can weigh the correlativity of data itself from non-linear angle, provide guarantee to the accurately predicting of model.

The variables choice forecasting procedure of blast furnace molten iron silicon content comprises the following steps:

1) pre-service is carried out, to the normalization of sample data to the correlated variables of the blast furnace molten iron silicon content forecasting model obtained;

2) use multivariate correlation analytical approach and Spearman rank correlation analytical approach to the variables choice of sample data;

3) determine that the mode input variable that blast furnace molten iron silicon content forecasts, input variable comprise CO, CO in furnace top pressure, top temperature, material speed, stock gas ₂with the silicone content of a upper stove;

4) algorithm of support vector machine is used to set up the model of blast furnace molten iron silicon content forecast;

5) model parameter of particle cluster algorithm to blast furnace silicon content prediction is adopted to be optimized.

The described correlated variables for the blast furnace molten iron silicon content forecasting model obtained carries out pre-service, is: x to the method for normalizing of sample data _i ^*=(x _i-min x _i)/(max x _i-min x _i), wherein x _iand x _i ^*represent the value before and after normalization respectively, max x _iwith min x _irepresent the minimum and maximum value in sample data respectively.

Described for use multivariate correlation analytical approach and Spearman rank correlation analytical approach to the variables choice of sample data, the method for wherein multivariate correlation analysis is:

$r(x,y)=\frac{\mathrm{cov}(x,y)}{{\mathrm{\σ}}_x{\mathrm{\σ}}_y}$

Wherein x and y represents any Two Variables respectively, and cov (x, y) represents the covariance matrix of variable x and y.As x=y, r (x, y)=1 represents to have complete linear dependence, and r (x, y)=0 represents to have nonfull dependency, and r (x, y)=-1 represents to have complete negative linear correlation.The method of multivariate correlation coefficient analysis is used to can be good at weighing the linear dependence between blast furnace variable.Spearman rank correlation analytical approach is:

$\mathrm{\ρ}(x,y)=1-\frac{6\underset{i=1}{\overset{N}{\mathrm{\Σ}}}d{(x_i,y_i)}^2}{N(N^2-1)}$

Wherein N is the number of variable sample, d (x _i, y _i) represent the difference of the grade of Two Variables every pair of sample.

The described methods combining multivariate correlation analytical approach for variables choice and Spearman rank correlation analytical approach, select multiple variable average correlation coefficient to be less than the variable of 0.15 as input variable.

The model that described use algorithm of support vector machine sets up blast furnace molten iron silicon content forecast is:

Note mode input variable data sample is (x _i, y _i), i=1,2 ... N, x _i∈ R ⁿ, wherein x _ifor input variable, y _ibe corresponding output variable, N is the number of variable sample. be Nonlinear Mapping input variable x being mapped to high-dimensional feature space, carry out at feature space the non-linear regression that linear regression just corresponds to the low-dimensional input space.A linear regression function is built in feature space here ω is the vector of feature space.If all training datas by linear function fit under precision ε, can be considered the situation allowing matching, introduce relaxation factor ξ, ξ ^*>=0, according to structural risk minimization, we can obtain:

$\min \frac{1}{2}{\left|\right|\mathrm{\ω}\left|\right|}^2+C(\mathrm{\ν\ϵ}+\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}({\mathrm{\ξ}}_i+{\mathrm{\ξ}}_i^{*}))$

ε≥0，ξ _i，ξ _i ^*≥0，i＝1，2，…N

Wherein, parameter ν ∈ [0,1] is used for controlling the number of support vector or the number of error sample point, and C >=0 represents penalty factor, is used for punishing the point exceeding fit Plane.The dual problem that we can obtain above-mentioned optimization problem is:

$s.t.\underset{i=1}{\overset{N}{\mathrm{\Σ}}}({\mathrm{\α}}_i-{\mathrm{\α}}_i^{*})=0$

$0\≤{\mathrm{\α}}_i\≤C/N,0\≤{\mathrm{\α}}_i^{*}\≤C/N,i=\mathrm{1,2},...N$

Wherein α _iwith represent Lagrange multiplier.In the process addressed this problem, the training sample that the Lagrange multiplier of non-zero is corresponding is called support vector, and the super dawn of matching has support vector to determine completely:

$f\left(x\right)=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}({\mathrm{\α}}_i^{*}-{\mathrm{\α}}_i)k(x_i,x)+b$

Wherein be defined as kernel function, the function meeting not this condition can be called as kernel function, and here, we use gaussian kernel function, are defined as follows:

$k(x_i,x)=\exp (-\frac{{\left|\right|x_i-x\left|\right|}^2}{2{\mathrm{\σ}}^2})$

The described model parameter of employing particle cluster algorithm to support vector machine is optimized:

Parameter ν in model, the parameter σ in penalty factor and kernel function has a great impact the precision of prediction of model, adopts the particle cluster algorithm with whole search capability to select this three parameters.With the root mean square index of the checking sample set of regression model for fitness function, wherein y _ifor the actual value of checking collection sample, the predicted value that the model that checking sample evidence trains obtains, N is the number of checking sample.

Embodiment

For the validity of checking institute of the present invention extracting method, adopt certain steel mill 2500m ³the actual production data of blast furnace carry out the experiment of molten iron silicon content forecast application.Choose CO, CO of comprising in furnace top pressure, top temperature, material speed, stock gas ₂with the silicone content of the upper stove input variable as Silicon Content Prediction in Process of Iron model.The sampled data of all variablees used in model training and model predictive process, all adopts the measurement mean value of coming out of the stove in units of heat using molten iron as sampling and forecast cycle.

Acquire the sample data of 1000 stoves in experiment altogether, wherein continuous 900 stoves are as training sample during modeling, and other 100 stoves more late on the time are as test sample book.

Next in conjunction with this detailed process, implementation step of the present invention is explained in detail:

1) pre-service is carried out, to the normalization of sample data to the correlated variables of the blast furnace molten iron silicon content forecasting model obtained;

2) use multivariate correlation analytical approach and Spearman rank correlation analytical approach to the variables choice of sample data;

3) determine that the mode input variable that blast furnace molten iron silicon content forecasts, input variable comprise CO, CO in furnace top pressure, top temperature, material speed, stock gas ₂with the silicone content of a upper stove;

4) algorithm of support vector machine is used to set up the model of blast furnace molten iron silicon content forecast;

5) model parameter of particle cluster algorithm to blast furnace silicon content prediction is adopted to be optimized.

The described correlated variables for the blast furnace molten iron silicon content forecasting model obtained carries out pre-service, is: x to the method for normalizing of sample data _i ^*=(x _i-min x _i)/(max x _i-min x _i), wherein x _iand x _i ^*represent the value before and after normalization respectively, max x _iwith min x _irepresent the minimum and maximum value in sample data respectively.

Described for use multivariate correlation analytical approach and Spearman rank correlation analytical approach to the variables choice of sample data, the method for wherein multivariate correlation analysis is:

$r(x,y)=\frac{\mathrm{cov}(x,y)}{{\mathrm{\σ}}_x{\mathrm{\σ}}_y}$

Wherein x and y represents any Two Variables respectively, and cov (x, y) represents the covariance matrix of variable x and y.As x=y, r (x, y)=1 represents to have complete linear dependence, and r (x, y)=0 represents to have nonfull dependency, and r (x, y)=-1 represents to have complete negative linear correlation.The method of multivariate correlation analysis is used to can be good at weighing the linear dependence between blast furnace variable.Spearman rank correlation property coefficient method is:

$\mathrm{\ρ}(x,y)=1-\frac{6\underset{i=1}{\overset{N}{\mathrm{\Σ}}}d{(x_i,y_i)}^2}{N(N^2-1)}$

Wherein N is the number of variable sample, d (x _i, y _i) represent the difference of the grade of Two Variables every pair of sample.

The described methods combining multivariate correlation analytical approach for variables choice and Spearman rank correlation analytical approach, select multiple variable average correlation coefficient to be less than the variable of 0.15 as input variable.

