CN102890144B - Method for predicting coke quality through nonlinear optimization coal blending based on coal rock vitrinite total reflectance - Google Patents

Abstract

The invention discloses a method for predicting coke quality through nonlinear optimization coal blending based on coal rock vitrinite total reflectance, thereby providing important technical assurance for the improvement of the stability and quality of coke produced by a coke making enterprise. The method provided by the invention comprises the following steps of: establishing a coking coal resource information database, and inputting the caking property index of coking blended coal and the coal rock index of single coking coal into the coking coal resource information database; and establishing a coke quality prediction model through a support vector machine, and then predicting the quality index of the coke according to the coal-quality caking property index level of the coking blended coal, wherein the coal-quality caking property index level comprises two factors, namely the maximum thickness value Y of a gelatinous layer and a caking index value G, and comprises coal rock vitrinite total-component reflectance and the liver-inert ratio of macerals. The method provided by the invention is capable of characterizing the maximum thickness value Y of the gelatinous layer indicating the quantity of metaplast in the softening process of the coking coal and indicating the caking property quality of the metaplast, thereby realizing the prediction process with the goal of predicting the mechanical strength and thermal state performance of the coke.

Description

The method of coal petrography vitrinite total reflectivity nonlinear optimization coal blending prediction coke quality

Technical field

The present invention relates to the Forecasting Methodology of coke quality in technical field of coal chemical industry coking industry coking production process, be specifically related to coking and produce mixed coal cohesiveness index level used, comprise maximum thickness of plastic layer Y value and two factors of caking index G value; Coal petrography index level, comprises two factors of vitrinite's total reflectivity and micropetrological unit lazy ratio alive, and two levels four of formation are because usually predicting the utility system of coke strength and hot performance.The method of especially a kind of coal petrography vitrinite total reflectivity nonlinear optimization coal blending prediction coke quality.

Background technology

The raising of blast furnace maximization and Oygen And Coal Blowing Technology, to burnt quality and the stable requirements at the higher level that proposed for blast furnace.In, the unsettled situation of ature of coal various at current coking coal type change, how to pass through physical strength and the hot performance of coking coal ature of coal property prediction coke, to stabilizing and increasing coke quality, and reduce coal chemical enterprise and produce coal blending cost etc. and all there is very important practical significance, this also at present iron and steel metallurgy enterprise coke-oven plant need an actual production problem of solution badly.

At present, adopt traditional experience coal blending or adopt separately Blending of Coal Petrography method to predict that coke quality index is mostly to plant coal coal index and corresponding blending ratio by list, and utilize method of weighting to predict coke strength (M ₄₀, M ₁₀) and hot performance (CRI, CSR), or adopt the conventional cohesiveness index Y value of a kind of ature of coal or G value again in conjunction with degree of coal metamorphism index volatile matter V _dafor the single index of the average maximum reflectivity of coal vitrinite is predicted coke strength and hot performance.These methods can estimate coke strenth and thermal behavior to a certain extent.But for single coal ature of coal of planting while having mixture phenomenon, when coke quality index prediction that the method is carried out, can produce larger error with actual coke quality index.Chinese patent CN1749358A discloses a kind of employing multiple linear regression analysis method prediction coke quality and has referred to calibration method, and the method feature is simple, but one is because adopting linear regression relation, being difficult to obtain determinacy numerical value when the method prediction; It two is ature of coal complexity due to coking coal, the coking coal volatile matter V of similarity condition _dafwith G value, in the time of prediction coke strength and hot performance, can produce very big-difference, this is because single coal of planting, by after mixture, can adopt the coal of varying number and kind to obtain identical volatile matter V after mixture _dafwith G value.Meanwhile, because coking coal and coke quality index are complicated relations in process of coking, adopt linear relationship to substitute the nonlinear relationship in coking coal coking process, must make the error producing with reality that predicts the outcome of coke quality index increase.Chinese patent CN101661026A adopts BP neural network to predict coke quality, in fact the method is to adopt empirical risk minimization principle, but actual result can not make expected risk minimize, and has in theory defect, easily be absorbed in local minimum point, the shortcoming such as generalization ability is not strong.In addition, determining of the hidden layer number of neural network and the number of hidden nodes generally comments experience to determine, the also clear and definite algorithm of neither one, and this,, with regard to making the precision of prediction of neural network be subject to certain impact, is subject to certain restrictions in actual applications.Simultaneously, Chinese patent CN101661026A adopts coal petrography vitrinite reflectance to be divided into 6 sections as input quantity, both vitrinite reflectance has been divided into and has been less than 0.60%, 0.60%～0.65%, 0.65%～1.25%, 1.25%～1.75%, 1.75%～1.85%, has been greater than 1.85% totally 6 reflectivity distributions section.The input parameter as prediction of coke quality at each section of employing aggregate-value has been considered in this segmentation.Meanwhile, distribute section containing having covered bottle coal, gas-fat coal, rich coal and part coking coal class ature of coal component at 0.65%～1.25% reflectivity, contained part coking coal, thin coking coal, part lean coal class ature of coal component in 1.25%～1.75% reflectivity distribution section.And in the situation that at present coking coal market mixed coal phenomenon is serious, need to be to the classification of the prediction coal petrography method division that becomes more meticulous, to improve the concentration degree of coking coal prediction index.Therefore two level four factors that, the present invention proposes will play larger facilitation to coking coal prediction coke quality.

Summary of the invention

The object of this invention is to provide the method for a kind of coal petrography vitrinite total reflectivity nonlinear optimization coal blending prediction coke quality, coal chemical enterprise is produced to the stability of coke and the raising of coke quality the guarantee of important technology can be provided.

