CN100385204C - Measuring method of on line key paramotor based on new type generalized predictive control - Google Patents

Abstract

The present invention discloses an on-line key parameter measurement method based on general predictive control. The real time measurement of key parameters is an important problem in the process control, and the general predictive control can provide excellent accuracy and trend to the key parameter measurement. The real time on-line measurement model suggested by the present invention can make use of the ideas in the general predictive control and can realize the real time measurement of the key parameters of some time extending or hard measurement by ordinary sensors, and the real time on-line measurement model provides fundamental basis for the on-line key parameter measurement method. The present invention realizes the key parameter real-time on-line measurement of industrial processes and solves the time-domain adaptation problem of the measurement and control in boiler control, the suggested oxygen content soft measurement technique based on generalized prediction also opens up a new approach for the accurate, real-time and on-line soft measurement of other key parameters, such as flue-gas temperature in the process of boiler combustion, and the present invention can realize the closed-loop control and the optimal operation of combustion systems. The present invention is also suitable for measuring the oxygen content of the flue gas and CO flue gas content in blank heating boilers, coal gas production boiler and other production processes.

Description

Online crucial parameter measurement method based on the novel broad sense PREDICTIVE CONTROL

Technical field

The present invention relates to the process industry measurement and control area, is a kind of online crucial parameter measurement method based on the novel broad sense PREDICTIVE CONTROL.

Background technology

In the process industry field, particularly in thermal technology's industry, there is a large amount of key parameters measured that is difficult to, the situation that generally can only lean on artificial means or artificial experience to judge realizes that online crucial parameter measurement is necessary.So main at present employing soft-sensing model carries out real-time measurement, promptly the auxiliary variable that is easy to detect by measurement estimates the variable that is difficult to even can't measures based on estimation model.

The measuring equipment precision is not high in present thermal technology's industry, and investment is big, and serviceable life is short, and the measurement hysteresis is bigger, is difficult to realize the online in real time observing and controlling of combustion process.Progressively adopted soft-measuring technique to solve this type of problem in recent years.Adopting the thinking of measuring indirectly, utilize other parameter that is easy to obtain, by calculating the estimation that realizes tested measurement, is a kind of new technology-soft-measuring technique (Soft-Sensing Technique) that emerges at process control and detection range recently.Soft-measuring technique is also referred to as soft instrument technology, utilize exactly and easily survey process variable (as procedure parameters such as the pressure that obtains easily in the industrial process, temperature), easily survey soft-sensing model between process variable and the process variable to be measured that are difficult to directly measurement (as the flue gas oxygen content in the boiler combustion process etc.) according to these, by various calculating and method of estimation, thereby realization is to the measurement of process variable to be measured.

Key parameters is meant that those are requisite and be difficult to obtain in real time the accurately parameter of data with the routine measurement method in control, as fire box temperature, flue gas oxygen content (O ₂), hydrocarbon content (CO ₂, CO), oxynitrides content (NOx), water vapour content (H ₂O), dust concentration and combustion process time lag constant (τ) etc.Wherein, fire box temperature is the key parameters that guarantees burning and boiler operatiopn; Oxygen level is to be related to wind coal proportion, the good and bad significant process parameter of evaluation burning, also is important target component; And SO ₂, CO ₂, CO, H ₂O, NOx, dust concentration etc. then are that heat waste is calculated and the requisite parameter of assessment boiler efficiency.The conventional detection method of these parameters employings is difficult to guarantee in real time, requirement accurately.Though and for those be control must but the parameter that can obtain easily as parameters such as combustion chamber draft, discharge pressure, flue-gas temperatures, then do not list in the key parameters.

At present, flue gas oxygen content mainly relies on zirconia sensor and realizes measurement on the engineering.SO ₂, CO ₂Flue gas analyzer provides on the chimney by being installed in etc. parameter.Owing to reasons such as the time lag of course of reaction, installation site restriction, instrument self-characteristics, make these parameters not satisfy and measure and the adaptive requirement of control time domain.At present, why some high-caliber theoretical researches can not be applied to engineering reality, and one of them important reasons is exactly the influence of having ignored this factor.

The disadvantage that zirconia is measured flue gas oxygen content mainly contains, and the one, intoxicating phenomenon happens occasionally, and causes the measurement misalignment; The 2nd, because causing to measure, the time delay of the transient process of combustion reaction, flue gas circulation lags behind; The 3rd, stickiness causes to measure and lags behind during zirconia self.In fact, the zirconia on the engineering is measured oxygen content only just may be reacted operating mode under stable situation real result now, similarly, and SO ₂Measurement also have same problem.

The real-time measurement of key parameters is a vital problem in the boiler combustion process control.Have the large dead time characteristic signals for solving those, the key parameters that can not measure with direct method maybe adopts the method for soft measurement usually.

Generalized predictive control is a kind of advanced algorithm with forecast model, rolling optimization and feedback compensation and self-adaptation characteristics, the novel algorithm of paper studies has more the characteristics of fast convergence, can be applied to the Generalized Prediction technology among flue gas oxygen content measures, realize the soft measurement of boiler combustion process key parameters.

Summary of the invention

The purpose of this invention is to provide a kind of online crucial parameter measurement method based on the novel broad sense PREDICTIVE CONTROL, it utilizes the advantage of generalized predictive control, soft measurement key parameter model is carried out rolling optimization and feedback compensation, make soft measurement key parameter model output have good precision and trend.

The technical solution adopted for the present invention to solve the technical problems is: according to the measuring method of the time lag characteristic and the key parameter of process control, utilize the novel broad sense predictive controller, set up soft measurement forecast model based on the Generalized Prediction algorithm, realize the real-time online measuring of process control key parameters, specifically comprised measurement model foundation, initialization, novel forecast Control Algorithm, feedback compensation and based on four steps of online crucial parameter measurement method of novel broad sense PREDICTIVE CONTROL:

(1) sets up measurement model: set up measurement model according to generalized predictive control technical characterictic and process control time lag characteristic;

(2) initialization and find the solution optimal control law: set controlled variable according to measurement model, utilize the novel broad sense forecast Control Algorithm to find the solution optimal control law;

(3) model feedback compensation: become two sections method for parameter estimation of forgetting factor during employing and carry out the model feedback compensation, the optimal control parameter;

(4) based on the online crucial parameter measurement method of generalized predictive control: calculate, export the crucial parameter measurement value according to the novel broad sense predictive controller.

The measurement model of described foundation, specific as follows:

The objective function of determining the process control of system is

$J=E\{{(\mathrm{\ψ}-w)}^T(\mathrm{\ψ}-w)+\mathrm{\λ}{\mathrm{\ψ}}^T\mathrm{\ψ}\}$

$=E\{{(G\stackrel{\→}{u}+\stackrel{\→}{f}+\stackrel{\→}{e}-\stackrel{\→}{w})}^T(G\stackrel{\→}{u}+\stackrel{\→}{f}+\stackrel{\→}{e}-\stackrel{\→}{w})+\mathrm{\λ}{(P\stackrel{\→}{u}+\stackrel{\→}{\mathrm{\δ}})}^T(P\stackrel{\→}{u}+\stackrel{\→}{\mathrm{\δ}})\}$

In order to try to achieve optimal control law Can get by target function type

$(G^{T}G+\mathrm{\λ}{P}^{T}P)\stackrel{\→}{u}=G^T(\stackrel{\→}{w}-\stackrel{\→}{f}-\stackrel{\→}{e})-\mathrm{\λ}{P}^T\stackrel{\→}{\mathrm{\δ}}$

Try to achieve by equation as can be known by mathematical analysis

Be the J that makes in the target function type and obtain the optimum solution of minimum value.

