CN104765955A - Online soft measurement method - Google Patents

Abstract

The invention provides an online soft measurement method. The method comprises the following steps: acquiring operating data as a sample sequence; building a CM-LSSVM-PLS online soft measurement predicting model to train the model; using a model prediction measured value, and dynamically updating the model if a prediction error does not meet the requirements. According to the online soft measurement method provided by the invention, the measured value can be predicted by a CM-LSSVM-PLS method, the prediction precision can be improved, and the modeling time can be reduced; furthermore, the prediction precision can be further improved by dynamically updating the model; meanwhile, the matrix inverse operation can be carried out by a simple matrix, an algorithm is relatively simple, a result can be quickly calculated, and the model can be ensured to be applicable to online monitoring requirements; in addition, the soft measurement system is driven by real-time data, the model can be continuously updated according to the real-time data, and the online soft measurement method has the advantages of being likely to acquire data, low in extra hardware investment, high in model prediction precision, strong in self-adaption and the like.

Description

A kind of online soft sensor method

Technical field

What the present invention relates to is a kind of heat power engineering and computer monitoring interleaving techniques field, specifically a kind of online soft sensor method, based on coal-burning boiler burner hearth exit gas temperature online soft sensor method under the full load of least square method supporting vector machine.

Background technology

Furnace outlet gas temperature is the flue-gas temperature at furnace outlet place, online soft sensor system is used for large-sized station boiler furnace outlet section temperature distribution measuring, it is the turning point of radiation heat transfer and convection heat transfer in boiler, be all a very important variable in boiler all the time, affect the safe and stable operation of boiler and overall efficiency.

The height of furnace outlet gas temperature directly affects the thermal efficiency of boiler.Furnace outlet gas temperature is too high, then the exhaust gas temperature of boiler raises, and heat loss due to exhaust gas strengthens, and then has influence on the overall thermal efficiency of boiler.Furnace outlet gas temperature is also unsuitable too low, if on the low side, flue gas medial temperature then in burner hearth must reduce, affect the after-flame of the stable burning of flame, thermal radiation effect so between flue gas and furnace wall cooling is deteriorated, and the Radiant exothermicity of water-cooling wall reduces, and flue-gas temperature is on the low side, heat transfer effect between the heating surfaces such as flue gas and follow-up superheater is also deteriorated, and then the overall thermal efficiency of boiler is reduced.

Meanwhile, furnace outlet gas temperature is also the key variables of a core in burner hearth fouling monitoring system, also has a great impact the fouling and slagging of heating surface.A large amount of lime-ash and dust is had in boiler combustion product.Lime-ash is in furnace wall cooling and the meeting slagging scorification of hyperthermia radiation heat-transfer surface surface; Dust can, along with high-temperature flue gas, along flue stream through each convection heat transfer face, make heat-transfer surface produce dust stratification.The process of Slag and accumulating ash and flue-gas temperature have larger relation.In radiation heat transfer regions such as burner hearths, lime-ash melting under high-temperature flue gas effect, the surface sticking to water-cooling wall forms lime-ash; In convection heating surface, flue-gas temperature height also can aggravate the adhesion of ash content to heating surface.Furnace outlet gas temperature is too high, lime-ash and ash content can be made more easily to stick to heating surface surface, increase the weight of the degree of fouling of heating surface slagging scorification.The heat transmission resistance of ash dirt is comparatively large, is attached on heating surface, boiler cannot be run by fully loaded, can only increase the coal-supplying amount of boiler, further increase furnace outlet gas temperature, thus form vicious cycle.

The safe operation of furnace outlet gas temperature to boiler also has a great impact.When furnace outlet gas temperature higher than setting value a lot of time, the caloric receptivity showed increased of superheater at different levels can be made, superheater tube wall temperature is caused to exceed setting value, increase the weight of high temperature corrosion and the Temperature Deviation of heating surface, the security incidents such as the explosion of superheater tube wall, coking can be caused time serious, seriously affect safely and steadily running of boiler.On the other hand, flue-gas temperature is too high, also can aggravate the high temperature corrosion of heating surface, and for the heating surface of Slag and accumulating ash, the situation of high temperature corrosion is more serious.Find according to research, under the effect of high-temperature flue gas, sticking to the materials such as the lime-ash on heating surface surface and dust stratification can there is complicated chemical reaction with heating surface, causes the high temperature corrosion of heating surface, therefore, composition and the surface temperature on the mainly heating surface surface of high temperature corrosion is affected.High temperature corrosion can make the thickness of heating surface from outer toward interior thinning, under permanent effect, and the serious safety affecting heating surface.

