CN103593578A - Flue suction force feedback setting method in coke oven heating combustion process - Google Patents

Abstract

The invention discloses a flue suction force feedback setting method in the coke oven heating combustion process. The flue suction force feedback setting method is characterized in that the method comprises the following steps that 1 an air coefficient prediction model is established by the adoption of an LSSVM based on a maximum linear independent set; 2 based on the air coefficient prediction model, a flue suction force feedback set value under the typical air door opening degree is calculated according to the mixed gas combustion mechanism; 3 feed-forward correction and dynamic compensation are conducted on a flue suction force compensation value according to the experience of experts so as to realize feed-forward correction and dynamic compensation of the flue suction force set value, and therefore the control requirements for flue suction force under different air door baffle opening values can be met. In the flue suction force online setting method in the coke oven heating combustion process, due to the fact that the air coefficient prediction module is established by the adoption of the LSSVM based on the maximum linear independent set, the method has the advantages of being high in model establishing speed and high in precision. In addition, real-time feedback setting of flue suction force in the coke oven heating combustion process is realized through online prediction of the air coefficient and flue suction force setting under different air door baffle opening degrees is realized according to the expert rules.

Description

Coke oven heating flame process flue suction force feedback establishing method

Technical field

The present invention relates to process monitoring and control technology field that Ferrous Metallurgy is produced, a kind of capacity coke oven heating flame process flue suction force feedback establishing method is particularly provided.

Background technology

As primary raw material and the fuel of the industries such as metallurgy, machinery, chemical industry, coke has been widely used in the aspects such as blast furnace ironmaking, calcium carbide, gasification, casting and non-ferrous metal metallurgy, occupies very consequence in the national economic development.

Coke oven is one of industrial furnace very important in coal chemical industry, is the crucial production equipment of coke.Coke oven heating flame process joins coal gas amount and the air capacity of coke oven combustion chamber by control, maintain the stable of coke oven temperature, makes carbonization chamber keep good destructive distillation condition, to produce qualified coke.In heating flame process, the main mixed gas that adopts heats, and in blast furnace gas, sneaks into the coke-oven gas of 3%-7% as main fuel.By setting flue suction force adjusting, maintain suitable air coefficient, for coal gas can fully be burnt, maintain coke oven temperature stable, improve coke quality, guarantee that the steady suitable operation of coke oven is very crucial.

In actual coke oven heating flame process, the quality of air coefficient is the important indicator of judgement combustion process quality, for coal gas can fully be burnt, must maintain suitable air coefficient by regulating flue suction force to regulate.Industry spot generally obtains the variation of gas flow according to the deviation of temperature, and the setting of flue suction force often obtains by the mode of feedforward, according to current gas flow, changes, and determines the variation of flue suction force according to artificial experience.This is mainly because oxygen analyser easily damages, and industry spot is not generally installed oxygen analyser, causes flue suction force to carry out real-time feedback control according to current oxygen level.

In the situation that production scene lacks calorific value pick-up unit and waste gas oxygen analyser table, waste gas oxygen level obtains by off-line manual measurement, the waste gas oxygen level obtaining by off-line estimates that whether air coefficient is suitable, cause time-lag effect larger, can not adjust timely the air-fuel ratio of coking production process.So the online Prediction of carrying out to air coefficient, realizes the Real-time Feedback of flue suction force in coke oven heating flame process and set, the energy-saving effect that improves coke oven has very important meaning.

Summary of the invention

In order to realize the reasonable setting of flue suction force in coke oven heating flame process, improve burning efficiency, guarantee the steady along operation of coke oven heating flame process, the present invention proposes the real-time establishing method of a kind of flue suction force based on oxygen level On-line Estimation.Utilize the weighting LSSVM method of maximal system of linear independence to set up air coefficient prediction model, carry out air coefficient On-line Estimation, and the several online updatings based on space projection and increment recursive algorithm; The estimated value of air coefficient of take is feedback, according to combustion mechanism, obtain typical damper opening value lower flue suction feedback setting value, and adopt Expert Rules to the correction that feedovers of the flue suction force under different damper opening values, thereby form the flue suction force setting value real-time optimization scheme that a kind of feedforward and feedback combine.