Described initialization and find the solution optimal control law, specific as follows:

Because novel control algolithm adopts matrix recursion alternative manner, avoided classic method to find the solution the calculating of inverse matrix in the optimal control law, the descended square order of M (matrix exponent number) of computational complexity, calculated amount of saving and memory space just can enlarge the scope of choosing of prediction time domain and control time domain like this, increase the stability of TT﹠C system, can realize the soft measurement of boiler key parameter real-time online;

The recursion optimized Algorithm is as follows:

Note $G^{T}G+\mathrm{\λ}{P}^{T}P=\left[\begin{array}{cccc}{a}_{11}& a_{12}& \·\·\·& a_{1M}\\ a_{21}& a_{22}& \·\·\·& a_{2M}\\ \·\·\·& \·\·\·& \·\·\·& \·\·\·\\ a_{M1}& a_{M2}& \·\·\·& a_{\mathrm{MM}}\end{array}\right]=\left[\begin{array}{c}{{\stackrel{\→}{a}}_1}^t\\ {{\stackrel{\→}{a}}_2}^t\\ \·\\ \·\\ \·\\ {{\stackrel{\→}{a}}_M}^t\end{array}\right]$

$G^T(\stackrel{\→}{w}-\stackrel{\→}{f})-\mathrm{\λ}{P}^T\stackrel{\→}{\mathrm{\δ}}={\stackrel{\→}{b}}^t=[{\stackrel{\→}{b}}_1,...,{\stackrel{\→}{b}}_M]$

(1) read in M, t, ${\stackrel{\&RightArrow;}{a}}_i^t=(a_{\mathrm{<figure-callout\; id="1"\; label="i"\; filenames="C20051011579400181.png"\; state="\{\{state\}\}">i</figure-callout>}1},...,a_{\mathrm{iM}}),i=\mathrm{1,2},...,M$

e ₁=(1，0，…，0)，…，e _j=(0，…，0，1，0，…，0)，…，e _M＝(0，0，…，1)， ${\stackrel{\&RightArrow;}{b}}^t=[{\stackrel{\&RightArrow;}{b}}_1,...,{\stackrel{\&RightArrow;}{b}}_M],$ ${\stackrel{\&RightArrow;}{u}}^T=(u_1,...,u_M),$ u _i＝△u _f(t+i-1)，i＝1，2，…，M

(2) appoint and get ${\stackrel{\&RightArrow;}{u}}^1\&Element;E^M,$ Here might as well get ${\stackrel{\&RightArrow;}{u}}^1=(0,...,0)\&Element;E^M,$ Get H ₁∈ E ^{M * M}For any nonsingular symmetrical matrix, might as well get here $H_1^T=({\mathrm{<figure-callout\; id="1"\; label="e"\; filenames="C20051011579400181.png"\; state="\{\{state\}\}">e</figure-callout>}}_1,...,e_M),$ Put i=1, iflag=0;

(3) calculate I component of place's residual vector, ${\mathrm{\&tau;}}_i={\stackrel{\&RightArrow;}{a}}_i^T{\stackrel{\&RightArrow;}{u}}^i-{\stackrel{\&RightArrow;}{b}}_i,$ Calculate the search vector $S^i=H_i{\stackrel{\&RightArrow;}{a}}_i;$

(4) if $S^i\&NotEqual;\stackrel{\&RightArrow;}{0},$ Change the step (5); If $S^i=\stackrel{\&RightArrow;}{0},$ Simultaneously ${\mathrm{\&tau;}}_i=\stackrel{\&RightArrow;}{0},$ Then put ${\stackrel{\&RightArrow;}{u}}^{i+1}={\stackrel{\&RightArrow;}{u}}^i,$ H _I+1=H _i, iflag=iflag+1 is if i＜M changes the step (7); Otherwise stop to calculate; At this moment

Separate for equational.If $S^i=\stackrel{\&RightArrow;}{0},$ But ${\mathrm{\&tau;}}^i\&NotEqual;\stackrel{\&RightArrow;}{0},$ Then put iflag=-i and stop to calculate, this moment, the system of equations formula was incompatible, also was that target function type does not have optimum solution;

(5) approximate value of modified solutions, ${\stackrel{\&RightArrow;}{u}}^{i+1}={\stackrel{\&RightArrow;}{u}}^i-{\mathrm{\&lambda;}}_i{S}^i,$ Step-length wherein ${\mathrm{\&lambda;}}_i={\mathrm{\&tau;}}_i/{\stackrel{\&RightArrow;}{a}}_i^T{S}^i,$

If i=M stops to calculate,

Separate for equational;

(6) correction matrix sheet H _i, $H_{i+1}=H_i-S^i{\left(S^i\right)}^T/{\stackrel{\&RightArrow;}{a}}_i^T{S}^i;$

(7) put i=i+1, change the step (3);

$\stackrel{\&RightArrow;}{u}=u^{M+1}$ Stop, obtaining equational separating this moment, also promptly obtained making J in the target function type to obtain the optimum solution of minimum value.Thereby

$\mathrm{\&Delta;}{u}_f\left(t\right)=[\mathrm{1,0},...,0]\stackrel{\&RightArrow;}{u}$

Then be in t controlled quentity controlled variable constantly

u(t)=T(Z ^-1)[u _f(t-1)+△u _f(t)]

Described initialization and find the solution optimal control law, the feedback compensation of model, be to utilize the Generalized Prediction Model Calculation optimal value of this sampling instant and the deviation real-time online correction model parameter between the set-point, this based on the generalized predictive control model, actual output according to system is constantly carried out the rolling optimization correction to the PREDICTIVE CONTROL output valve, and utilized feedback information, constitute closed-loop optimization and proofread and correct; When adopting, the feedback compensation parameter estimation becomes the adaptive system optimizing control of Generalized Prediction of two sections parameter estimation performance index of forgetting factor weighting, specific as follows:

Forecast model adopts

A(z ^-1)△y _f(t)＝B(z ^-1)△u _f(t-1))+D(z ^-1)△v _f(t-1)+C(z ^-1)ξ _f(t)

If theorem is A (z ^-1) and C (z ^-1) be stable, then the forecast model formula can replace with following high-order CAR model approximation

${\mathrm{\&xi;}}_f\left(t\right)=\underset{j=0}{\overset{n_0}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_j\mathrm{\&Delta;}{y}_f(t-j)-\underset{j=0}{\overset{n_0}{\mathrm{\&Sigma;}}}{\mathrm{\&beta;}}_j\mathrm{\&Delta;}{u}_f(t-j)-\underset{j=0}{\overset{n_0}{\mathrm{\&Sigma;}}}{\mathrm{\&gamma;}}_j\mathrm{\&Delta;}{v}_f(t-j)$

α wherein ₀=1, β ₀=γ ₀=0

It has the LS structure

Δy _f(t)＝Φ ₁ ^T(t-1)θ ₁+ξ _f(t)

Φ wherein ₁ ^T(t-1)=[Δ y _f(t-1) ... ,-Δ y _f(t-n ₀), Δ u _f(t-1) ..., Δ u _f(t-n ₀)

Δv _f(t-1)，…，-Δv _f(t-n ₀)]

θ ₁＝[α ₁，…，α _n0，β ₁，…，β _n0，γ ₁，…，γ _n0] ^T

So can get θ with following improved RLS algorithm ₁The LS valuation

${\widehat{\mathrm{\&theta;}}}_1\left(t\right)={\widehat{\mathrm{\&theta;}}}_1(t-1)+\frac{P(t-1){\mathrm{\&Phi;}}_1(t-1)[\mathrm{\&Delta;}{y}_f\left(t\right)-{{\mathrm{\&Phi;}}_1}^T(t-1){\widehat{\mathrm{\&theta;}}}_1(t-1)]}{\mathrm{\&beta;}\left(t\right)+{{\mathrm{\&Phi;}}_1}^T(t-1)P(t-1){\mathrm{\&Phi;}}_1(t-1)}$