In prior art, usually comprise hard ware measure and hard measurement two different developing direction.

In hard ware measure, the general instrument such as thermocouple thermometer, burner hearth cigarette temperature probe being arranged on burner hearth the right and left that adopts is to measure furnace outlet gas temperature.But due to burner hearth cigarette temperature too high (generally at about 1000 DEG C), these surveying instruments are often difficult to obtain long stably measured data, dismounting is often needed to carry out the maintenance and repair of off-line.In recent years, in the coal-burning boiler that some are newly-built, the advanced surveying instruments such as acoustics pyrometer, optics pyrometer are used to measure furnace outlet gas temperature.Through finding existing literature search, patent documentation " boiler furnace outlet smoke temperature on-line monitoring system " (notification number CN201532256U), disclose a kind of boiler furnace outlet smoke temperature on-line monitoring system, be included near boiler smokestack and arrange that target is for temperature-sensitive, near infrared imaging instrument is installed for gathering target emittance at horizontal flue, and installation data computing machine processes the data that near infrared imaging instrument gathers, and carries out the drafting of section temperature distribution curve." the boiler furnace fouling monitoring method based on acoustic thermometry and least square method supporting vector machine " in " power science and engineering " describes the application of acoustic thermometry system and least square method supporting vector machine, and this acoustic thermometry system and device mainly comprises the input-output device and temperature display devices etc. of sound wave transceiver, acoustic waveguide tube, signal processor, power amplifier, temperature.The monitor and feedback of this kind of method mainly depends on pneumatic noise near boiler horizontal gas pass and sound emission, and will to overcome numerous interference be very difficult.The surveying instrument of these advanced persons often expensive, installation and maintenance costly, traditional coal fired power plant does not have condition to carry out changing and add these surveying instruments.

Hard measurement aspect, through finding existing literature search, document " reckoning of boiler furnace outlet smoke temperature " adopts calculation of thermodynamics, by the thermodynamic equilibrium of each heat-transfer surface of boiler flue, calculates furnace outlet gas temperature.The people such as the Zhou Huaichun of the Central China University of Science and Technology utilize the image procossing of furnace flame, characterize the distribution situation in temperature field, carry out modeling to furnace outlet gas temperature.These methods are all often the boilers for a certain type, and mostly rely on experience, do not have generalization and real-time.

Summary of the invention

For the defect existed in prior art, the invention provides a kind of online soft sensor system, prediction coal-burning boiler burner hearth exit gas temperature, overcome the shortcoming such as precision, time in existing furnace outlet gas temperature soft-measuring technique, under load wide variation, on-line continuous can monitor furnace outlet section medial temperature.

Technical solution of the present invention is as follows: first, a kind of new C mean cluster-least square method supporting vector machine-offset minimum binary (C Means-Least Support Vector Machine-Partial Least Squares, CM-LSSVM-PLS) method is proposed to predict furnace outlet gas temperature.

The main thought of the method is that the modeling problem of complexity is resolved into multiple simple subproblem, mainly in three steps: the first step, uses C means clustering method that original data are divided into several subdata class; Second step, applies LSSVM method respectively to each subdata class and carries out modeling; Finally, apply deflected secondary air and each submodel is synthesized a final mask.Relative to modeling methods such as traditional LSSVM, the method has obvious advantage at model accuracy with on the modeling time.

Then, in actual industrial process, along with the time constantly changes during the state of industrial system, especially operating mode situation along with the time significantly change time, this requires that the model set up has online updating ability.And the process of online updating must be quick, model accuracy wants high, can the dynamic perfromance of tracing system.Therefore, the CM-LSSVM-PLS model of foundation needs efficient online updating.

In the present invention, for the CM-LSSVM-PLS model set up, propose a kind of online updating algorithm and be mainly divided into two steps: first, judge the precision of prediction of initial model, if precision meets the demands, then model is still suitable for, and does not upgrade; If do not met the demands, then carry out model modification, old data are removed, and new data are replaced, Renewal model parameter.This update algorithm is derived by matrix theory and is drawn, replaced the matrix inversion operation of original complexity by simple matrix plus-minus, algorithm is relatively simple, can calculate result within the rational time, and then ensures that model adapts to the requirement of on-line monitoring.Finally, watch-dog is connected with computing machine and shows flue gas temperature of hearth outlet data message in real time.