Technical solution of the present invention is as follows:

The air coefficient estimation model of foundation based on maximal system of linear independence weighting LSSVM, by to the collection of gas flow, coke-oven gas flow and flue suction force and pre-service, and by the maximal system of linear independence of structure history data set, utilize weighting specifically to comprise that LSSVM method sets up air coefficient prediction model to improve modeling speed, and by the renewal of sample and the self-adaptation of model, guarantee the real-time that air coefficient is estimated, specifically comprise:

(1) by Analysis on Mechanism, determine the input and output of air coefficient computation model, form modeling history data set;

(2) blast furnace gas flow, coke-oven gas flow and flue suction force detected value data in historical data base are carried out to pre-service;

(3) maximal system of linear independence of structure history data set, to form training sample set, utilizes weighting LSSVM method to set up air coefficient prediction model, improves the speed of modeling in the situation that not affecting model accuracy.

(4) volume content of oxygen operating personnel being gathered in burnt gas exit, carbon monoxide, carbon dioxide is added up, utilize space projection method to judge whether these data contain enough effective informations, upgrade training sample set, and utilize increment recursive algorithm with the parameter of new prediction model, obtain online adaptive optimization model, guarantee that current air coefficient data should be able to be reflected in air coefficient calculating timely and effectively.

By above-mentioned online weighting LSSVM model, can obtain the air coefficient discreet value of current time combustion process waste gas, using this value as value of feedback and optimum air coefficient compare, based on mixed gas combustion mechanism, obtain typical damper opening value lower flue suction feedback set amount, according to expertise and a large amount of field data, the flue suction force under other damper apertures is proofreaied and correct again, realization is set and feedforward compensation the feedback of flue suction force setting value, reduce the adverse effect of mixed gas calorific value fluctuation and random disturbance, make mixed gas in optimal combustion state.Specifically comprise:

(1) according to current coke-oven gas, blast furnace gas flow, utilize air coefficient prediction model to obtain the air coefficient discreet value of current time combustion process waste gas, this value and optimum air coefficient are compared, using deviation as basis, based on mixed gas combustion mechanism, calculate typical damper opening value lower flue suction setting value;

(2) the flue suction force setting value calculating according to expertise is proofreaied and correct, thereby realizes to the feedback modifiers of flue suction force setting value and dynamic compensation, to adapt to the control requirement of different damper opening value lower flue suction.

The invention has the advantages that: utilize the weighting LSSVM air coefficient method of estimation based on maximal system of linear independence, and the online updating method based on space projection method assurance air coefficient prediction model, the flue suction force of utilization based on combustion mechanism and Expert Rules obtains flue suction force feedback setting value and feedforward modified value, thereby form the real-time control program of flue suction force that a kind of feedforward and feedback combine, compare with the existing flue suction force computing method that just feedforward is set according to expertise, there is flue suction force setting more reasonable, reduced the adverse effect of mixed gas calorific value fluctuation and random disturbance, for the optimization operation that guarantees heating flame process, have great importance.

Accompanying drawing explanation

The original fire path temperature intelligent optimization of Fig. 1 control structure figure;

Fig. 2 fire path temperature intelligent optimization of the present invention control structure figure;

The prediction effect figure of Fig. 3 air coefficient estimation model;

The approximate error figure of Fig. 4 air coefficient estimation model;

Fig. 5 improves the operational effect figure of rear air coefficient.