$\mathrm{\&beta;}\left(t\right)=1-\frac{1}{N\left(t\right)}$

$N\left(t\right)=1+\frac{1+{\mathrm{\&Phi;}}_1(t-1)P(t-1){{\mathrm{\&Phi;}}_1}^T(t-1)}{{[\mathrm{\&Delta;}{y}_f\left(t\right)-{{\mathrm{\&Phi;}}_1}^T(t-1){\widehat{\mathrm{\&theta;}}}_1(t-1)]}^2}\mathrm{\&eta;}\left(t\right)$

Note $\mathrm{\&psi;}\left(t\right)=P(t-1)-\frac{P(t-1){\mathrm{\&Phi;}}_1(t-1){{\mathrm{\&Phi;}}_1}^T(t-1)P^T(t-1)}{1+{{\mathrm{\&Phi;}}_1}^T(t-1)P(t-1){\mathrm{\&Phi;}}_1(t-1)}$

If trace[is ψ (t)]/β (t)≤ρ

P (t)=ψ (t)/β (t) then

Otherwise P (t)=ψ (t)

Parameter η (t) is as follows with the selection of ρ

η(t)＝mη _σ(t)

ρ＞α ²

M is a memory span, generally can be taken as 1000 and more than, η _σBe the mean square deviation of η (t), α ²Be initial covariance matrix p (0)=α ²Value among the I (I is a unit matrix), $\widehat{\mathrm{\&theta;}}\left(0\right)=\mathrm{\&epsiv;}$ (α is big as far as possible number, and is general desirable more than 100, and ε is fully little real vector)

Generally get

$\left\{\begin{array}{c}\widehat{\mathrm{\&theta;}}\left(0\right)=\stackrel{\&RightArrow;}{0}\\ p\left(0\right)={\mathrm{\&alpha;}}^{2}I\end{array}\right.$

Trace[ψ (t)] be the mark of matrix ψ (t).

Thereby can get ξ _f(j) level and smooth valuation.

${\widehat{\mathrm{\&xi;}}}_f\left(j\right)=\mathrm{\&Delta;}{y}_f\left(j\right)-{{\mathrm{\&Phi;}}_1}^T\left(j\right){\widehat{\mathrm{\&theta;}}}_1(t+1),j=t-1,...,t-n_c$

${\widehat{\mathrm{\&sigma;}}}_{{\mathrm{\&xi;}}_f}^2\left(t\right)={\widehat{\mathrm{\&sigma;}}}_{{\mathrm{\&xi;}}_f}^2(t-1)+\frac{1}{t}[{{\widehat{\mathrm{\&xi;}}}_f}^2\left(t\right)-{\widehat{\mathrm{\&sigma;}}}_{{\mathrm{\&xi;}}_f}^2(t-1)]$

With level and smooth valuation

Substitution Φ ^T(t-1) in, obtain the parameter estimation that two sections parameter estimation methods of time spent change forgetting factor provide

The slack-off problem of parameter estimation when the time becomes two sections parameter estimation of forgetting factor and has solved controlled variable and noise tight coupling, algorithm increases or reduces memory span along with the variation adjustment forgetting factor of system dynamics, has improved the stability of system.

Described based on the online crucial parameter measurement method of generalized predictive control, utilize the generalized predictive control self-correcting algorithm of band filter exactly, calculate, export the crucial parameter measurement value, specific as follows:

(1) reads in parameter N, M, λ, P (z ^-1), Q (z ^-1), T (z ^-1) and parameter estimation algorithm in initial value α, η _σ(t), m unit matrix I.

(2)t＝0，P(t)＝α ²I， $\widehat{\mathrm{\&theta;}}\left(t\right)=0;$

(3) use formula

${\widehat{\mathrm{\&theta;}}}_1\left(t\right)={\widehat{\mathrm{\&theta;}}}_1(t-1)+\frac{P(t-1){{\mathrm{\&Phi;}}_1}^T(t-1)[\mathrm{\&Delta;}{y}_f\left(t\right)-{{\mathrm{\&Phi;}}_1}^T(t-1){\widehat{\mathrm{\&theta;}}}_1(t-1)]}{\mathrm{\&beta;}\left(t\right)+{\mathrm{\&Phi;}}_1(t-1)P(t-1){{\mathrm{\&Phi;}}_1}^T(t-1)}$

$\mathrm{\&beta;}\left(t\right)=1-\frac{1}{N\left(t\right)}$

$N\left(t\right)=1+\frac{1+{{\mathrm{\&Phi;}}_1}^T(t-1)P(t-1){\mathrm{\&Phi;}}_1(t-1)}{{[\mathrm{\&Delta;}{y}_f\left(t\right)-{{\mathrm{\&Phi;}}_1}^T(t-1){\widehat{\mathrm{\&theta;}}}_1(t-1)]}^2}\mathrm{\&eta;}\left(t\right)$

Note $\mathrm{\&psi;}\left(t\right)=P(t-1)-\frac{P(t-1){\mathrm{\&Phi;}}_1(t-1){{\mathrm{\&Phi;}}_1}^T(t-1)P(t-1)}{1+{{\mathrm{\&Phi;}}_1}^T(t-1)P(t-1){\mathrm{\&Phi;}}_1(t-1)}$

If trace[is ψ (t)]/β (t)≤ρ

P (t)=ψ (t)/β (t) then

Otherwise P (t)=ψ (t)

Parameter η (t) is the same with the selection of ρ, and initial value P (0) reaches

Select method the same, thereby can get ξ _f(j) level and smooth valuation.

${\widehat{\mathrm{\&xi;}}}_f\left(j\right)=\mathrm{\&Delta;}{y}_f\left(j\right)-{{\mathrm{\&Phi;}}_1}^T\left(j\right){\widehat{\mathrm{\&theta;}}}_1(t+1)$

j＝t-1，…，t-n _c

Provide

(4) will

The substitution formula

Φ ^T(t-1)＝[-Δy _f(t-1)，…，-Δy _f(t-n _a)，Δu _f(t-1)，Δu _f(t-2)，…，

Δ u _f(t-n _b-1), Δ v _f(t-1) ..., Δ v _f(t-n _d-1),

Φ ^T(t-1) in.

Utilize formula

$N\left(t\right)=1+\frac{1+{\mathrm{\&Phi;}}^T(t-1)p(t-1)\mathrm{\&Phi;}(t-1)}{{[\mathrm{\&Delta;}{y}_f\left(t\right)-{\mathrm{\&Phi;}}^T(t-1)\widehat{\mathrm{\&theta;}}(t-1)]}^2}\mathrm{\&eta;}\left(t\right)$

$\mathrm{\&beta;}\left(t\right)=1-\frac{1}{N\left(t\right)}$

$\widehat{\mathrm{\&theta;}}\left(t\right)=\widehat{\mathrm{\&theta;}}(t-1)+\frac{p(t-1)\mathrm{\&Phi;}(t-1)[\mathrm{\&Delta;}{y}_f\left(t\right)-{\mathrm{\&Phi;}}^T(t-1)\widehat{\mathrm{\&theta;}}(t-1)]}{\mathrm{\&beta;}\left(t\right)+{\mathrm{\&Phi;}}^T(t-1)p(t-1)\mathrm{\&Phi;}(t-1)}$

Introduce $\mathrm{\&psi;}\left(t\right)=p(t-1)-\frac{p(t-1)\mathrm{\&Phi;}(t-1){\mathrm{\&Phi;}}^T(t-1)p(t-1)}{1+{\mathrm{\&Phi;}}^T(t-1)p(t-1)\mathrm{\&Phi;}(t-1)}$

If trace[is ψ (t)]/β (t)≤ρ

P (t)=ψ (t)/β (t) then

Estimated parameter

(5) utilize the recursion optimized Algorithm to obtain optimal value

(6) utilize formula

$\mathrm{\&Delta;}{u}_f\left(t\right)=[\mathrm{1,0},...,0]\stackrel{\&RightArrow;}{u}$

Then be in t controlled quentity controlled variable constantly:

u(t)＝T(Z ^-1)[u _f(t-1)+Δu _f(t)]

Calculate t controlled quentity controlled variable (blow rate required) u (t) constantly, u (t) as history data store;

Utilize the t value of u (t) value calculating measured (flue gas oxygen content) y (t) constantly;

The soft measured value of output flue gas oxygen content;

(7) put t=t+1, return the step (2).