The method step of coal petrography provided by the invention vitrinite total reflectivity nonlinear optimization coal blending prediction coke quality is as follows:

One. set up coal for coking resource information database

By the cohesiveness index of coking mixed coal, comprise that maximum thickness of plastic layer Y value, caking index G value are entered into coal resource information database; By single coal petrography index of planting coking coal, comprise the lazy parameter that compares of work that the meticulous paragraph data of coal vitrinite total reflectivity, coal petrography maceral composition calculate, in the mixed coal ature of coal input information coal resource information database that formed thus, set up coal for coking resource information database;

Two. set up Coke Quality Prediction Models

According to the ature of coal cohesiveness index level of coking mixed coal, comprise maximum thickness of plastic layer Y value and two factors of caking index G value; Coal petrography index level, comprises the lazy quality index of recently predicting coke of work of vitrinite's full constituent reflectivity and micropetrological unit, and the prediction main body of mixed coal is the shatter strength M of coke ₄₀, scuff resistance M ₁₀and reactive CRI and post-reaction strength CSR;

(1) determine coking mixed coal quality index

According to the mixed coal cohesiveness index forming after coal preparation technology comminutor of Experiment Coke Oven or the actual use of coke-oven plant's commercial coke oven or Experiment Coke Oven pulverize and coordinate after coal cohesiveness index, comprise maximum thickness of plastic layer Y value and two factors of caking index G value; Coal petrography index, comprises that the lazy quality of recently predicting coke of work of vitrinite's full constituent reflectivity and micropetrological unit, this prediction index system are to adopt support vector machine technology.Carry out in Computing process in the actual data that detect of employing, above-mentioned two cohesiveness factors according to coal characteristic to coking coal and two coal petrography index factors adopt support vector machine technology to predict, in coke making and coal blending cohesiveness index maximum thickness of plastic layer Y value and caking index G value and coal petrography index prediction coke quality index, in maceral, active component comprises vitrinite, half vitrinite and stable group, inert constituent comprises inertinite, half vitrinite and mineral, wherein:

Active component=vitrinite+1/3 × half vitrinite+stablize group

Inert constituent=inertinite+2/3 × half vitrinite+mineral

Lazy ratio=active component/inert constituent alive

Undertaken by following provisions:

(1) the maximum thickness of colloidal matter layer Y value of mixed coal and caking index G value are as the criterion with actual analysis value

(2) the lazy computation model that compares of the work of coking coal vitrinite total reflectivity and micropetrological unit

D _jbe in model mixed coal vitrinite in the frequency of j point reflection rate:

$D_j=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}(\mathrm{DSingl}{e}_j\×P_i)$

Wherein, DSingle _j(j=1,2,3 ... 60) frequency distributing in j point reflection rate for each the Dan Zhong coal vitrinite gathering is continuity reflectivity Distribution Value; P _iit is the proportioning of i kind list kind coal.

The lazy ratio of work of mixed coal is made as B:

$B=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}(A_i/I_i\×P_i)$

Wherein, A _ibe the active component of i kind coal, I _ibe the inert constituent of i kind coal, P _ibe i kind coal shared number percent in Coal Blending Schemes.N is the single total number of planting coal in Coal Blending Schemes.

(2) prediction coke quality index

Compare according to the check analysis data of Experiment Coke Oven coking tests or commercial coke oven coke and actual coal detection analysis, Mechanical Strength of Coke and hot performance index are predicted, and adopt support vector machine technology, support vector machine is the structural risk minimization based on Statistical Learning Theory, utilize maximum interphase sorter thought and the method based on core to combine, can solve preferably the practical problemss such as small sample, non-linear, high dimension drawn game portion minimal point, effectively avoid " cross and fit ", following sample is had to good generalization ability.

$L^{\mathrm{\ϵ}}(x,y,f)={|y-f\left(x\right)|}_{\mathrm{\ϵ}}=\left\{\begin{array}{cc}0& |y-f\left(x\right)|\≤\mathrm{\ϵ}\\ |y-f\left(x\right)|-\mathrm{\ϵ}& \mathrm{else}\end{array}\right.---\left(1\right)$

Wherein f is the real-valued function on the X of territory.Its meaning is if the difference between predicted value and actual value is while being less than ε, loss equals 0, otherwise be that the absolute value of predicted value and actual value difference is poor with it, finds a kind of estimation regression function in linear function set, f (x)=(wx)+bw, x ∈ R ⁿ, b ∈ R, wherein, (x ₁, y ₁) ... (x _m, y _m) be independent identically distributed data, b is amount of bias.Regression estimation problem is to ask parameter w and b, makes, for the input x beyond sample, to meet | f (x)-(wx)-b|≤ε, asks parameter w and b to be equivalent to the minimum value of asking following formula:

$\min \mathrm{\Φ}\left(w\right)=\frac{1}{2}{\left|\right|w\left|\right|}^2=\frac{1}{2}(w\·w)---\left(2\right)$

$\begin{array}{cc}\mathrm{subject}& \mathrm{to}\left\{\begin{array}{c}{y}_i-((w\·x_i)+b)\≤\mathrm{\ϵ}\\ ((w\·x_i)+b)-y_i\≤\mathrm{\ϵ}\end{array}\right.\end{array},i=\mathrm{1,2},...,m$

For guaranteeing that above-mentioned optimization problem has solution, introduce slack variable ξ, optimization problem is converted into the minimum problems solving under following formula constraint:

$\min \mathrm{\Φ}\left(w\right)=\frac{1}{2}{\left|\right|w\left|\right|}^2+C\underset{i}{\overset{m}{\mathrm{\Σ}}}({\mathrm{\ξ}}_i+{\widehat{\mathrm{\ξ}}}_i)---\left(3\right)$

$\begin{array}{cc}\mathrm{subject}& \mathrm{to}\left\{\begin{array}{c}{y}_i-((w\·x_i)+b)\≤\mathrm{\ϵ}+{\mathrm{\ξ}}_i\\ ((w\·x_i)+b)-y_i\≤\mathrm{\ϵ}+{\widehat{\mathrm{\ξ}}}_i\\ {\mathrm{\ξ}}_i,{\widehat{\mathrm{\ξ}}}_i\≥0,i=\mathrm{1,2},...,l\end{array}\right.\end{array}$

Wherein C is the constant of specifying, and C>0 is used for the smoothness of representative function f and permissible error and is greater than the compromise between the numerical value of ε, is mainly improving generalization ability and is reducing to rise between error regulating and controlling effect.ε is a positive number, need to set in advance, is mainly to wish for control algolithm the precision reaching;

Ask optimum solution by former problem is converted into dual problem, set up Lagrange function according to objective function and constraint condition, the dual form of optimization problem is:

$\begin{array}{c}L(w,b,{\mathrm{\ξ}}_i,{\widehat{\mathrm{\ξ}}}_i)=\frac{1}{2}{\left|\right|w\left|\right|}^2+C\underset{i}{\overset{m}{\mathrm{\Σ}}}({\mathrm{\ξ}}_i+{\widehat{\mathrm{\ξ}}}_i)-\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\α}}_i[\mathrm{\ϵ}+{\mathrm{\ξ}}_i-y_i+((w\·x_i)+b)]-\\ \underset{i=1}{\overset{m}{\mathrm{\Σ}}}{\widehat{\mathrm{\α}}}_i[\mathrm{\ϵ}+{\widehat{\mathrm{\ξ}}}_i+y_i-((w\·x_i)+b]-\underset{i=1}{\overset{m}{\mathrm{\Σ}}}(n_i{\mathrm{\ξ}}_i+{\widehat{\mathrm{\η}}}_i{\widehat{\mathrm{\ξ}}}_i)\end{array}---\left(4\right)$

Wherein, w, b, ξ _i, for former variable; α _i, η _i, for dual variable, and meet α _i, η _i, variable in above formula is asked to local derviation, can obtain

$\left\{\begin{array}{c}\frac{\∂L}{\∂b}=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}({\widehat{\mathrm{\α}}}_i-{\mathrm{\α}}_i)=0\\ \frac{\∂L}{\∂w}=w-\underset{i=1}{\overset{m}{\mathrm{\Σ}}}({\mathrm{\α}}_i-{\widehat{\mathrm{\α}}}_i)\\ \frac{\∂L}{\∂{\widehat{\mathrm{\ξ}}}_i}=C-{\widehat{\mathrm{\α}}}_i-{\widehat{\mathrm{\η}}}_i=0\end{array}{x}_i=0\right.---\left(5\right)$

(5) substitution (4) can be obtained to its primal-dual optimization problem is: under constraint condition (7), right the maximal value of solved function formula (6).

$\begin{array}{c}\max Q(\mathrm{\α}-\widehat{\mathrm{\α}})=-\frac{1}{2}\underset{i,j=1}{\overset{m}{\mathrm{\Σ}}}({\mathrm{\α}}_i-{\widehat{\mathrm{\α}}}_i)({\mathrm{\α}}_j-{\widehat{\mathrm{\α}}}_j)(x_i,x_j)-\\ \mathrm{\ϵ}\underset{i=1}{\overset{m}{\mathrm{\Σ}}}({\mathrm{\α}}_i+{\widehat{\mathrm{\α}}}_i)+\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{y}_i({\mathrm{\α}}_i-{\widehat{\mathrm{\α}}}_i)\end{array}---\left(6\right)$

$\begin{array}{cc}\mathrm{subject}& \mathrm{to}\left\{\begin{array}{c}\underset{i=1}{\overset{m}{\mathrm{\Σ}}}({\mathrm{\α}}_i-{\widehat{\mathrm{\α}}}_i)=0\\ {\mathrm{\α}}_i,{\widehat{\mathrm{\α}}}_i\∈[0,C]\end{array}\right.\end{array}---\left(7\right)$

The KKT covering condition of its correspondence is

$\left\{\begin{array}{c}{\mathrm{\α}}_i(y_i-(w\·x_i)-b-\mathrm{\ϵ}-{\mathrm{\ξ}}_i)=0\\ {\widehat{\mathrm{\α}}}_i((w\·x_i)+b-y_i-\mathrm{\ϵ}-{\widehat{\mathrm{\ξ}}}_i)=0\\ {\mathrm{\ξ}}_i{\widehat{\mathrm{\ξ}}}_i=0\\ {\mathrm{\α}}_i{\widehat{\mathrm{\α}}}_i=0\\ ({\mathrm{\α}}_i-C){\mathrm{\ξ}}_i=0\\ ({\widehat{\mathrm{\α}}}_i-C){\widehat{\mathrm{\ξ}}}_i=0\end{array}\right.---\left(8\right)$

Solve after the problems referred to above, can obtain w and treat estimation function

$\left\{\begin{array}{c}w=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}({\mathrm{\α}}_i-{\widehat{\mathrm{\α}}}_i)x_i\\ f\left(x\right)=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}({\mathrm{\α}}_i-{\widehat{\mathrm{\α}}}_i)(x_i,x)+b\end{array}\right.---\left(9\right)$

Utilize support vector machine to solve nonlinear problem, first utilize a Nonlinear Mapping by training dataset Nonlinear Mapping to high-dimensional feature space, it is Hilbert space, non-linear function regression problem is converted into the linear function regression problem in high-dimensional feature space, the method of conversion is to introduce the thought of kernel function, as radial basis function;

The prediction of support vector machine is using representative mixed coal Y value and G Value Data and predicts that the coal petrography index obtaining is as input parameter, using coke quality index as output parameter, support vector machine is trained, form the nonlinear relationship of input and output parameter, then using mixed coal achievement data to be predicted as input parameter, obtain the index of coke;

(3) set up coke shatter strength and scuff resistance forecast model

Coking coal vitrinite reflectance no matter in distribution graph of reflectivity in what position, no matter be continuity reflectivity distribution curve or interruption class mixture coal reflectivity distribution curve, vitrinite's total reflectivity of participating in Coal Blending is all participated in this coal blending model and prediction coke quality index system, and the each point value of vitrinite used total reflectivity is all dependent variable, therefore this factor is remarkable to the result accuracy of prediction.From technology machine system, extract vitrinite reflectance data, according to the division on 0.05 rank, by long-flame coal to the meager lean coal stage, leave altogether 60 data segments, be no matter that the coking of coalingging of tamping coking or top is very wide in range to the applicability of coal, practical enterprise expanded to coking coal resource and there is important actual directive significance.Meanwhile, any one the coal vitrinite reflection that participates in coal blending is distributed all can fall in these 60 data segments, and use respectively D ₁, D ₂d ₆₀represent.

The forecast model version of Coke Quality is constructed as follows:

M ₄₀＝SvmM40(D ₁,D ₂...D ₆₀,B _mix,G _mix,Y _mix)

M ₁₀＝SvmM10(D ₁,D ₂...D ₆₀,B _mix,G _mix,Y _mix)

Wherein, the anticipation function that SvmM40 and SvmM10 are self-defining support vector machine.