Embodiment

Embodiment 1

For carbonization chamber overall height, be 6m, pusher side (side that pusher machine operates) width is 0.42m, the width of coke side (side of the discharging of the coke) is 0.48m, the mean breadth of carbonization chamber is the capacity coke oven of 0.45m, the present invention is according to blast furnace gas flow, the actual detected value of coke-oven gas flow and flue suction force, the weighting LSSVM air coefficient computing method of employing based on maximal system of linear independence, air coefficient online updating method based on space projection and increment recursive algorithm obtains air coefficient discreet value, by obtaining flue suction force feedback setting value based on combustion mechanism and Expert Rules, the flue suction force setting value calculating according to expertise is proofreaied and correct, realization is set the on-line optimization of coke oven heating flame process flue suction force.

1, the air coefficient based on maximal system of linear independence weighting LSSVM calculates

According to coke oven heating flame process mechanism, choose blast furnace gas flow, coke-oven gas flow and flue suction force detected value as auxiliary variable, utilize maximal system of linear independence method to build training sample set, adopt weighting LSSVM method to calculate the air coefficient of current time, its calculation procedure is as follows:

(1) composition of data sample:

Air coefficient α refers to the air quality of burning 1kg fuel institute effective supply and the required theoretical air mass ratio of perfect combustion 1kg fuel.In production run, for guaranteeing the perfect combustion of mixed gas, the air capacity of effective supply must be more than theoretical required air quantity.Usually, during mixed gas heating, air coefficient α is taken as 1.18～1.25, and the present invention is according to the actual conditions of production scene, by air coefficient optimal setting α _optbe taken as 1.22.In original production run, operating personnel utilize waste-gas analysis instrument to detect the volume content of oxygen, carbon monoxide, carbon dioxide in burnt gas exit, detect once every day, and calculate air coefficient, add up weekly, rule of thumb the suction in follow-up heating flame process is adjusted.The formula that calculates air coefficient is as follows:

$\mathrm{\α}=1+K\frac{{x_O}_2-0.5x_{\mathrm{CO}}}{{x_{\mathrm{CO}}}_2+x_{\mathrm{CO}}}---\left(1\right)$

In formula,

x _cOfor each composition volume content in dry waste gas, K is scale factor, generates the ratio of carbon dioxide volume and requisite oxygen tolerance while representing every cubic metre of coal gas perfect combustion.K value changes with the difference of composition of coal gas, considers that the ratio that coke-oven gas in coke oven heating flame process accounts for mixed gas is 3%～7%, during mixed gas heating, K is taken as to 2.39.

Carbon monoxide in dry waste gas, carbon dioxide are mainly produced through burning by blast furnace gas, coke-oven gas, can directly reflect by blast furnace gas flow, coke-oven gas flow detection value, oxygen in dry waste gas is mainly for participating in the remaining oxygen of burning, can indirectly reflecting by flue suction force detected value.Coke-oven gas and blast furnace gas are Large Time Delay Process to the effect of coke oven fire path temperature, value in a certain moment or short time can not be as the influence factor that changes fire path temperature, and coke-oven gas is higher than blast furnace gas calorific value, burning fast, short to the retardation time of fire path temperature.By process mechanism and historical data analysis, every 4 hours of coke oven fire path temperature detects once, find that coke-oven gas is to being the retardation time of fire path temperature 4 hours, and is 8 hours the retardation time of blast furnace gas.Therefore, the input variable of air coefficient forecast model is taken as blast furnace gas flow u last, two cycles _{b (k-1)}, u _{b (k-2)}, the coke-oven gas flow u in current gas flow and previous cycle _{c (k)}, u _{c (k-1)}, and current and last, two cycle flue suction force u _a(k-2), u _a(k-1), u _{a (k)}, output variable air coefficient y _k.Choose N group historical process data and set up Sample Storehouse { X, Y}, X={x _k}={ u _{b (k-1)}, u _{b (k-2)}, u _{c (k)}, u _{c (k-1)}, u _{a (k-2)}, u _{a (k-1)}, u _{a (k)}, Y={y _k, k=1,2 ..., N.