The model that described use algorithm of support vector machine sets up blast furnace molten iron silicon content forecast is:

Note mode input variable data sample is (x _i, y _i), i=1,2 ... N, x _i∈ R ⁿ, wherein x _ifor input variable, y _ibe corresponding output variable, N is the number of variable sample. be Nonlinear Mapping input variable x being mapped to high-dimensional feature space, carry out at feature space the non-linear regression that linear regression just corresponds to the low-dimensional input space.A linear regression function is built in feature space here ω is the vector of feature space.If all training datas by linear function fit under precision ε, can be considered the situation allowing matching, introduce relaxation factor ξ, ξ ^*>=0, according to structural risk minimization, we can obtain:

$\min \frac{1}{2}{\left|\right|\mathrm{\ω}\left|\right|}^2+C(\mathrm{\ν\ϵ}+\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}({\mathrm{\ξ}}_i+{\mathrm{\ξ}}_i^{*}))$

ε≥0，ξ _i，ξ _i ^*≥0，i＝1，2，…N

Wherein, parameter ν ∈ [0,1] is used for controlling the number of support vector or the number of error sample point, and C >=0 represents penalty factor, is used for punishing the point exceeding fit Plane.The dual problem that we can obtain above-mentioned optimization problem is:

$s.t.\underset{i=1}{\overset{N}{\mathrm{\Σ}}}({\mathrm{\α}}_i-{\mathrm{\α}}_i^{*})=0$

$0\≤{\mathrm{\α}}_i\≤C/N,0\≤{\mathrm{\α}}_i^{*}\≤C/N,i=\mathrm{1,2},...N$

Wherein α _iwith represent Lagrange multiplier.In the process addressed this problem, the training sample that the Lagrange multiplier of non-zero is corresponding is called support vector, and the super dawn of matching has support vector to determine completely:

$f\left(x\right)=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}({\mathrm{\α}}_i^{*}-{\mathrm{\α}}_i)k(x_i,x)+b$

Wherein be defined as kernel function, the function meeting not this condition can be called as kernel function, and here, we use gaussian kernel function, are defined as follows:

$k(x_i,x)=\exp (-\frac{{\left|\right|x_i-x\left|\right|}^2}{2{\mathrm{\σ}}^2})$

The described model parameter of employing particle cluster algorithm to support vector machine is optimized:

Parameter ν in model, the parameter σ in penalty factor and kernel function has a great impact the precision of prediction of model, adopts the particle cluster algorithm with whole search capability to select this three parameters.With the root mean square index of the checking sample set of regression model for fitness function, wherein y _ifor the actual value of checking collection sample, the predicted value that the model that checking sample evidence trains obtains, N is the number of checking sample.

μ=0.44 is being obtained, C=4.90, σ=0.6 with PSO Optimal Parameters.The production data new with continuous 100 stoves carries out the test case of molten iron silicon content forecast as shown in Figure 1 to the model set up.Fig. 1 gives the tracking effect of molten iron silicon content predicted value to true laboratory values, as can be seen from the prediction error curve of Fig. 2, the ratio that the error of the silicon content prediction of 100 stove molten iron is less than 0.1 is 86%, and higher than other Variable Selections, and its variation tendency has well approached truth.

Claims (6)

1. a variables choice forecasting procedure for blast furnace molten iron silicon content, is characterized in that, comprise the steps:

1) pre-service is carried out, to the normalization of sample data to the correlated variables of the blast furnace molten iron silicon content forecasting model obtained;

2) method of multivariate correlation analysis and Spearman rank correlation analysis is used to carry out correlation analysis to sample data variable;

3) determine that the mode input variable that blast furnace molten iron silicon content forecasts, input variable comprise CO, CO in furnace top pressure, top temperature, material speed, stock gas ₂with the silicone content of a upper stove;

4) algorithm of support vector machine is used to set up the model of blast furnace molten iron silicon content forecast;

5) model parameter of particle cluster algorithm to blast furnace silicon content prediction is adopted to be optimized.