(4) set up coke reactivity CRI and post-reaction strength CSR forecast model

Forecast model version is as follows:

CRI＝SvmCRI(D ₁,D ₂...D ₆₀,B _mix,G _mix,Y _mix)

CSR＝SvmCSR(D ₁,D ₂...D ₆₀,B _mix,G _mix,Y _mix)

Wherein, the anticipation function that SvmCRI and SvmCSR are self-defining support vector machine.

Three. adopt the cohesiveness index level of coking mixed coal, comprise that maximum thickness of plastic layer Y value, caking index G value are two factors, and be entered into ature of coal property information database; Adopt single coal petrography index level of planting coking coal, comprise lazy two factors that form than parameter of work that the meticulous paragraph data of coal vitrinite total reflectivity, coal petrography maceral composition calculate, and the mixed coal ature of coal information database that formed thus, and adopt support vector base technology to predict the shatter strength M of coke ₄₀, scuff resistance M ₁₀and reactive CRI and post-reaction strength CSR.

The present invention compares with existing similar technology, and its significant beneficial effect is embodied in:

1. the features such as the inventive method and the system that forms have that the coking coal of covering affects level and factor is many, and the inherent index parameter contained is many.There is especially the Forecasting Methodology of being practicality, result precision high.The mixed coal factor of selecting from the viewpoint of Coke Quality prediction and predict the outcome etc., the method and system are than employing V _dafwith G value, divide 6 sections etc. all to there is obvious advantage than G and coal vitrinite reflectance.Prognoses system has important directive function to Experiment Coke Oven platform and commercial coke oven coke quality etc.

2. the present invention is with the relevant ature of coal cohesiveness index level (comprising maximum thickness of plastic layer Y value and two factors of caking index G value) of mixed coal, coal petrography index level is predicted (comprising the lazy ratio of work of vitrinite's total reflectivity and micropetrological unit) cold strength and the hot performance of coke, wherein for predicting the coal petrography vitrinite total reflectivity difference of coke quality and the method for other piecewise predictions, for the prognoses system that coaling in top or tamping coking technique adopts, therefore in Forecasting Methodology and system formation, all coking coal kinds of coal-blending coking will likely be participated in, comprise long-flame coal, bottle coal, 1/3 coking coal, rich coal, gas-fat coal, coking coal, lean coal, meager coals etc. can embody the basic underlying variables of all conducts prediction of numerical value in coal vitrinite reflectance, simultaneously in conjunction with commercial coke oven coke quality data or Experiment Coke Oven coke quality experimental data, adopt Nonlinear Support Vector Machines technology to set up the model of prediction coke quality index, and form prediction module.Along with continuous collection the input database of coking coal and coke quality data, then by system self training, can complete automatic adjustment predictive equation, realize the optimization object of prediction coke quality model.

3. the present invention has adopted and in process of coking, can characterize the maximum thickness of plastic layer Y value and expression plastic mass cohesiveness quality that coking coal represents plastic mass quantity in softening process, the G value of cohering ability is two factors, the intrinsic properties of coking coal is taking coal petrography index as basal expression simultaneously, therefore the efficiency index that reflects coking coal metamorphic grade is the maximum vitrinite of coal total reflectivity factor, and form the active component of plastic mass and the ratio of inert constituent in conjunction with can soften melting in macerals in coking process, the i.e. lazy factor that compares alive, this factor forms coke quality to coking coal and has important control action.Therefore above-mentioned two level four factors form the method for mixed coal prediction coke quality, can realize the forecasting process that prediction coke strength and hot performance are target.Therefore, the present invention is taking ature of coal cohesiveness index level (comprising maximum thickness of plastic layer Y value and two factors of caking index G value), coal petrography index level (comprising two factors of the lazy ratio of work of vitrinite's full constituent reflectivity and micropetrological unit) as main body, form Nonlinear Prediction Models by the support vector machine technology of setting up between coking mixed coal index and prediction coke quality index, realized the physical strength M of the two level four factor prediction coke that form with coking coal ature of coal cohesiveness level and coal petrography level ₄₀, M ₁₀with hot performance CRI, CSR.The present invention is to provide the high forecast model of multiparameter accuracy that can form ature of coal quality index and coke quality index, there is real-time update or manual intervention function simultaneously.

Brief description of the drawings

The method model schematic diagram of Tu1Shi coal petrography vitrinite total reflectivity nonlinear optimization coal blending prediction coke quality.

Fig. 2 is typical coking coal distribution graph of reflectivity instance graph.

Fig. 3 is utilized coal petrography vitrinite full constituent value, G, the Y of mixed coal and is lived lazy than the coke M of prediction by support vector machine ₄₀and graph of a relation between measured value.

Fig. 4 is utilized coal petrography vitrinite full constituent value, G, the Y of mixed coal and is lived lazy than the coke M of prediction by support vector machine ₁₀and graph of a relation between measured value.

Fig. 5 is utilized coal petrography vitrinite full constituent value, G, the Y of mixed coal and is lived lazy than the graph of a relation between coke CRI and the measured value of prediction by support vector machine.

Fig. 6 is utilized coal petrography vitrinite full constituent value, G, the Y of mixed coal and is lived lazy than the graph of a relation between coke CSR and the measured value of prediction by support vector machine.

Embodiment

Be described in more detail by reference to the accompanying drawings the present invention below by embodiment.

The present invention is the quality index of predicting coke by support vector machine technology, by the ature of coal cohesiveness index of coking mixed coal check analysis, comprises plastometer indices and caking index two factors, both input parameters; Coal petrography index, comprises the lazy ratio of work of vitrinite's full constituent reflectivity and micropetrological unit, and as two factors of independent variable, both input parameters, by the physical strength M of coke ₄₀, M ₁₀as output parameter, by the training to support vector machine, obtain the nonlinear relationship of input and output parameter with hot performance CRI, CSR.Lazy the work of the plastometer indices of mixed coal to be predicted, caking index, vitrinite's full constituent reflectivity and the micropetrological unit input parameter that is compared to is input to the support vector machine training, just can obtains physical strength and the hot performance of the coke of prediction.