(2) processing of data

Sampling period due to each sensor in real system is inconsistent, in heating flame process, because the speed of topworks reaction is different, and gas flow and flue suction force be detection time per minute once, and oxygen content in air wants detect once for 8 hours.Abnormal data refers to the data that are not obviously inconsistent with actual conditions, coking production process work condition environment is complicated, heat wave in electromagnetic noise and coke pushing operation etc. all exists compared with strong jamming, makes measurement data unavoidably with various measuring error, comprises two kinds of human error and stochastic errors.Human error comprises deviation and the fault of routine measurement instrument, and the probability that in actual production, human error occurs is very little, but therefore its existence meeting severe exacerbation data quality must detect and delete in time.The generation of stochastic error is the impact that is subject to enchancement factor, is generally inevitably, but meets certain statistical law, can eliminate by digital filtering mode.In order to eliminate human error, the scope of regulation verification and measurement ratio, surpasses following scope, abandons reorganizing data.

Pusher side flue suction force [230Pa, 285Pa], coke side flue suction force ([240Pa, 295Pa]).Pusher side blast furnace gas flow [20900m ³/ h, 32600m ³/ h], coke side blast furnace gas flow [21000m ³/ h, 33500m ³/ h], pusher side coke-oven gas flow [200m ³/ h, 1500m ³/ h], coke side coke-oven gas flow [250m ³/ h, 1550m ³/ h], pusher side mixed gas pressure [400Pa, 900Pa], coke side mixed gas pressure [450Pa, 950Pa], pusher side flue suction force [200Pa, 260Pa], coke side flue suction force [205Pa, 265Pa].

The present invention, to blast furnace gas, coke-oven gas and flue suction force per minute once sampling, can obtain 240 data for 4 hours, and establishing current period is k, adopts mean filter method to process:

$u_{B(k-1)}=\left(\underset{i=0}{\overset{239}{\mathrm{\Σ}}}{u}_{B(k-1,i)}\right)/240$

$u_{B(k-2)}=\left(\underset{i=0}{\overset{239}{\mathrm{\Σ}}}{u}_{B(k-2,i)}\right)/240$

$u_{C\left(k\right)}=\left(\underset{i=0}{\overset{239}{\mathrm{\Σ}}}{u}_{C(k,i)}\right)/240$

$u_{C(k-1)}=\left(\underset{i=0}{\overset{239}{\mathrm{\Σ}}}{u}_{C(k-1,i)}\right)/240---\left(2\right)$

$u_{A(k-2)}=\left(\underset{i=0}{\overset{239}{\mathrm{\Σ}}}{u}_{A(k-2,i)}\right)/240$

$u_{A(k-1)}=\left(\underset{i=0}{\overset{239}{\mathrm{\Σ}}}{u}_{A(k-1,i)}\right)/240$

$u_{A\left(k\right)}=\left(\underset{i=0}{\overset{239}{\mathrm{\Σ}}}{u}_{A(k,i)}\right)/240$

Wherein: u _{b (k-1)}, u _{b (k-2)}blast furnace gas flows last, two cycles, u _{c (k)}, u _{c (k-1)}current and the coke-oven gas flow in last cycle,

it is the flue suction force in front 3 cycles; u _{b (k-1, i)}, u _{b (k-2, i)}, (i=0 ..., 239) and be respectively the per minute sampled value of k-1 cycle, the k-2 blast furnace gas flow in the cycle, u _{c (k, i)}, u _{c (k-1, i)}, (i=0 ..., 239) and be respectively the per minute sampled value of k, the k-1 coke-oven gas flow in the cycle, u _{a (k, i)}, u _{a (k-1, i)}, u _{a (k-2, i)}, (i=0 ..., 239) and be respectively cycle k, k-1, the k-2 flue suction force per minute sampled value in the cycle.

(3) foundation of the air coefficient prediction model based on LSSVM

When structure air coefficient weighting LSSVM higher dimensional space linear regression function, if can obtain the very big linear independence group of input vector in feature space, just can reduce the number of training data and support vector, make weighting LSSVM there is good sparse property; Computing cost due to weighting LSSVM training process is mainly used in matrix inversion operation simultaneously, and the dimension of matrix equals the number of training data, thereby can reduce the computing cost of training process.