The present invention can obtain the real-time online value of furnace outlet gas temperature, after computing machine obtains monitoring result, be presented on monitored picture, or result is uploaded to DCS by microcomputer interface by it, for the concrete operations of operating personnel or unit cooperative control to provide data basis.

A kind of online soft sensor method of the present invention, comprises the following steps:

(1) N group service data is obtained, as N group sample sequence

(2) according to described sample sequence set up CM-LSSVM-PLS online soft sensor forecast model, and use described sample sequence to train described CM-LSSVM-PLS online soft sensor forecast model;

(3) by new sample (x _j,y _j) in x _jinput described CM-LSSVM-PLS online soft sensor forecast model, obtain the measured value predicted computational prediction error if predicated error ε is greater than prediction error threshold ε ₀, dynamically update described CM-LSSVM-PLS online soft sensor forecast model.

Further, the method setting up CM-LSSVM-PLS online soft sensor forecast model in step (2) comprises the following steps:

(21) apply C means clustering method by described sample sequence cluster, be divided into T sub-data class

L ₁..., L _t, and acquisition is subordinate to angle value accordingly:

Wherein m _klrepresent and be subordinate to angle value, k=1,2 ..., T, l=1,2 ..., T;

(22) for each described subdata class L ₁..., L _t, set up submodel h based on LSSVM ₁() ..., h _t();

(23) according to each sample in calculate all submodels respectively prediction output valve by described prediction output valve the input value of angle value μ as offset minimum binary is subordinate to, actual measured value y=[y with described ₁..., y _n] ^tas output valve, set up partial least square model, wherein regression function is F (), and described CM-LSSVM-PLS online soft sensor forecast model is:

y＝F(h ₁(x),...h _T(x),μ)。

Further, the method dynamically updating described CM-LSSVM-PLS online soft sensor forecast model in step (3) comprises the following steps:

(31) by C means clustering method, the subdata class needing to upgrade is found out;

(32) according to similarity, select and new sample (x _j, y _j) immediate sample is as being replaced sample (x _i, y _i), and with described new sample (x _j, y _j) be replaced sample described in replacement;

(33) corresponding to the described subdata class needing to upgrade submodel upgrades.

Further, step (32), according to similarity, is selected and new sample (x _m, y _m) immediate sample is as being replaced sample (x _i, y _i), comprise the following steps:

$i=\arg \left(\min {\left|\right|x_k-x_m\left|\right|}_{k=1}^N\right).$

Further, in step (32) with described new sample (x _m, y _m) be replaced sample (x described in replacement _i, y _i) method comprise the following steps:

(321) sample (x will be replaced _i, y _i) with described sample sequence in last sample (x _n, y _n) switch;

(322) with new sample (x _m, y _m) replace the sample (x being in described last sample position of sample sequence _i, y _i).

Further, step (33) comprises the following steps the described method needing the submodel upgraded to upgrade:

(331) step (321) will be replaced sample (x _i, y _i) with described sample sequence in last sample (x _n, y _n) switch, fisrt feature matrix Φ ₁for:

Wherein, c _i=K (x _i, x _i)+1/2 γ,

P _i＝[K(x ₁,x _i),...,K(x _i-1,x _i),K(x _N,x _i),K(x _i+1,x _i),...,K(x _N-1,x _i)] ^T；

(332) fisrt feature matrix Φ is calculated ₁inverse matrix:

$\begin{array}{c}{\mathrm{\Φ}}_1^{-1}={\left[\begin{array}{cc}P& P_i\\ P_i& c_i\end{array}\right]}^{-1}\\ =\left[\begin{array}{cc}{P}^{-1}& 0\\ 0& 0\end{array}\right]+\left[\begin{array}{cc}{P}^{-1}{P}_i{s}_i^{-1}{P}_i^T{P}^{-1}& -P^{-1}{P}_i{s}_i^{-1}{P}_i^T\\ s_i^{-1}{P}_i^T{P}^{-1}& s_i^{-1}\end{array}\right]\end{array},$

Wherein $s_i=c_i-P_i^T{P}^{-1}{P}_i;$

(333) establish ${\mathrm{\Φ}}_1^{-1}=\left[\begin{array}{cc}{\widetilde{\mathrm{\Ψ}}}_{11}& {\widetilde{\mathrm{\Ψ}}}_{12}\\ {\widetilde{\mathrm{\Ψ}}}_{21}& {\widetilde{\mathrm{\Ψ}}}_{22}\end{array}\right],$ The inverse matrix of the matrix P then in formula is:

$P^{-1}={\widetilde{\mathrm{\Ψ}}}_{11}-{\widetilde{\mathrm{\Ψ}}}_{12}{\widetilde{\mathrm{\Ψ}}}_{22}^{-1}{\widetilde{\mathrm{\Ψ}}}_{21};$

(334) the described new sample (x of step (322) _m, y _m) replace the sample (x being in described last sample position of sample sequence _i, y _i), second characteristic matrix Φ ₂for:

${\mathrm{\Φ}}_2=\left[\begin{array}{cccc}K(x_1,x_1)+\frac{1}{2\mathrm{\γ}}& K(x_1,x_N)& \·\·\·& K(x_1,x_j)\\ K({x_N}^{},x_1)& K(x_N,x_N)+\frac{1}{2\mathrm{\γ}}& \·\·\·& K(x_N,x_j)\\ \·& \·& \·& \·\\ \·& \·& \·& \·\\ \·& \·& \·& \·\\ K(x_j,x_1)& K(x_j,x_N)& \·\·\·& K(x_j,x_j)+\frac{1}{2\mathrm{\γ}}\end{array}\right]=\left[\begin{array}{cc}P& P_j\\ P_j& c_j\end{array}\right],$

Wherein P _j=[K (x ₁, x _j) ..., K (x _i-1, x _j), K (x _n, x _j), K (x _i+1, x _j) ..., K (x _n-1, x _j)] ^t,

C _j=K (x _j, x _j)+1/2 γ, γ represent punishment parameter;

(335) second characteristic matrix Φ is calculated ₂inverse matrix be:

${\mathrm{\Φ}}_2^{-1}={\left[\begin{array}{cc}P& P_j\\ P_j& c_j\end{array}\right]}^{-1}=\left[\begin{array}{cc}{P}^{-1}+P^{-1}{P}_j{s}_j{P}_j^T{P}^{-1}& -P^{-1}{P}_j{s}_j^{-1}\\ -s_j^{-1}{P}_j^T{P}^{-1}& s_j^{-1}\end{array}\right];$

(336) model parameter of the submodel of needs renewal is:

$\left\{\begin{array}{c}{\mathrm{\α}}^{\′}={\mathrm{\Φ}}_2^{-1}{y}^{\′}-{\mathrm{\Φ}}_2^{-1}e\·\frac{e^T{\mathrm{\Φ}}_2^{-1}{y}^{\′}}{e^T{\mathrm{\Φ}}_2^{-1}{y}^{\′}}\\ b^{\′}=\frac{e^T{\mathrm{\Φ}}_2^{-1}{y}^{\′}}{e^T{\mathrm{\Φ}}_2^{-1}{y}^{\′}}\end{array}\right..$

Further, described measured value is coal-fired power station boiler furnace outlet gas temperature.

Compared with prior art, a kind of online soft sensor method provided by the invention, has following beneficial effect:

(1) C mean cluster-least square method supporting vector machine-offset minimum binary (CM-LSSVM-PLS) method is adopted to predict measured value, the modeling problem of complexity is resolved into multiple simple subproblem, precision and there is obvious advantage the modeling time;

(2) CM-LSSVM-PLS online soft sensor forecast model is dynamically updated, can the dynamic perfromance of tracing system, improve the precision of online soft sensor;

(3), when dynamically updating CM-LSSVM-PLS online soft sensor forecast model, replaced the matrix inversion operation of original complexity by simple matrix plus-minus, algorithm is relatively simple, can calculate result fast, ensures that model adapts to the requirement of on-line monitoring;

(4) hard measurement system is by Real-time data drive, according to real time data constantly Renewal model, can have data and easily obtain, and extra hardware drops into the advantages such as little and model prediction accuracy high self-adaptation is strong.

Accompanying drawing explanation

Fig. 1 is coal-fired power station boiler block diagram;

Fig. 2 is the modeling process of CM-LSSVM-PLS online soft sensor forecast model in the online soft sensor method of measurement furnace outlet gas temperature in one embodiment of the present of invention;

Fig. 3 dynamically updates flow process for the CM-LSSVM-PLS online soft sensor forecast model shown in Fig. 2;

Fig. 4 is the predicted value of CM-LSSVM-PLS online soft sensor forecast model shown in Fig. 2 and the graph of a relation of actual measured value;

Fig. 5 contrasted for the modeling time of the CM-LSSVM-PLS online soft sensor forecast model shown in Fig. 2 and traditional LSSVM model;

The measured value of CM-LSSVM-PLS online soft sensor forecast model prediction shown in the actual measurement data that Fig. 6 is certain day boiler furnace outlet cigarette temperature, Fig. 2 and the contrast figure of model modification predicted value.