1. extract data

According to the Experiment Coke Oven coking database of setting up, coding extracts the data that need, from experiment, extract 86 effective coke making and coal blending schemes and relevant ature of coal quality index plastometer indices, caking index, and coal vitrinite reflectance and live lazy ratio and corresponding coke quality indication information, comprise physical strength M ₄₀, M ₁₀with and hot performance CRI, CSR.

In order to make experimental data representative, every 7 data, with wherein front 6 as training data, rear 1 as test data.The like, extract altogether 74 groups of data and 12 groups of data for testing for training.

2. training

For different input parameters, for fear of the larger less parameter generating negative effect of unusual sample data parameter logarithm value of numerical value, produce the phenomenon that large number is eaten decimal, similar parameter is all adopted to normalized, and so-called unusual sample data refers to respect to the large or especially little especially sample vector of other input samples.After normalization, the training time reduces, and avoids the phenomenon that causes that training cannot restrain, and normalizing equation is:

y＝(ymax-ymin)*(x-xmin)/(xmax-xmin)+ymin；

Wherein: ymax and ymin default value are respectively 1 and-1, the maximal value of the input parameter that xmax is same scheme, the minimum value of the input parameter that xmin is same scheme.

Support vector machine is the structural risk minimization based on Statistical Learning Theory, utilize maximum interphase sorter thought and the method based on core to combine, can solve preferably the practical problemss such as small sample, non-linear, high dimension drawn game portion minimal point, effectively avoid " cross and fit ", following sample is had to good generalization ability

$L^{\mathrm{\ϵ}}(x,y,f)={|y-f\left(x\right)|}_{\mathrm{\ϵ}}=\left\{\begin{array}{cc}0& |y-f\left(x\right)|\≤\mathrm{\ϵ}\\ |y-f\left(x\right)|-\mathrm{\ϵ}& \mathrm{else}\end{array}\right.---\left(1\right)$

Wherein f is the real-valued function on the X of territory.Its meaning is if the difference between predicted value and actual value is while being less than ε, loss equals 0, otherwise be that the absolute value of predicted value and actual value difference is poor with it, finds a kind of estimation regression function in linear function set, f (x)=(wx)+bw, x ∈ R ⁿ, b ∈ R, wherein, (x ₁, y ₁) ... (x _m, y _m) be independent identically distributed data, b is amount of bias.Regression estimation problem is to ask parameter w and b, makes, for the input x beyond sample, to meet | f (x)-(wx)-b|≤ε, asks parameter w and b to be equivalent to the minimum value of asking following formula:

$\min \mathrm{\Φ}\left(w\right)=\frac{1}{2}{\left|\right|w\left|\right|}^2=\frac{1}{2}(w\·w)---\left(2\right)$

$\begin{array}{cc}\mathrm{subject}& \mathrm{to}\left\{\begin{array}{c}{y}_i-((w\·x_i)+b)\≤\mathrm{\ϵ}\\ ((w\·x_i)+b)-y_i\≤\mathrm{\ϵ}\end{array}\right.\end{array},i=\mathrm{1,2},...,m$

For guaranteeing that above-mentioned optimization problem has solution, introduce slack variable ξ, optimization problem is converted into the minimum problems solving under following formula constraint:

$\min \mathrm{\Φ}\left(w\right)=\frac{1}{2}{\left|\right|w\left|\right|}^2+C\underset{i}{\overset{m}{\mathrm{\Σ}}}({\mathrm{\ξ}}_i+{\widehat{\mathrm{\ξ}}}_i)---\left(3\right)$

$\begin{array}{cc}\mathrm{subject}& \mathrm{to}\left\{\begin{array}{c}{y}_i-((w\·x_i)+b)\≤\mathrm{\ϵ}+{\mathrm{\ξ}}_i\\ ((w\·x_i)+b)-y_i\≤\mathrm{\ϵ}+{\widehat{\mathrm{\ξ}}}_i\\ {\mathrm{\ξ}}_i,{\widehat{\mathrm{\ξ}}}_i\≥0,i=\mathrm{1,2},...,l\end{array}\right.\end{array}$

Wherein C is the constant of specifying, and C>0 is used for the smoothness of representative function f and permissible error and is greater than the compromise between the numerical value of ε, is mainly improving generalization ability and is reducing to rise between error regulating and controlling effect.ε is a positive number, need to set in advance, is mainly to wish for control algolithm the precision reaching;

Ask optimum solution by former problem is converted into dual problem, set up Lagrange function according to objective function and constraint condition, the dual form of optimization problem is:

$\begin{array}{c}L(w,b,{\mathrm{\ξ}}_i,{\widehat{\mathrm{\ξ}}}_i)=\frac{1}{2}{\left|\right|w\left|\right|}^2+C\underset{i}{\overset{m}{\mathrm{\Σ}}}({\mathrm{\ξ}}_i+{\widehat{\mathrm{\ξ}}}_i)-\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\α}}_i[\mathrm{\ϵ}+{\mathrm{\ξ}}_i-y_i+((w\·x_i)+b)]-\\ \underset{i=1}{\overset{m}{\mathrm{\Σ}}}{\widehat{\mathrm{\α}}}_i[\mathrm{\ϵ}+{\widehat{\mathrm{\ξ}}}_i+y_i-((w\·x_i)+b]-\underset{i=1}{\overset{m}{\mathrm{\Σ}}}(n_i{\mathrm{\ξ}}_i+{\widehat{\mathrm{\η}}}_i{\widehat{\mathrm{\ξ}}}_i)\end{array}---\left(4\right)$

Wherein, w, b, ξ _i, for former variable; α _i, η _i, for dual variable, and meet α _i, η _i, variable in above formula is asked to local derviation, can obtain

$\left\{\begin{array}{c}\frac{\∂L}{\∂b}=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}({\widehat{\mathrm{\α}}}_i-{\mathrm{\α}}_i)=0\\ \frac{\∂L}{\∂w}=w-\underset{i=1}{\overset{m}{\mathrm{\Σ}}}({\mathrm{\α}}_i-{\widehat{\mathrm{\α}}}_i)\\ \frac{\∂L}{\∂{\widehat{\mathrm{\ξ}}}_i}=C-{\widehat{\mathrm{\α}}}_i-{\widehat{\mathrm{\η}}}_i=0\end{array}{x}_i=0\right.---\left(5\right)$

(5) substitution (4) can be obtained to its primal-dual optimization problem is: under constraint condition (7), right the maximal value of solved function formula (6).