The invention has the beneficial effects as follows: the present invention organically combines according to the advantage separately of generalized predictive control and soft-measuring technique, to give full play to advantage separately; The real-time online measuring model that proposes has well utilized the thought in the generalized predictive control, prolong when realizing some or be difficult to real-time measurement with the key parameters of common sensor measurement, provide fundamental basis for online crucial parameter measurement method, also further guaranteed the application of in real process, succeeding of on-line parameter measuring technique; The oxygen level soft-measuring technique that the present invention proposes based on Generalized Prediction be in the boiler combustion process other key parameters such as flue-gas temperature accurately, in real time, online soft sensor opened up new approach, to the closed-loop control that realizes combustion system with optimize operation and have great importance; Realize industrial process key parameters real-time online measuring, solved the adaptive problem of measuring in the boiler control with control of time domain.

Description of drawings

The present invention is further described below in conjunction with drawings and Examples;

Fig. 1 is flue gas oxygen content observing and controlling figure;

Fig. 2 is a model tuning parameter estimation method block diagram;

Fig. 3 is a Generalized Prediction self-correcting process of measurement process flow diagram;

Fig. 4 is novel algorithm oxygen level test emulation figure;

Fig. 5 is traditional algorithm oxygen level test emulation figure.

Specific embodiments

Embodiment: based on the On-line Measuring Method of key parameter in the boiler combustion process of novel broad sense PREDICTIVE CONTROL.

1, the foundation of flue gas oxygen content observing and controlling model

Keeping the optimum condition and the economy of boiler combustion process is the boiler combustion process vital task of control automatically.In boiler operatiopn, must control the air capacity of boiler and the ratio of fuel quantity of entering well.If air capacity is relative less with the ratio of fuel quantity, then chemical imperfect combustion loss is big, otherwise then heat loss due to exhaust gas increases.In order to make boiler keep the best combustion operating mode, must make the ratio of air capacity and quantity combusted suitable, this ratio is called excess air coefficient, and numerical value should be between 1.20～1.30.Oxygen level and excess air coefficient have the funtcional relationship of definite (monodrome) in the flue gas

$\mathrm{\&alpha;}=\frac{{\mathrm{\&beta;}}_0}{{\mathrm{\&beta;}}_0-\mathrm{\&beta;}}$

β in the formula ₀=20.9%, be the volume fraction of oxygen in the normal air.

The oxygen content of flue gas is the key parameters of boiler control, and its characterizes combustion conditions quality, also is the control air capacity that enters burner hearth, keeps best coal-air ratio, reaches the fundamental basis of optimized combustion.

In the various boiler combustion process of computer optimization, come the regulating and controlling air quantity according to flue gas oxygen content usually, simultaneously with reference to fuel metering amount m, noise interferences, in order to control coefficient of excess air α or best oxygen content β, flue gas oxygen content observing and controlling is as shown in Figure 1.

Provide the flue gas oxygen content observing and controlling model that makes up based on Generalized Prediction CARIMA forecast model below:

$\stackrel{\&RightArrow;}{A}\left(z^{-1}\right)y\left(t\right)=B\left(z^{-1}\right)\mathrm{\&Delta;u}(t-1)+D\left(z^{-1}\right)\mathrm{\&Delta;v}(t-1)+C\left(z^{-1}\right)\mathrm{\&xi;}\left(t\right)$

Y, u, v, ξ represent that respectively output quantity (measured-flue gas oxygen content), input quantity (controlled quentity controlled variable-blow rate required), feedforward (can survey disturbance-coal input quantity) and normal white noise disturb in the formula.

Introduce auxiliary output

It is as follows to consider that Diophantine (Diophantine equation) provides objective function:

$J=E\{{(\mathrm{\&psi;}-w)}^T(\mathrm{\&psi;}-w)+\mathrm{\&lambda;}{\mathrm{\&psi;}}^T\mathrm{\&psi;}\}$

$=E\{{(G\stackrel{\&RightArrow;}{u}+\stackrel{\&RightArrow;}{f}+\stackrel{\&RightArrow;}{e}-\stackrel{\&RightArrow;}{w})}^T(G\stackrel{\&RightArrow;}{u}+\stackrel{\&RightArrow;}{f}+\stackrel{\&RightArrow;}{e}-\stackrel{\&RightArrow;}{w})$

$+\mathrm{\&lambda;}{(P\stackrel{\&RightArrow;}{u}+\stackrel{\&RightArrow;}{\mathrm{\&delta;}})}^T(P\stackrel{\&RightArrow;}{u}+\stackrel{\&RightArrow;}{\mathrm{\&delta;}})\}$

2, the design of Generalized Prediction soft instrument

The design of Generalized Prediction soft instrument is as follows, and parameter identification design and soft instrument are measured shown in Fig. 2,3.

(1) reads in parameter N=20, M=20, λ=12, P (z ^-1), Q (z ^-1), T (z ^-1) and parameter estimation algorithm in initial value α=100, m=100 unit matrix I, η _σ(t).

(2)t＝0，P(t)＝α ²I， $\widehat{\mathrm{\&theta;}}\left(t\right)=0;$

(3) use formula

${\widehat{\mathrm{\&theta;}}}_1\left(t\right)={\widehat{\mathrm{\&theta;}}}_1(t-1)+\frac{P(t-1){{\mathrm{\&Phi;}}_1}^T(t-1)[\mathrm{\&Delta;}{y}_f\left(t\right)-{{\mathrm{\&Phi;}}_1}^T(t-1){\widehat{\mathrm{\&theta;}}}_1(t-1)]}{\mathrm{\&beta;}\left(t\right)+{\mathrm{\&Phi;}}_1(t-1)P(t-1){{\mathrm{\&Phi;}}_1}^T(t-1)}$

$\mathrm{\&beta;}\left(t\right)=1-\frac{1}{N\left(t\right)}$

$N\left(t\right)=1+\frac{1+{{\mathrm{\&Phi;}}_1}^T(t-1)P(t-1){\mathrm{\&Phi;}}_1(t-1)}{{[\mathrm{\&Delta;}{y}_f\left(t\right)-{{\mathrm{\&Phi;}}_1}^T(t-1){\widehat{\mathrm{\&theta;}}}_1]}^2}\mathrm{\&eta;}\left(t\right)$

Note $\mathrm{\&psi;}\left(t\right)=P(t-1)-\frac{P(t-1){\mathrm{\&Phi;}}_1(t-1){{\mathrm{\&Phi;}}_1}^T(t-1)P(t-1)}{1+{{\mathrm{\&Phi;}}_1}^T(t-1)P(t-1){\mathrm{\&Phi;}}_1(t-1)}$

If trace[is ψ (t)]/β (t)≤ρ

P (t)=ψ (t)/β (t) then

Otherwise P (t)=ψ (t)

Parameter η (t) is the same with the selection of ρ, and initial value P (0) reaches

Select method the same, thereby can get ξ _f(j) level and smooth valuation.