2. method according to claim 1, is characterized in that, step 1) described in pre-service is carried out for the correlated variables of blast furnace molten iron silicon content forecasting model obtained, to the method for normalizing of sample data be: x _i ^*=(x _i-minx _i)/(maxx _i-minx _i), wherein x _iand x _i ^*represent the value before and after normalization respectively, maxx _iand minx _irepresent the minimum and maximum value in sample data respectively.

3. method according to claim 1, it is characterized in that, step 2) described in for use multivariate correlation analytical approach and Spearman rank correlation analytical approach to the variables choice of sample data, the method for wherein multivariate correlation analysis is: wherein x and y represents any Two Variables respectively, cov (x, y) covariance matrix of variable x and y is represented, as x=y, r (x, y)=1 represents to have complete linear dependence, r (x, y)=0 represents to have nonfull dependency, and r (x, y)=-1 represents to have complete negative linear correlation; Spearman rank correlation analytical approach is:

$\mathrm{\ρ}(x,y)=1-\frac{6\underset{i=1}{\overset{N}{\mathrm{\Σ}}}d{(x_i,y_i)}^2}{N(N^2-1)}$

Wherein N is the number of variable sample, d (x _i, y _i) represent the difference of the grade of Two Variables every pair of sample.

4. method according to claim 1, is characterized in that, step 3) described in the Input variable selection average correlation coefficient variable that is less than 0.15.

5. method according to claim 1, is characterized in that, step 4) model that uses algorithm of support vector machine to set up blast furnace molten iron silicon content forecast is: note mode input variable data sample is (x _i, y _i), i=1,2 ... N, x _i∈ R ⁿ, wherein x _ifor input variable, y _ibe corresponding output variable, N is the number of variable sample. be Nonlinear Mapping input variable x being mapped to high-dimensional feature space, carry out at feature space the non-linear regression that linear regression just corresponds to the low-dimensional input space, in feature space, build a linear regression function here ω is the vector of feature space, if all training datas by linear function fit under precision ε, can be considered the situation allowing matching, introduce relaxation factor ξ, ξ ^*>=0, according to structural risk minimization, obtain optimization problem:

$\min \frac{1}{2}{\left|\right|\mathrm{\ω}\left|\right|}^2+C(\mathrm{v\ϵ}+\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}({\mathrm{\ξ}}_i+{\mathrm{\ξ}}_i^{*}))$

$\mathrm{\ϵ}\≥0,{\mathrm{\ξ}}_1,{\mathrm{\ξ}}_i^{*}\≥0,i=\mathrm{1,2},...N$

Wherein, parameter be used for controlling the number of support vector or the number of error sample point, C>=0 represents penalty factor, is used for punishing the point exceeding fit Plane; The dual problem of above-mentioned optimization problem is:

$\begin{array}{cc}s.t.& \underset{i=1}{\overset{N}{\mathrm{\Σ}}}({\mathrm{\α}}_i-{\mathrm{\α}}_i^{*})=0\end{array}$

$0\≤{\mathrm{\α}}_i\≤C/N,0\≤{\mathrm{\α}}_i^{*}\≤C/N,i=\mathrm{1,2},...N$

Wherein α _iwith represent Lagrange multiplier, in the process addressed this problem, the training sample that the Lagrange multiplier of non-zero is corresponding is called support vector, and the lineoid of matching is determined by support vector completely:

$f\left(x\right)=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}({\mathrm{\α}}_i^{*}-{\mathrm{\α}}_i)k(x_i,x)+b$

Wherein be defined as kernel function, described kernel function is gaussian kernel function, is defined as follows:

$k(x_i,x)=\exp (-\frac{{\left|\right|x_i-x\left|\right|}^2}{{2\mathrm{\σ}}^2}).$

6. method according to claim 5, it is characterized in that, step 5) described in the model parameter of employing particle cluster algorithm to support vector machine be optimized: the parameter ν in model, parameter σ in penalty factor and kernel function adopts the particle cluster algorithm with whole search capability to select, with the root mean square index of the checking sample set of regression model for fitness function, wherein y _ifor the actual value of checking collection sample, the predicted value that the model that checking sample evidence trains obtains, N is the number of checking sample.