$\begin{array}{c}\max Q(\mathrm{\α}-\widehat{\mathrm{\α}})=-\frac{1}{2}\underset{i,j=1}{\overset{m}{\mathrm{\Σ}}}({\mathrm{\α}}_i-{\widehat{\mathrm{\α}}}_i)({\mathrm{\α}}_j-{\widehat{\mathrm{\α}}}_j)(x_i,x_j)-\\ \mathrm{\ϵ}\underset{i=1}{\overset{m}{\mathrm{\Σ}}}({\mathrm{\α}}_i+{\widehat{\mathrm{\α}}}_i)+\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{y}_i({\mathrm{\α}}_i-{\widehat{\mathrm{\α}}}_i)\end{array}---\left(6\right)$

$\begin{array}{cc}\mathrm{subject}& \mathrm{to}\left\{\begin{array}{c}\underset{i=1}{\overset{m}{\mathrm{\Σ}}}({\mathrm{\α}}_i-{\widehat{\mathrm{\α}}}_i)=0\\ {\mathrm{\α}}_i,{\widehat{\mathrm{\α}}}_i\∈[0,C]\end{array}\right.\end{array}---\left(7\right)$

The KKT covering condition of its correspondence is

$\left\{\begin{array}{c}{\mathrm{\α}}_i(y_i-(w\·x_i)-b-\mathrm{\ϵ}-{\mathrm{\ξ}}_i)=0\\ {\widehat{\mathrm{\α}}}_i((w\·x_i)+b-y_i-\mathrm{\ϵ}-{\widehat{\mathrm{\ξ}}}_i)=0\\ {\mathrm{\ξ}}_i{\widehat{\mathrm{\ξ}}}_i=0\\ {\mathrm{\α}}_i{\widehat{\mathrm{\α}}}_i=0\\ ({\mathrm{\α}}_i-C){\mathrm{\ξ}}_i=0\\ ({\widehat{\mathrm{\α}}}_i-C){\widehat{\mathrm{\ξ}}}_i=0\end{array}\right.---\left(8\right)$

Solve after the problems referred to above, can obtain w and treat estimation function

$\left\{\begin{array}{c}w=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}({\mathrm{\α}}_i-{\widehat{\mathrm{\α}}}_i)x_i\\ f\left(x\right)=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}({\mathrm{\α}}_i-{\widehat{\mathrm{\α}}}_i)(x_i,x)+b\end{array}\right.---\left(9\right)$

Utilize support vector machine to solve nonlinear problem, first utilize a Nonlinear Mapping by training dataset Nonlinear Mapping to high-dimensional feature space, it is Hilbert space, non-linear function regression problem is converted into the linear function regression problem in high-dimensional feature space, the method of conversion is to introduce the thought of kernel function, as radial basis function.

Support vector machine is trained, obtain the vector value of w and b, then predicted data.

3. prediction

With the lazy ratio of work of 12 groups of plastometer indicess that obtain, caking index, vitrinite's full constituent reflectivity and micropetrological unit, as input parameter, by support vector machine computing, just can obtain the predicted value of coke, and then by renormalization function, obtain the actual prediction value of coke.

Below Forecasting Methodology of the present invention and existing several typical Forecasting Methodology by coking coal prediction coke quality are made comparisons, pass through support vector machine, with the total reflectivity distribution of mixed coal vitrinite, the lazy ratio of living, G value and Y value prediction, concrete grammar and result of calculation example are compared as follows:

(1) by support vector machine, with the total reflectivity distribution of mixed coal vitrinite, the lazy ratio of living, G value and Y value prediction coke M ₄₀and M ₁₀, predict the outcome as Fig. 3 and Fig. 4, M ₄₀and M ₁₀average error is respectively 2.13% and 0.43%.

(2) by support vector machine, distribute, live with the each rank of mixed coal vitrinite reflectance and lazyly predict coke CRI and CSR than G value and Y value, predict the outcome as Fig. 5 and Fig. 6, CRI and CSR average error are respectively 1.48% and 2.60%.

The tables of data obtaining from above-mentioned forecasting process, can find out and adopt support vector machine technology, in conjunction with cohesiveness index maximum thickness of plastic layer Y value and the caking index G value of coking coal, and coal petrography vitrinite total reflectivity with live lazy ratio can be predicted preferably coke quality result, no matter be coke M ₄₀and M ₁₀or coke is all very little at CRI and CSR average error value, and can meets coke quality and detect and analyze GB requirement, and along with the increase of Experiment Coke Oven test figure or commercial coke oven coke quality data volume, the precision predicting the outcome will further reduce, and more approach true value.

Claims (1)

1. a method for coal petrography vitrinite total reflectivity nonlinear optimization coal blending prediction coke quality, is characterized in that the method realizes according to the following steps:

One. set up coal for coking resource information database

By the cohesiveness index of coking mixed coal, comprise that maximum thickness of plastic layer Y value, caking index G value are two factors of a level, and be entered into ature of coal property information database; By single coal petrography index of planting coking coal, comprise lazy level two factors that form than parameter of work that the meticulous paragraph data of coal vitrinite total reflectivity, coal petrography maceral composition calculate, and the mixed coal ature of coal input information that formed is thus in ature of coal property information database, sets up coal for coking resource information database;

Two. set up Coke Quality Prediction Models

According to the ature of coal cohesiveness index level of coking mixed coal, comprise maximum thickness of plastic layer Y value and two factors of caking index G value; Coal petrography index level, comprises that four indexs of the lazy ratio of work of vitrinite's full constituent reflectivity and micropetrological unit predict the quality index of coke, and the prediction main body of mixed coal is the shatter strength M of coke ₄₀, scuff resistance M ₁₀and reactive CRI and post-reaction strength CSR;