If input vector X={x _k}={ u _{b (k-1)}, u _{b (k-2)}, u _{c (k)}, u _{c (k-1)}, u _{a (k-2)}, u _{a (k-1)}, u _{a (k)}, k=1,2 ..., N, x _k∈ R ^p, p is the dimension of input data, p=7.By following steps, solve maximum irrelevant group:

1) set up set

x _a=φ;

2) for k=1,2 ..., N's

minimizing

3) if the minimal value of trying to achieve

just corresponding

add set X _ain; Otherwise corresponding add X _l.ε is defined as linear dependence truncated error.

4) get back to step 2), until complete

calculating.

X={x ₁..., x _nby Nonlinear Mapping function

being mapped to high-dimensional feature space can be expressed as

by above step 1)-4) X ' is divided into two set, X _land X _a, X _l={ x ₁, x ₂..., x _lin element linear independence, X _acan use X _lin the linear approximate representation of element, so X _lbe approximate maximum irrelevant group, L represents the number of element in maximum irrelevant group.Can obtain base data sample set { x _i, y _i, i=1,2 ..., L.

Introduce Nonlinear Mapping training set sample is mapped to a high-dimensional feature space from luv space, thereby makes the nonlinear function estimation problem in luv space be converted into the linear function estimation problem in high-dimensional feature space, regression function is:

In formula, w, is weight vector, and b is threshold value.

Weighted Least Squares Support Vector Machines utilizes structural risk minimization structure to have the objective function that minimizes of equality constraint, and training error is increased to corresponding weights, and regression problem is become:

In formula, γ is regularization parameter, μ _ifor weight factor, e _ifor error variance.

Solve for convenience above-mentioned optimization problem, introduce by Lagrange multiplier α=[α ₁, α ₂..., α _l] ^t, convert constrained optimization problem to following unconstrained problem:

According to Karush-Kuhn-Tucker optimal condition, ask respectively L about variable (w, b, e _i, α _i) partial differential, obtain

By the arrangement to above formula, cancellation variable w and e _i, obtain

$\left[\begin{array}{cc}0& {\mathrm{\ψ}}^T\\ \mathrm{\ψ}& \mathrm{\Ω}+U_{\mathrm{\γ}}\end{array}\right]\left[\begin{array}{c}b\\ \mathrm{\α}\end{array}\right]=\left[\begin{array}{c}0\\ y\end{array}\right]---\left(7\right)$

In formula, the unit column vector that ψ is N * 1,

y=[y ₁, y ₂..., y _l] ^t, the nuclear matrix that Ω is L * L, the element of the capable j row of its i is

for kernel function, the present invention adopts gaussian kernel function , wherein σ is gaussian kernel function width parameter.The expression formula that can obtain weighting LSSVM model is:

$y\left(x\right)=\underset{i=1}{\overset{L}{\mathrm{\Σ}}}{\mathrm{\α}}_{i}K(x_i,x)+b---\left(8\right)$

The method of employing based on Euclidean distance realizes the weighting to training error, obtains weight factor μ _i.If input vector x _kwith data sample

euclidean distance be d _ik=|| x _i-x _k||, the weight of this input vector is:

${\mathrm{\&mu;}}_k=f\left(d_{\mathrm{ik}}\right)=(1-\mathrm{\&tau;})\left(\frac{1/d_{\mathrm{ik}}-1/{\mathrm{<figure-callout\; id="1"\; label="d\; max"\; filenames="HSA0000098110240000013.png,HSA0000098110240000021.png"\; state="\{\{state\}\}">d</figure-callout>}}_{\max }}{1/d_{\mathrm{mm}}-1/d_{\max }}\right)+\mathrm{\&tau;}---\left(9\right)$

In formula, d _mmfor this input vector and the corresponding minimum euclidean distance of data sample, d _maxthe corresponding maximum Euclidean distance of this input vector and data sample, and establish μ _mm=f (d _max)=0.001. μ _max=f (d _mm)=1.