${\widehat{\mathrm{\&xi;}}}_f\left(j\right)=\mathrm{\&Delta;}{y}_f\left(j\right)-{{\mathrm{\&Phi;}}_1}^T\left(j\right){\widehat{\mathrm{\&theta;}}}_1(t+1),j=t-1,...,t-n_c$

Provide

(4) will

The substitution formula

Φ ^T(t-1)＝[-Δy _f(t-1)，…，-Δy _f(t-n _a)，Δu _f(t-1)，Δu _f(t-2)，…，

Δ u _f(t-n _b-1), Δ v _f(t-1) ..., Δ v _f(t-n _d-1),

Φ ^T(t-1) in, utilize formula

$N\left(t\right)=1+\frac{1+{\mathrm{\&Phi;}}^T(t-1)p(t-1)\mathrm{\&Phi;}(t-1)}{{[\mathrm{\&Delta;}{y}_f\left(t\right)-{\mathrm{\&Phi;}}^T(t-1)\widehat{\mathrm{\&theta;}}(t-1)]}^2}\mathrm{\&eta;}\left(t\right)$

$\mathrm{\&beta;}\left(t\right)=1-\frac{1}{N\left(t\right)}$

$\widehat{\mathrm{\&theta;}}\left(t\right)=\widehat{\mathrm{\&theta;}}(t-1)+\frac{p(t-1)\mathrm{\&Phi;}(t-1)[\mathrm{\&Delta;}{y}_f\left(t\right)-{\mathrm{\&Phi;}}^T(t-1)\widehat{\mathrm{\&theta;}}(t-1)]}{\mathrm{\&beta;}\left(t\right)+{\mathrm{\&Phi;}}^T(t-1)p(t-1)\mathrm{\&Phi;}(t-1)}$

Introduce $\mathrm{\&psi;}\left(t\right)=p(t-1)-\frac{p(t-1)\mathrm{\&Phi;}(t-1){\mathrm{\&Phi;}}^T(t-1)p(t-1)}{1+{\mathrm{\&Phi;}}^T(t-1)p(t-1)\mathrm{\&Phi;}(t-1)}$

If trace[is ψ (t)]/β (t)≤ρ

P (t)=ψ (t)/β (t) then

Estimated parameter

(5) utilize the recursion optimized Algorithm to obtain optimal value

Note $G^{T}G+\mathrm{\&lambda;}{P}^{T}P=\left[\begin{array}{cccc}{a}_{11}& a_{12}& \&CenterDot;\&CenterDot;\&CenterDot;& a_{1M}\\ a_{21}& a_{22}& \&CenterDot;\&CenterDot;\&CenterDot;& a_{2M}\\ \&CenterDot;\&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;\&CenterDot;\\ a_{M1}& a_{M2}& \&CenterDot;\&CenterDot;\&CenterDot;& a_{\mathrm{MM}}\end{array}\right]=\left[\begin{array}{c}{{\stackrel{\&RightArrow;}{a}}_1}^t\\ {{\stackrel{\&RightArrow;}{a}}_2}^t\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ {{\stackrel{\&RightArrow;}{a}}_M}^t\end{array}\right]$

$G^T(\stackrel{\&RightArrow;}{w}-\stackrel{\&RightArrow;}{f})-\mathrm{\&lambda;}{P}^T\stackrel{\&RightArrow;}{\mathrm{\&delta;}}={\stackrel{\&RightArrow;}{b}}^t=[{\stackrel{\&RightArrow;}{b}}_1,...,{\stackrel{\&RightArrow;}{b}}_M]$

1. read in M=20, t=1, ${\stackrel{\&RightArrow;}{a}}_i^t=(a_{\mathrm{<figure-callout\; id="1"\; label="i"\; filenames="C20051011579400181.png"\; state="\{\{state\}\}">i</figure-callout>}1},...,a_{\mathrm{iM}}),i=\mathrm{1,2},...,M$

e ₁＝(1，0，…，0)，…，e _j＝(0，…，0，1，0，…，0)，…，e _M＝(0，0，…，1)， ${\stackrel{\&RightArrow;}{b}}^t=[{\stackrel{\&RightArrow;}{b}}_1,...,{\stackrel{\&RightArrow;}{b}}_M],$ ${\stackrel{\&RightArrow;}{u}}^T=({\mathrm{<figure-callout\; id="1"\; label="u"\; filenames="C20051011579400181.png"\; state="\{\{state\}\}">u</figure-callout>}}_1,...,u_M),$ u _i＝△u _f(t+i-1)，i＝1，2，…，M

2. appoint and get ${\stackrel{\&RightArrow;}{u}}^1\&Element;E^M,$ Here might as well get ${\stackrel{\&RightArrow;}{u}}^1=(0,...,0)\&Element;E^M,$ Get H ₁∈ E ^{M * M}For any nonsingular symmetrical matrix, might as well get here $H_1^T=({\mathrm{<figure-callout\; id="1"\; label="e"\; filenames="C20051011579400181.png"\; state="\{\{state\}\}">e</figure-callout>}}_1,\&CenterDot;\&CenterDot;\&CenterDot;e_M),$ Put i=1, iflag=0;

3. calculate

I component of place's residual vector, ${\mathrm{\&tau;}}_i={\stackrel{\&RightArrow;}{a}}_i^T{\stackrel{\&RightArrow;}{u}}^i-{\stackrel{\&RightArrow;}{b}}_i,$ Calculate the search vector $S^i=H_i{\stackrel{\&RightArrow;}{a}}_i;$

4. if $S^i\&NotEqual;\stackrel{\&RightArrow;}{0},$ Change the step 5.; If $S^i=\stackrel{\&RightArrow;}{0},$ Simultaneously ${\mathrm{\&tau;}}_i=\stackrel{\&RightArrow;}{0},$ Then put ${\stackrel{\&RightArrow;}{u}}^{i+1}={\stackrel{\&RightArrow;}{u}}^i,$ H _I+1=H _i, iflag=iflag+1 is if 7. i＜M changes the step; Otherwise stop to calculate; At this moment

Separate for equational.If $S^i=\stackrel{\&RightArrow;}{0},$ But ${\mathrm{\&tau;}}^i\&NotEqual;\stackrel{\&RightArrow;}{0},$ Then put iflag=-i and stop to calculate, this moment, the system of equations formula was incompatible, also was that target function type does not have optimum solution;

5. the approximate value of modified solutions, ${\stackrel{\&RightArrow;}{u}}^{i+1}={\stackrel{\&RightArrow;}{u}}^i-{\mathrm{\&lambda;}}_i{S}^i,$ Step-length wherein ${\mathrm{\&lambda;}}_i={\mathrm{\&tau;}}_i/{\stackrel{\&RightArrow;}{a}}_i^T{S}^i,$

If i=M stops to calculate,

Separate for equational;

6. correction matrix H _i, $H_{i+1}=H_i-S^i{\left(S^i\right)}^T/{\stackrel{\&RightArrow;}{a}}_i^T{S}^i;$

7. put i=i+1, change the step 3.;

$\stackrel{\&RightArrow;}{u}=u^{M+1}$ Stop, obtaining equational separating this moment, also promptly obtained making J in the target function type to obtain the optimum solution of minimum value.

(6) utilize formula $\mathrm{\&Delta;}{u}_f\left(t\right)=[\mathrm{1,0},...,0]\stackrel{\&RightArrow;}{u}$

Be u (t)=T (Z then in t controlled quentity controlled variable constantly ^-1) [u _f(t-1)+△ u _f(t)]

Calculate t controlled quentity controlled variable (blow rate required) u (t) constantly, u (t) as history data store; Utilize the t value of u (t) value calculating measured (flue gas oxygen content) y (t) constantly; The soft measured value of output flue gas oxygen content:

The flue gas oxygen content control interval is 33%-4.0%.

(7) put t=t+1, return the step (2).