(1) determine coking mixed coal quality index

According to the mixed coal cohesiveness index forming after coal preparation technology comminutor of Experiment Coke Oven or the actual use of coke-oven plant's commercial coke oven or Experiment Coke Oven pulverize and coordinate after coal cohesiveness index, comprise maximum thickness of plastic layer Y value and two factors of caking index G value, coal petrography index, comprise the lazy quality of recently predicting coke of work of vitrinite's full constituent reflectivity and micropetrological unit, this prediction index system is to adopt support vector machine technology, carry out in Computing process in the actual data that detect of employing, above-mentioned two cohesiveness factors according to coal characteristic to coking coal and two coal petrography index factors adopt support vector machine technology to predict, in coke making and coal blending cohesiveness index maximum thickness of plastic layer Y value and caking index G value and coal petrography index prediction coke quality index, in maceral, active component comprises vitrinite, part half vitrinite and stable group, inert constituent comprises inertinite, half vitrinite and mineral, wherein:

Active component=vitrinite+1/3 × half vitrinite+stablize group

Inert constituent=inertinite+2/3 × half vitrinite+mineral

Lazy ratio=active component/inert constituent alive

Undertaken by following provisions:

(1) the maximum thickness of colloidal matter layer Y value of mixed coal and caking index G value are as the criterion with actual analysis value

(2) the lazy ratio of the work of coking coal vitrinite total reflectivity and micropetrological unit is pressed computation model

D _jbe in model mixed coal vitrinite in the frequency of j point reflection rate:

$D_j=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}(\mathrm{DSingl}{e}_j\×P_i)$

Wherein, DSingle _j(j=1,2,3 ... 60) frequency distributing in j point reflection rate for each the Dan Zhong coal vitrinite gathering is continuity reflectivity Distribution Value; P _iit is the proportioning of i kind list kind coal;

The lazy ratio of work of mixed coal is made as B:

$B=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}(A_i/I_i\×P_i)$

Wherein, A _ibe the active component of i kind coal, I _ibe the inert constituent of i kind coal, P _ibe i kind coal shared number percent in Coal Blending Schemes, n is the single total number of planting coal in Coal Blending Schemes;

(2) prediction coke quality index

Compare according to the coke quality check analysis data of Experiment Coke Oven coking tests or commercial coke oven and actual coal detection analysis, Mechanical Strength of Coke and hot performance index are predicted, and adopt support vector machine technology, support vector machine is the structural risk minimization based on statistical theory, utilize maximum interphase sorter thought and the method based on core to combine

$L^{\mathrm{\ϵ}}(x,y,f)={|y-f\left(x\right)|}_{\mathrm{\ϵ}}=\left\{\begin{array}{cc}0& |y-f\left(x\right)|\≤\mathrm{\ϵ}\\ |y-f\left(x\right)|-\mathrm{\ϵ}& \mathrm{else}\end{array}\right.---\left(1\right)$

Wherein f is the real-valued function on the X of territory, its meaning is if the difference between predicted value and actual value is while being less than ε, loss equals 0, otherwise the absolute value that is predicted value and actual value difference is poor with it, in linear function set, find a kind of regression function of estimating, f (x)=(wx)+bw, x ∈ R ⁿ, b ∈ R, wherein, (x ₁, y ₁) ... (x _m, y _m) be independent identically distributed data, b is amount of bias, regression estimation problem is to ask parameter w and b, makes, for the input x beyond sample, to meet | f (x)-(wx)-b|≤ε, asks parameter w and b to be equivalent to the minimum value of asking following formula:

$\min \mathrm{\Φ}\left(w\right)=\frac{1}{2}{\left|\right|w\left|\right|}^2=\frac{1}{2}(w\·w)---\left(2\right)$

$\begin{array}{cc}\mathrm{subject}& \mathrm{to}\left\{\begin{array}{c}{y}_i-((w\·x_i)+b)\≤\mathrm{\ϵ}\\ ((w\·x_i)+b)-y_i\≤\mathrm{\ϵ}\end{array}\right.\end{array},i=\mathrm{1,2},...,m$

For guaranteeing that above-mentioned optimization problem has solution, introduce slack variable ξ, optimization problem is converted into the minimum problems solving under following formula constraint:

$\min \mathrm{\Φ}\left(w\right)=\frac{1}{2}{\left|\right|w\left|\right|}^2+C\underset{i}{\overset{m}{\mathrm{\Σ}}}({\mathrm{\ξ}}_i+{\widehat{\mathrm{\ξ}}}_i)---\left(3\right)$

$\begin{array}{cc}\mathrm{subject}& \mathrm{to}\left\{\begin{array}{c}{y}_i-((w\·x_i)+b)\≤\mathrm{\ϵ}+{\mathrm{\ξ}}_i\\ ((w\·x_i)+b)-y_i\≤\mathrm{\ϵ}+{\widehat{\mathrm{\ξ}}}_i\\ {\mathrm{\ξ}}_i,{\widehat{\mathrm{\ξ}}}_i\≥0,i=\mathrm{1,2},...,l\end{array}\right.\end{array}$

Wherein C is the constant of specifying, C>0, being used for the smoothness of representative function f and permissible error is greater than the compromise between the numerical value of ε, mainly improving generalization ability and reducing to rise between error regulating and controlling effect, ε is a positive number, need to set in advance, be used for control algolithm and wish the precision reaching;

Ask optimum solution by former problem is converted into dual problem, set up Lagrange function according to objective function and constraint condition, the dual form of optimization problem is:

$\begin{array}{c}L(w,b,{\mathrm{\ξ}}_i,{\widehat{\mathrm{\ξ}}}_i)=\frac{1}{2}{\left|\right|w\left|\right|}^2+C\underset{i}{\overset{m}{\mathrm{\Σ}}}({\mathrm{\ξ}}_i+{\widehat{\mathrm{\ξ}}}_i)-\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\α}}_i[\mathrm{\ϵ}+{\mathrm{\ξ}}_i-y_i+((w\·x_i)+b)]-\\ \underset{i=1}{\overset{m}{\mathrm{\Σ}}}{\widehat{\mathrm{\α}}}_i[\mathrm{\ϵ}+{\widehat{\mathrm{\ξ}}}_i+y_i-((w\·x_i)+b]-\underset{i=1}{\overset{m}{\mathrm{\Σ}}}(n_i{\mathrm{\ξ}}_i+{\widehat{\mathrm{\η}}}_i{\widehat{\mathrm{\ξ}}}_i)\end{array}---\left(4\right)$