Model complexity in the direct Controlling object function of regularization parameter γ in weighting LSSVM model and the weight proportion of training error, nuclear parameter σ is controlling the radius of action of kernel function, is the local function of bounded.The performance of weighting LSSVM forecast model depends on whether regularization parameter γ and nuclear parameter σ can Rational choices to a great extent.Using the square error minimum of weighting LSSVM forecast model discreet value and actual detected value as Optimization goal, it is defined as follows:

$f=\frac{1}{L}\underset{i=1}{\overset{L}{\mathrm{\&Sigma;}}}{({\hat{y}}_i-y_i)}^2=\frac{1}{L}\underset{i=1}{\overset{L}{\mathrm{\&Sigma;}}}{(y_i-(\underset{i=1}{\overset{L}{\mathrm{\&Sigma;}}}{\mathrm{\&beta;}}_{i}k(x_i,x)+b))}^2---\left(10\right)$

In formula, y _ifor the actual detected value of air coefficient,

for the model pre-estimating value of air coefficient, L is the number that training sample is concentrated data.Adopt intelligent optimization algorithm to realize the optimization of regularization parameter γ and nuclear parameter σ in forecast model and choose, regularization parameter γ span is [0.1,120], kernel function σ span is [0.1,10], after intelligent optimization algorithm optimizing, γ is taken as to 21.35, σ and is taken as 1.46.

(4) the air coefficient online updating based on space projection and increment recursive algorithm

Restriction due to Site Detection means and checkout equipment, air coefficient detected value data sample point is less, fluctuation is larger, for the consideration to coke oven heating flame process dynamics time-varying characteristics, the air coefficient data that new assay obtains should be able to be reflected in air coefficient calculating timely and effectively simultaneously.Based on space projection thought, adopt increment recursive algorithm to realize the online updating of weighting LSSVM air coefficient, guarantee that air coefficient discreet value has higher precision of prediction.

Obtaining through the new modeling data sample of the pretreated air coefficient of data x _n+1after, projected to feature space and obtained

if try to achieve minimal value

accordingly

add set X _ain; Otherwise corresponding add X _l.ε is defined as linear dependence truncated error.

As base data sample set X _ladd a new data x _n+1after, need to dynamically update weighting LSSVM air coefficient, recalculate Lagrange multiplier β and threshold value b in formula (10).

2, the flue suction force establishing method based on prediction model and Expert Rules

By above-mentioned online weighting LSSVM model, obtain the air coefficient discreet value of current time combustion process waste gas, this value and optimum air coefficient are compared, using deviate as basis, based on mixed gas combustion mechanism, set up the mapping relations of air coefficient and flue suction force under typical damper opening value, according to expertise and a large amount of field data, the flue suction force offset calculating is proofreaied and correct again, thereby realize the feedback modifiers of flue suction force setting value and dynamic compensation, thereby reduce the adverse effect of mixed gas calorific value fluctuation and random disturbance, make mixed gas in optimal combustion state.Its calculation procedure is as follows:

(1) typical damper aperture lower flue suction feedback setting value is calculated

The variation of coke-oven gas, blast furnace gas blending ratio in mixed gas, can make every cubic metre of mixed gas needed air capacity of fully burning that certain variation occurs, by the analysis of process mechanism and a large amount of Field Production Data, the variation of air capacity is directly proportional to air coefficient, can be similar to and think, at typical damper opening value (60,70] and in the constant situation of mixed gas flow, square being directly proportional of flue suction force and air coefficient, at k constantly, their pass is:

$\frac{u_{A\left(\stackrel{\&OverBar;}{k}\right)}+u_{\mathrm{\&Delta;A}\_\mathrm{fb}\left(k\right)}}{u_{A\left(\stackrel{\&OverBar;}{k}\right)}}=\frac{y_{\mathrm{opt}}^2}{{y_k}^2}---\left(11\right)$

In formula,

the mean value of the flue suction force in front 3 cycles, u _{Δ A_fb}(k) the flue suction force offset calculating according to the discreet value of air coefficient,

current air coefficient discreet value, y _kfor air coefficient discreet value, y _optfor air coefficient optimal value, be taken as 1.22.