3, simulation result

Application is measured flue gas oxygen content based on the soft instrument of novel broad sense prediction algorithm and traditional Generalized Prediction algorithm, situation such as Fig. 4, shown in Figure 5.The result who studies under same computer environment is as follows: M got 20 o'clock, and the computing time of new and old algorithm is 0.02s and 0.209s respectively.Obviously, the dimension of system is big more, predicts that promptly time domain and control time domain strengthen, and the time that new algorithm is saved is many more, and stability is better.Around set-point, traditional algorithm shown in Figure 5 is measured flue gas oxygen content obviously hysteresis, lag behind about about 100 steps, and fluctuation range is bigger, and simulation result shows that the soft instrument that utilizes the novel broad sense prediction algorithm to make up has quick tracking and exports stable advantage.

If adopt zirconia sensor to measure, system lags behind generally near more than the 60s so.When measuring in 0～60s, because the change of controlled quentity controlled variable (air quantity) and the disturbance of coal supply, and the time lag characteristics of stove combustion process, make the zirconia measured value can not truly reflect stove internal combustion situation, implement control according to the zirconia measured value, for step disturbance, overshoot is up to 50%～300%, and time lag is for up to more than the 60s; And it is very little to adopt the Generalized Prediction flexible measurement method to record the data overshoot, can reflect the actual conditions that oxygen level changes.When measuring in 60s～120s, because the stove internal combustion has reached steady state (SS), the measured value of soft or hard instrument is close, and its error is negligible.

Therefore, the soft instrument that utilization novel broad sense predictive control algorithm makes up can objectively respond the truth of any period in the boiler combustion process, has improved the measuring accuracy and the real-time of flue gas oxygen content, has optimized the performance of boiler combustion control system.Along with the development of generalized predictive control theory and the improvement of prediction algorithm, measuring accuracy also can further improve, new means are provided based on the oxygen level soft-measuring technique of Generalized Prediction for the oxygen content measurement of Industrial Boiler, also for other key parameters in the boiler combustion process such as flue-gas temperature accurately, in real time, online soft sensor opened up new approach, to the closed-loop control that realizes combustion system with to optimize operation significant.

The present invention is equally applicable to the measurement of flue gas oxygen content, CO flue gas content in the production runes such as blank heating furnace, Gas Production stove.

Claims (5)

1. online crucial parameter measurement method based on the novel broad sense PREDICTIVE CONTROL, adopt this method to realize the real-time online measuring of flue gas oxygen content and temperature, it is characterized in that: according to the measuring method of the time lag characteristic and the key parameter of process control, utilize the novel broad sense predictive controller, set up soft measurement forecast model based on the Generalized Prediction algorithm, realized the real-time online measuring of process control key parameters, specifically comprised and set up measurement model, initialization and find the solution optimal control law, model feedback compensation and based on four steps of the online crucial parameter measurement method of generalized predictive control:

(1) sets up measurement model: set up measurement model according to generalized predictive control technical characterictic and process control time lag characteristic;

(2) initialization and find the solution optimal control law: set controlled variable according to measurement model, utilize the novel broad sense forecast Control Algorithm to find the solution optimal control law;

(3) model feedback compensation: become two sections method for parameter estimation of forgetting factor during employing and carry out the model feedback compensation, the optimal control parameter;

(4) based on the online crucial parameter measurement method of generalized predictive control: calculate, export the crucial parameter measurement value according to the novel broad sense predictive controller.

2. the online crucial parameter measurement method based on the novel broad sense PREDICTIVE CONTROL according to claim 1 is characterized in that: the measurement model of described foundation, specific as follows:

The objective function of determining the process control of system is:

$=E\{{(G\stackrel{\&RightArrow;}{u}+\stackrel{\&RightArrow;}{f}+\stackrel{\&RightArrow;}{e}-\stackrel{\&RightArrow;}{w})}^T(G\stackrel{\&RightArrow;}{u}+\stackrel{\&RightArrow;}{f}+\stackrel{\&RightArrow;}{e}-\stackrel{\&RightArrow;}{w})+\mathrm{\&lambda;}{(P\stackrel{\&RightArrow;}{u}+\stackrel{\&RightArrow;}{\mathrm{\&delta;}})}^T(P\stackrel{\&RightArrow;}{u}+\stackrel{\&RightArrow;}{\mathrm{\&delta;}})\}$

In order to try to achieve optimal control law

, can get by target function type

$(G^{T}G+{\mathrm{\&lambda;P}}^{T}P)\stackrel{\&RightArrow;}{u}=G^T(\stackrel{\&RightArrow;}{w}-\stackrel{\&RightArrow;}{f}-\stackrel{\&RightArrow;}{e})-{\mathrm{\&lambda;P}}^T\stackrel{\&RightArrow;}{\mathrm{\&delta;}}$

Try to achieve by equation as can be known by mathematical analysis Be the J that makes in the target function type and obtain the optimum solution of minimum value.

3. the online crucial parameter measurement method based on the novel broad sense PREDICTIVE CONTROL according to claim 1 is characterized in that: initialization and find the solution optimal control law, specific as follows:

Because novel control algolithm adopts matrix recursion alternative manner, avoided classic method to find the solution the calculating of inverse matrix in the optimal control law, the descended square order of matrix exponent number M of computational complexity, calculated amount of saving and memory space just can enlarge the scope of choosing of prediction time domain and control time domain like this, increase the stability of TT﹠C system, can realize the soft measurement of boiler key parameter real-time online;

The recursion optimized Algorithm is as follows:

Note $G^{T}G+\mathrm{\&lambda;}{P}^{T}P=\left[\begin{array}{cccc}{a}_{11}& a_{12}& \&CenterDot;\&CenterDot;\&CenterDot;& a_{1M}\\ a_{21}& a_{22}& \&CenterDot;\&CenterDot;\&CenterDot;& a_{2M}\\ \&CenterDot;\&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;\&CenterDot;\\ a_{M1}& a_{M2}& \&CenterDot;\&CenterDot;\&CenterDot;& a_{\mathrm{MM}}\end{array}\right]=\left[\begin{array}{c}{{\stackrel{\&RightArrow;}{a}}_1}^t\\ {{\stackrel{\&RightArrow;}{a}}_2}^t\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ {{\stackrel{\&RightArrow;}{a}}_M}^t\end{array}\right]$

$G^T(\stackrel{\&RightArrow;}{w}-\stackrel{\&RightArrow;}{f})-{\mathrm{\&lambda;P}}^T\stackrel{\&RightArrow;}{\mathrm{\&delta;}}={\stackrel{\&RightArrow;}{b}}^t=[{\stackrel{\&RightArrow;}{b}}_1,\&CenterDot;\&CenterDot;\&CenterDot;,{\stackrel{\&RightArrow;}{b}}_M]$

(1) read in M, t, ${\stackrel{\&RightArrow;}{a}}_i^t=(a_{i1},\&CenterDot;\&CenterDot;\&CenterDot;,a_{\mathrm{iM}}),$ i＝1，2，…，M

e ₁＝(1，0，…，0)，…，e _j＝(0，…，0，1，0，…，0)，…，e _M＝(0，0，…，1)， ${\stackrel{\&RightArrow;}{b}}^t=[{\stackrel{\&RightArrow;}{b}}_1,\&CenterDot;\&CenterDot;\&CenterDot;,{\stackrel{\&RightArrow;}{b}}_M],$ ${\stackrel{\&RightArrow;}{u}}^T=(u_i,\&CenterDot;\&CenterDot;\&CenterDot;,u_M),$ u _i＝Δu _f(t+i-1)，i＝1，2，…，M；