Wherein, w, b, ξ _i, for former variable; α _i, η _i, for dual variable, and meet α _i, η _i, variable in above formula is asked to local derviation, can obtain

$\left\{\begin{array}{c}\frac{\∂L}{\∂b}=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}({\widehat{\mathrm{\α}}}_i-{\mathrm{\α}}_i)=0\\ \frac{\∂L}{\∂w}=w-\underset{i=1}{\overset{m}{\mathrm{\Σ}}}({\mathrm{\α}}_i-{\widehat{\mathrm{\α}}}_i)\\ \frac{\∂L}{\∂{\widehat{\mathrm{\ξ}}}_i}=C-{\widehat{\mathrm{\α}}}_i-{\widehat{\mathrm{\η}}}_i=0\end{array}{x}_i=0\right.---\left(5\right)$

(5) substitution (4) can be obtained to its primal-dual optimization problem is: under constraint condition (7), right the maximal value of solved function formula (6):

$\begin{array}{c}\max Q(\mathrm{\α}-\widehat{\mathrm{\α}})=-\frac{1}{2}\underset{i,j=1}{\overset{m}{\mathrm{\Σ}}}({\mathrm{\α}}_i-{\widehat{\mathrm{\α}}}_i)({\mathrm{\α}}_j-{\widehat{\mathrm{\α}}}_j)(x_i,x_j)-\\ \mathrm{\ϵ}\underset{i=1}{\overset{m}{\mathrm{\Σ}}}({\mathrm{\α}}_i+{\widehat{\mathrm{\α}}}_i)+\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{y}_i({\mathrm{\α}}_i-{\widehat{\mathrm{\α}}}_i)\end{array}---\left(6\right)$

$\begin{array}{cc}\mathrm{subject}& \mathrm{to}\left\{\begin{array}{c}\underset{i=1}{\overset{m}{\mathrm{\Σ}}}({\mathrm{\α}}_i-{\widehat{\mathrm{\α}}}_i)=0\\ {\mathrm{\α}}_i,{\widehat{\mathrm{\α}}}_i\∈[0,C]\end{array}\right.\end{array}---\left(7\right)$

The KKT covering condition of its correspondence is

$\left\{\begin{array}{c}{\mathrm{\α}}_i(y_i-(w\·x_i)-b-\mathrm{\ϵ}-{\mathrm{\ξ}}_i)=0\\ {\widehat{\mathrm{\α}}}_i((w\·x_i)+b-y_i-\mathrm{\ϵ}-{\widehat{\mathrm{\ξ}}}_i)=0\\ {\mathrm{\ξ}}_i{\widehat{\mathrm{\ξ}}}_i=0\\ {\mathrm{\α}}_i{\widehat{\mathrm{\α}}}_i=0\\ ({\mathrm{\α}}_i-C){\mathrm{\ξ}}_i=0\\ ({\widehat{\mathrm{\α}}}_i-C){\widehat{\mathrm{\ξ}}}_i=0\end{array}\right.---\left(8\right)$

Solve after the problems referred to above, can obtain w and treat estimation function

$\left\{\begin{array}{c}w=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}({\mathrm{\α}}_i-{\widehat{\mathrm{\α}}}_i)x_i\\ f\left(x\right)=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}({\mathrm{\α}}_i-{\widehat{\mathrm{\α}}}_i)(x_i,x)+b\end{array}\right.---\left(9\right)$

Utilize support vector machine to solve nonlinear problem, first utilize a Nonlinear Mapping by training dataset Nonlinear Mapping to high-dimensional feature space, it is Hilbert space, non-linear function regression problem is converted into the linear function regression problem in high-dimensional feature space, the method of conversion is to introduce the thought of kernel function, as radial basis function;

The prediction of support vector machine is using representative mixed coal Y value and G Value Data and predicts that the coal petrography index obtaining is as input parameter, using coke quality index as output parameter, support vector machine is trained, form the nonlinear relationship of input and output parameter, then using mixed coal achievement data to be predicted as input parameter, obtain the index of coke;

(3) set up coke shatter strength and scuff resistance forecast model

Compare according to the check analysis data of Experiment Coke Oven coking tests or commercial coke oven coke and actual coal detection analysis, Mechanical Strength of Coke and hot performance index are predicted, and adopt support vector machine technology, the prediction of support vector machine is using representative mixed coal Y value and G Value Data and predicts that the coal petrography index obtaining is as input parameter, using coke quality index as output parameter, support vector machine is trained, form the nonlinear relationship of input and output parameter, then using mixed coal achievement data to be predicted as input parameter, obtain the index of coke,

From technology machine system, extract vitrinite reflectance data, according to the division on 0.05 rank, by long-flame coal to the meager lean coal stage, leave altogether 60 data segments, meanwhile, any one the coal vitrinite reflection that participates in coal blending is distributed all can fall in these 60 data segments, and use respectively D ₁, D ₂d ₆₀represent;

The forecast model version of Coke Quality is constructed as follows:

M ₄₀＝SvmM40(D ₁,D ₂...D ₆₀,B _mix,G _mix,Y _mix)

M ₁₀＝SvmM10(D ₁,D ₂...D ₆₀,B _mix,G _mix,Y _mix)

Wherein, the anticipation function that SvmM40 and SvmM10 are self-defining support vector machine;

(4) set up coke reactivity CRI and post-reaction strength CSR forecast model

Forecast model version is as follows:

CRI＝SvmCRI(D ₁,D ₂...D ₆₀,B _mix,G _mix,Y _mix)

CSR＝SvmCSR(D ₁,D ₂...D ₆₀,B _mix,G _mix,Y _mix)

Wherein, the anticipation function that SvmCRI and SvmCSR are self-defining support vector machine;

Three. adopt the cohesiveness index level of coking mixed coal, comprise that maximum thickness of plastic layer Y value, caking index G value are two factors, and be entered into ature of coal property information database; Adopt single coal petrography index level of planting coking coal, comprise lazy two factors that form than parameter of work that the meticulous paragraph data of coal vitrinite total reflectivity, coal petrography maceral composition calculate, and the mixed coal ature of coal information database that formed thus, and adopt support vector base technology to predict the shatter strength M of coke ₄₀, scuff resistance M ₁₀and reactive CRI and post-reaction strength CSR.