According to formula (11), can obtain current flue suction force offset is:

$u_{\mathrm{\&Delta;A}\_\mathrm{fb}\left(k\right)}=u_{A\left(\stackrel{\&OverBar;}{k}\right)}(\frac{y_{\mathrm{opt}}^2}{{y_k}^2}-1)---\left(12\right)$

(2) the flue suction force setting value based on Expert Rules is proofreaied and correct

What more than obtain is (60 at damper opening value, the offset of flue suction force when 70] interval, due to the coupled relation between damper aperture and flue suction force, after setting up the mapping relations of air coefficient and flue suction force under typical damper opening value, need to the flue suction force offset calculating, carry out forward feedback correction according to expertise and a large amount of field data, to adapt to the control requirement of different damper opening value lower flue suction.If current time damper opening value is v _(k), flue suction force forward feedback correction value is u _{Δ A_ff (k)}, formulated following Expert Rules flue suction force offset proofreaied and correct:

By above-mentioned rule, flue suction force offset is proofreaied and correct, according to the flue suction force in front 3 cycles

add the feedback correction value u obtaining _{Δ A_fb (k+1)}with feedforward setting value u _{Δ A_ff (k+1)}, obtain final flue suction force Optimal Setting value

be handed down to valve positioner, realize the effective stabilization of flue suction force is followed the tracks of and controlled.

3, case verification

In weighting LSSVM air coefficient computation process, in order to obtain more accurate discreet value in the situation that training data is less, adopt maximal system of linear independence method to screen to form training sample set to raw data, σ is taken as 1.46.Choose 290 groups of field datas, the prediction effect of air coefficient and approximate error are as shown in Figure 3, Figure 4.

From Fig. 3 and Fig. 4, by adopting maximal system of linear independence method screening training sample set and utilizing space projection and increment recursive algorithm carries out the on-line correction of model, make air coefficient forecast model there is good prediction effect, by can be calculated, the average relative error value of forecast model is 1.05%, relative error maximal value is 3.51%, and model accuracy can meet the error requirements to air coefficient in coke oven fire path temperature optimal control.

Consider the large dead time of coke oven heating flame process, large inertial properties, for avoiding frequently adjusting flue suction force Optimal Setting value, cause production run fluctuation, the compensation cycle of flue suction force offset is made as to 10 minutes, detected value to this time period inner flue suction and blast furnace gas flow, coke-oven gas flow carries out arithmetic mean, input using the mean value of trying to achieve as air coefficient forecast model, utilizes flue suction force feedback modifiers model can obtain the feedback correction value of flue suction force.As shown in Figure 3, while adopting flue suction force setting value to simplify calculating model method calculating flue suction force setting value, the air coefficient and its optimal setting 1.22 that obtain have relatively large deviation, the optimisation strategy that adopts feedforward and feedback to combine realizes after the dynamic optimization of flue suction force setting value, and the air coefficient of coke oven heating flame process as shown in Figure 5.

As shown in Figure 5, after improving, air coefficient is basicly stable near its optimal setting 1.22, and maximum deviation is 0.1185, and average relative error is 3.41%.The flue suction force Optimal Setting strategy that feedforward and feedback combine can guarantee the abundant Reasonable Combustion of mixed gas, has realized the target of coke oven heating flame process energy conservation consumption reduction.

Claims (3)

1. a flue suction force on-line setup method in coke oven heating flame process, is characterized in that, comprises the following steps:

Step 1: set up air coefficient prediction model;

Step 2: based on air coefficient prediction model, calculate the flue suction force feedback setting value under typical throttle opening according to mixed gas combustion mechanism;

Step 3: according to expertise to flue suction force offset carry out forward feedback correction and, thereby realize feedforward correction and the dynamic compensation to flue suction force setting value, to adapt to the control requirement of different damper opening value lower flue suction.