(2) appoint and get ${\stackrel{\&RightArrow;}{u}}^1\&Element;E^M,$ Here get ${\stackrel{\&RightArrow;}{u}}^1=(0,\&CenterDot;\&CenterDot;\&CenterDot;,0)\&Element;E^M,$ Get H ₁∈ E ^{M * M}For any nonsingular symmetrical matrix, get here $H_1^T=(e_1,\&CenterDot;\&CenterDot;\&CenterDot;,e_M),$ Put i=1, iflag=0;

(3) calculate

I component of place's residual vector, ${\mathrm{\&tau;}}_i={\stackrel{\&RightArrow;}{a}}_i^T{\stackrel{\&RightArrow;}{u}}^i-{\stackrel{\&RightArrow;}{b}}_i,$ Calculate the search vector $S^i=H_i{\stackrel{\&RightArrow;}{a}}_i;$

(4) if $S^i\&NotEqual;\stackrel{\&RightArrow;}{0},$ Change the step (5); If $S^i=\stackrel{\&RightArrow;}{0},$ Simultaneously ${\mathrm{\&tau;}}_i=\stackrel{\&RightArrow;}{0},$ Then put ${\stackrel{\&RightArrow;}{u}}^{i+1}={\stackrel{\&RightArrow;}{u}}^i,$ H _I+1+ H _i, iflag=iflag+1 is if i＜M changes the step (7); Otherwise stop to calculate; At this moment

Separate for equational; If $S^i=\stackrel{\&RightArrow;}{0},$ But ${\mathrm{\&tau;}}^i\&NotEqual;\stackrel{\&RightArrow;}{0},$ Then put iflag=-i and stop to calculate, this moment, the system of equations formula was incompatible, also was that target function type does not have optimum solution;

(5) approximate value of modified solutions, ${\stackrel{\&RightArrow;}{u}}^{i+1}={\stackrel{\&RightArrow;}{u}}^i-{\mathrm{\&lambda;}}_i{S}^i,$ Step-length wherein ${\mathrm{\&lambda;}}_i={\mathrm{\&tau;}}_i/{\stackrel{\&RightArrow;}{a}}_i^T{S}^i,$

If i=M stops to calculate,

Separate for equational;

(6) correction matrix H _i, $H_{i+1}=H_i-S^i{\left(S^i\right)}^T/{\stackrel{\&RightArrow;}{a}}_i^T{S}^i;$

(7) put i=i+1, change the step (3);

$\stackrel{\&RightArrow;}{u}=u^{M+1}$ Stop, obtaining equational separating this moment, also promptly obtained making J in the target function type to obtain the optimum solution of minimum value, thereby

${\mathrm{\&Delta;u}}_f\left(t\right)=[\mathrm{1,0},\&CenterDot;\&CenterDot;\&CenterDot;,0]\stackrel{\&RightArrow;}{u}$

Then be in t controlled quentity controlled variable constantly

u(t)＝T(Z ^-1)[u _f(t-1)+Δu _f(t)]。

4. the online crucial parameter measurement method based on the novel broad sense PREDICTIVE CONTROL according to claim 1, it is characterized in that: initialization and find the solution optimal control law, the feedback compensation of model, be to utilize the Generalized Prediction Model Calculation optimal value of this sampling instant and the deviation real-time online correction model parameter between the set-point, this based on the generalized predictive control model, actual output according to system is constantly carried out the rolling optimization correction to the PREDICTIVE CONTROL output valve, and utilized feedback information, constituting closed-loop optimization proofreaies and correct, when adopting, the feedback compensation parameter estimation becomes the adaptive system optimizing control of Generalized Prediction of two sections parameter estimation performance index of forgetting factor weighting, specific as follows:

Forecast model adopts

A(z ^-1)Δy _f(t)＝B(z ^-1)Δu _f(t-1)+D(z ^-1)Δv _f(t ^-1)+C(z ^-1)ξ _f(t)

In the formula: Δ y _f(t), Δ u _f(t-1), Δ v _f(t-1), ξ _f(t) represent that respectively output quantity, input quantity, feedforward and normal white noise disturb;

If theorem is A (z ^-1) and C (z ^-1) be stable, then the forecast model formula can replace with following high-order CAR model approximation

${\mathrm{\&xi;}}_f\left(t\right)=\underset{j=0}{\overset{n_0}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_j\mathrm{\&Delta;}{y}_f(t-j)-\underset{j=0}{\overset{n_0}{\mathrm{\&Sigma;}}}{\mathrm{\&beta;}}_j\mathrm{\&Delta;}{u}_f(t-j)-\underset{j=0}{\overset{n_0}{\mathrm{\&Sigma;}}}{\mathrm{\&gamma;}}_j\mathrm{\&Delta;}{v}_f(t-j)$

α wherein ₀=1, β ₀=γ ₀=0

It has the LS structure

${\mathrm{\&Delta;y}}_f\left(t\right)={\mathrm{\&Phi;}}_1^T(t-1){\mathrm{\&theta;}}_1+{\mathrm{\&xi;}}_f\left(t\right)$

Wherein ${\mathrm{\&Phi;}}_1^T(t-1)=[-\mathrm{\&Delta;}{y}_f(t-1),\&CenterDot;\&CenterDot;\&CenterDot;,-{\mathrm{\&Delta;y}}_f(t-n_0),{\mathrm{\&Delta;u}}_f(t-1),\&CenterDot;\&CenterDot;\&CenterDot;,\mathrm{\&Delta;}{u}_f(t-n_0)$

${\mathrm{\&Delta;v}}_f(t-1),\&CenterDot;\&CenterDot;\&CenterDot;,-{\mathrm{\&Delta;v}}_f(t-n_0)]$

${\mathrm{\&theta;}}_1={[{\mathrm{\&alpha;}}_1,\&CenterDot;\&CenterDot;\&CenterDot;,{\mathrm{\&alpha;}}_{n_0},{\mathrm{\&beta;}}_1,\&CenterDot;\&CenterDot;\&CenterDot;,{\mathrm{\&beta;}}_{n_0},{\mathrm{\&gamma;}}_1,\&CenterDot;\&CenterDot;\&CenterDot;,{\mathrm{\&gamma;}}_{n_0}]}^T$

So can get θ with following improved RLS algorithm ₁The LS valuation

${\widehat{\mathrm{\&theta;}}}_1\left(t\right)={\widehat{\mathrm{\&theta;}}}_1(t-1)+\frac{P(t-1){\mathrm{\&Phi;}}_1(t-1)[{\mathrm{\&Delta;y}}_f\left(t\right)-{\mathrm{\&Phi;}}_1^T(t-1){\widehat{\mathrm{\&theta;}}}_1(t-1)]}{\mathrm{\&beta;}\left(t\right)+{\mathrm{\&Phi;}}_1^T(t-1)P(t-1){\mathrm{\&Phi;}}_1(t-1)}$

$\mathrm{\&beta;}\left(t\right)=1-\frac{1}{N\left(t\right)}$

$N\left(t\right)=1+\frac{1+{\mathrm{\&Phi;}}_1(t-1)P(t-1){\mathrm{\&Phi;}}_1^T(t-1)}{{[{\mathrm{\&Delta;y}}_f\left(t\right)-{\mathrm{\&Phi;}}_1^T(t-1){\widehat{\mathrm{\&theta;}}}_1(t-1)]}^2}\mathrm{\&eta;}\left(t\right)$

Note $\mathrm{\&psi;}\left(t\right)=P(t-1)-\frac{P(t-1){\mathrm{\&Phi;}}_1(t-1){\mathrm{\&Phi;}}_1^T(t-1)P^T(t-1)}{1+{\mathrm{\&Phi;}}_1^T(t-1)P(t-1){\mathrm{\&Phi;}}_1(t-1)}$

If trace[is ψ (t)]/β (t)≤ρ

P (t)=ψ (t)/β (t) then

Otherwise P (t)=ψ (t)

Parameter η (t) is as follows with the selection of ρ:

η(t)＝mη _σ(t)

ρ＞α ²

M is a memory span, generally can be taken as 1000 and more than, η _σBe the mean square deviation of η (t), α ²Be initial covariance matrix p (0)=α ²Value among the I, I is a unit matrix, $\widehat{\mathrm{\&theta;}}\left(0\right)=\mathrm{\&epsiv;},$ α is big as far as possible number, and is general desirable more than 100, and ε is fully little real vector;

Generally get $\left\{\begin{array}{c}\widehat{\mathrm{\&theta;}}\left(0\right)=\stackrel{\&RightArrow;}{0}\\ p\left(0\right)={\mathrm{\&alpha;}}^{2}I\end{array}\right.$

Trace[ψ (t)] be the mark of matrix ψ (t),

Thereby can get ξ _f(j) level and smooth valuation,

${\widehat{\mathrm{\&xi;}}}_f\left(j\right)={\mathrm{\&Delta;y}}_f\left(j\right)-{\mathrm{\&Phi;}}_1^T\left(j\right){\widehat{\mathrm{\&theta;}}}_1(t+1)$ j＝t-1，…，t-n _c

${\widehat{\mathrm{\&sigma;}}}_{{\mathrm{\&xi;}}_f}^2\left(t\right)={\widehat{\mathrm{\&sigma;}}}_{{\mathrm{\&xi;}}_f}^2(t-1)+\frac{1}{t}[{{\widehat{\mathrm{\&xi;}}}_f}^2\left(t\right)-{\widehat{\mathrm{\&sigma;}}}_{{\mathrm{\&xi;}}_f}^2(t-1)]$

With level and smooth valuation

Substitution Φ ^T(t-1) in, obtain the parameter estimation that two sections parameter estimation methods of time spent change forgetting factor provide

The slack-off problem of parameter estimation when the time becomes two sections parameter estimation of forgetting factor and has solved controlled variable and noise tight coupling, algorithm increases or reduces memory span along with the variation adjustment forgetting factor of system dynamics, has improved the stability of system.

5. the online crucial parameter measurement method based on the novel broad sense PREDICTIVE CONTROL according to claim 1, it is characterized in that: based on the online crucial parameter measurement method of generalized predictive control, utilize the generalized predictive control self-correcting algorithm of band filter exactly, calculate, export the crucial parameter measurement value, specific as follows:

(1) reads in parameter N, M, λ, P (z ^-1), Q (z ^-1), T (z ^-1) and parameter estimation algorithm in initial value α, η _σ(t), m unit matrix I;

(2)t＝0，P(t)＝α ²I， $\widehat{\mathrm{\&theta;}}\left(t\right)=0;$

(3) use formula

${\widehat{\mathrm{\&theta;}}}_1\left(t\right)={\widehat{\mathrm{\&theta;}}}_1(t-1)+\frac{P(t-1){\mathrm{\&Phi;}}_1^T(t-1)[{\mathrm{\&Delta;y}}_f\left(t\right)-{\mathrm{\&Phi;}}_1^T(t-1){\widehat{\mathrm{\&theta;}}}_1(t-1)]}{\mathrm{\&beta;}\left(t\right)+{\mathrm{\&Phi;}}_1(t-1)P(t-1){\mathrm{\&Phi;}}_1^T(t-1)}$

$\mathrm{\&beta;}\left(t\right)=1-\frac{1}{N\left(t\right)}$

$N\left(t\right)=1+\frac{1+{\mathrm{\&Phi;}}_1^T(t-1)P(t-1){\mathrm{\&Phi;}}_1(t-1)}{{[{\mathrm{\&Delta;y}}_f\left(t\right)-{\mathrm{\&Phi;}}_1^T(t-1){\widehat{\mathrm{\&theta;}}}_1(t-1)]}^2}\mathrm{\&eta;}\left(t\right)$

Note $\mathrm{\&psi;}\left(t\right)=P(t-1)-\frac{P(t-1){\mathrm{\&Phi;}}_1(t-1){\mathrm{\&Phi;}}_1^T(t-1)P(t-1)}{1+{\mathrm{\&Phi;}}_1^T(t-1)P(t-1){\mathrm{\&Phi;}}_1(t-1)}$

If trace[is ψ (t)]/β (t)≤ρ

P (t)=ψ (t)/β (t) then

Otherwise P (t)=ψ (t)

Parameter η (t) is the same with the selection of ρ, and initial value P (0) reaches

Select method the same, thereby can get ξ _f(j) level and smooth valuation,

${\widehat{\mathrm{\&xi;}}}_f\left(j\right)={\mathrm{\&Delta;y}}_f\left(j\right)-{\mathrm{\&Phi;}}_1^T\left(j\right){\widehat{\mathrm{\&theta;}}}_1(t+1)$

j＝t-1，…，t-n _c

Provide

(4) will

The substitution formula

Φ ^T(t-1)＝[-Δy _f(t-1)，…，-Δy _f(t-n _a)，Δu _f(t-1)，Δu _f(t-2)，…，

${\mathrm{\&Delta;u}}_f(t-n_b-1),{\mathrm{\&Delta;v}}_f(t-1),\&CenterDot;\&CenterDot;\&CenterDot;,{\mathrm{\&Delta;v}}_f(t-n_d-1),{\widehat{\mathrm{\&xi;}}}_f(t-1),\&CenterDot;\&CenterDot;\&CenterDot;,{\widehat{\mathrm{\&xi;}}}_f(t-n_c)]$

Φ ^T(t-1) in,

Utilize formula:

$N\left(t\right)=1+\frac{1+{\mathrm{\&Phi;}}^T(t-1)p(t-1)\mathrm{\&Phi;}(t-1)}{{[{\mathrm{\&Delta;y}}_f\left(t\right)-{\mathrm{\&Phi;}}^T(t-1)\widehat{\mathrm{\&theta;}}(t-1)]}^2}\mathrm{\&eta;}\left(t\right)$

$\mathrm{\&beta;}\left(t\right)=1-\frac{1}{N\left(t\right)}$

$\widehat{\mathrm{\&theta;}}\left(t\right)=\widehat{\mathrm{\&theta;}}(t-1)+\frac{p(t-1)\mathrm{\&Phi;}(t-1)[{\mathrm{\&Delta;y}}_f\left(t\right)-{\mathrm{\&Phi;}}^T(t-1)\widehat{\mathrm{\&theta;}}(t-1)]}{\mathrm{\&beta;}\left(t\right)+{\mathrm{\&Phi;}}^T(t-1)p(t-1)\mathrm{\&Phi;}(t-1)}$

Introduce $\mathrm{\&psi;}\left(t\right)=p(t-1)-\frac{p(t-1)\mathrm{\&Phi;}(t-1){\mathrm{\&Phi;}}^T(t-1)p(t-1)}{1+{\mathrm{\&Phi;}}^T(t-1)p(t-1)\mathrm{\&Phi;}(t-1)}$

If trace[is ψ (t)]/β (t)≤ρ

P (t)=ψ (t)/β (t) then

Estimated parameter

(5) utilize the recursion optimized Algorithm to obtain optimal value

(6) utilize formula

${\mathrm{\&Delta;u}}_f\left(t\right)=[\mathrm{1,0},\&CenterDot;\&CenterDot;\&CenterDot;0]\stackrel{\&RightArrow;}{u}$

Then be in t controlled quentity controlled variable constantly

u(t)＝T(Z ^-1)[u _f(t-1)+Δu _f(t)]

Calculating t controlled quentity controlled variable constantly is blow rate required u (t), u (t) as history data store;

Utilizing t u (t) value calculating constantly measured is the value of flue gas oxygen content y (t);

The soft measured value of output flue gas oxygen content;

(7) put t=t+1, return the step (2).

Images