2. step 1 according to claim 1, is characterized in that, comprises the following steps:

Step 1: blast furnace gas flow, coke-oven gas flow and flue suction force detected value data in historical data base are carried out to pre-service, and formation modeling history data set is chosen N group historical process data and set up Sample Storehouse { X, Y}, X={x _k}={ u _{b (k-1)}, u _{b (k-2)}, u _{c (k)}, u _{c (k-1)}, u _{a (k-2)}, u _{a (k-1)}, u _{a (k)}, Y={y _k, k=1,2 ..., N; In order to eliminate human error, regulation detects the scope of data.Pusher side flue suction force [230Pa, 285Pa], coke side flue suction force ([240Pa, 295Pa]).Pusher side blast furnace gas flow [20900m ³/ h, 32600m ³/ h], coke side blast furnace gas flow [21000m ³/ h, 33500m ³/ h], pusher side coke-oven gas flow [200m ³/ h, 1500m ³/ h], coke side coke-oven gas flow [250m ³/ h, 1550m ³/ h], pusher side mixed gas pressure [400Pa, 900Pa], coke side mixed gas pressure [450Pa, 950Pa], pusher side flue suction force [200Pa, 260Pa], coke side flue suction force [205Pa, 265Pa].The present invention, to blast furnace gas, coke-oven gas and flue suction force per minute once sampling, can obtain 240 data for 4 hours, and establishing current period is k, adopts mean filter method to process.

Step 2: the maximal system of linear independence X of structure history data set _l={ x ₁, x ₂..., x _lto form training sample set, utilizing LSSVM method to set up air coefficient prediction model, the dimension of maximum irrelevant group, lower than the dimension of history data set, can improve the speed of modeling in the situation that not affecting precision;

Step 3: utilize waste-gas analysis instrument to detect the volume content of oxygen, carbon monoxide, carbon dioxide in burnt gas exit, whether the data of utilizing space projection method to judge that this detection obtains contain enough effective informations, upgrade training sample set.Obtaining through the new modeling data sample of the pretreated air coefficient of data x _n+1after, projected to feature space and obtained if try to achieve minimal value

be illustrated in feature space, new samples and maximum irrelevant group linear dependence, can represent by the irrelevant group of maximum, accordingly

add set X _ain; Otherwise corresponding

add X _l.ε is defined as linear dependence truncated error.As base data sample set X _ladd a new data x _n+1after, need to dynamically update weighting LSSVM air coefficient, recalculate Lagrange multiplier β and threshold value b.

3. step 3 according to claim 1, it is characterized in that, according to expertise, flue suction force offset is carried out to forward feedback correction, thereby realize the feedforward correction to flue suction force setting value, to adapt to the control requirement of different damper opening value lower flue suction.In coke oven heating flame process, air is that the form by natural air exhaust enters combustion system, the number of air capacity directly affects the efficiency of burning, must avoid two kinds of situations in the adjusting of air capacity, and a kind of is that air capacity deficiency makes gas excess and is discharged in atmosphere; Another kind is that air excess makes waste gas take away a large amount of heats, and both of these case has directly caused the very big reduction of energy dissipation and burning efficiency.The adjusting that coke oven heats required air capacity divides two parts: the one, and the aperture of change air air inlet baffle plate, the cardinal principle balance of assurance coal gas and air capacity, i.e. coarse adjustment process; The 2nd, regulate flue suction force.In the constant situation of air door, can control air capacity by suction, suction is by regulating flue suction force valve to guarantee the stable of coke oven air capacity.

Bearing calibration is as follows:

If current time damper opening value is v _(k), flue suction force forward feedback correction value is u _{Δ A_ff (k)}, formulated following Expert Rules flue suction force offset proofreaied and correct:

Images