Embodiment

Elaborate to embodiments of the invention below, the present embodiment is implemented under premised on technical solution of the present invention, give detailed embodiment and concrete operating process, but protection scope of the present invention is not limited to following embodiment.

As shown in Figure 1, the present embodiment comprises: coal pulverizer 1, drum 2, downcomer 3, water-cooling wall 4, Secondary Air 5, First air 6, pendant superheater 7, high temperature superheater 8, high temperature reheater 9, low temperature superheater 10, low-temperature reheater 11, economizer 12.

This boiler is that 300MW is subcritical, Natural Circulation, resuperheat, two arch list burner hearth, " W " flame combustion mode, dry ash extraction, the coal-fired drum boiler of all steel frame suspension type.

Step 1: by coal fired power plant DCS system, obtains boiler actual operating data, comprises input parameter (primary air flow, secondary air flow, confluent, coal-supplying amount and main steam flow etc.) and output parameter (furnace outlet gas temperature).In reality, sample data is about 1,000, and according to actual load situation through row classification, the quantity of classification can not be too large, can not be too little, generally at 4-6.Application C means clustering method, by original training data cluster, is divided into T subclass L ₁..., L _t, obtain simultaneously and be subordinate to angle value accordingly: wherein,

Step 2: based on LSSVM, to each subclass L ₁..., L _t, set up submodel respectively, i.e. h ₁() ..., h _t();

Y=[y ₁..., y _n] ^tas output valve, set up partial least square model, the regression function obtained is F (),

Wherein:

Final CM-LSSVM-PLS model is as follows:

y＝F(h ₁(x),...h _T(x),μ) (1)

Step 5: setting prediction error threshold is ε ₀.For new online input data x _j, obtain the output valve of model, be designated as the predicated error of computation model if error is less than ε ₀, continue step 5, otherwise jump to step 6.

The precision of prediction of step 6:CM-LSSVM-PLS model is unacceptable, needs to upgrade model.To new online input data (x _j,y _j), according to C mean cluster, by Data Placement to certain subclass L _i, LSSVM model corresponding to corresponding subclass needs to upgrade.At subclass L _iin, find the data point (x be replaced _i, y _i).According to similarity formula, select and (x _j,y _j) immediate sample is as being replaced data, concrete replacement samples selection is as follows:

$i=\arg \left(\min {\left|\right|x_k-x_j\left|\right|}_{k=1}^N\right).$ (2)

Step 7: for subclass L _ilSSVM model upgrade.

The learning process of LSSVM is as follows:

$\left[\begin{array}{cc}0& e^T\\ e& \mathrm{\Φ}\end{array}\right]\left[\begin{array}{c}b\\ \mathrm{\α}\end{array}\right]=\left[\begin{array}{c}0\\ y\end{array}\right]---\left(3\right)$

The foundation of LSSVM model is mainly trained and is obtained model parameter, for:

$\left\{\begin{array}{c}\mathrm{\α}={\mathrm{\Φ}}^{-1}y-{\mathrm{\Φ}}^{-1}e\·\frac{e^T{\mathrm{\Φ}}^{-1}y}{e^T{\mathrm{\Φ}}^{-1}e}\\ b=\frac{e^T{\mathrm{\Φ}}^{-1}y}{e^T{\mathrm{\Φ}}^{-1}e}\end{array}\right.---\left(4\right)$

In online updating process, suppose the data point (x in a Geju City _i, y _i) by new online data point (x _j, y _j) replaced.If re-establish model, need to recalculate model parameter α and b according to formula (4), wherein relate to inverting of an eigenmatrix Φ, the dimension of this matrix is very high, the inversion calculation of matrix is chronic, is also the bottleneck of whole LSSVM algorithm.The present invention is derived by mathematical theory, avoids, directly to matrix inversion, obtaining LSSVM model parameter α and b.

The replacement process of online updating is divided into two steps: first, by sample (x _i, y _i) with sample sequence in last sample (x _n, y _n) switch; Second step, with new data point (x _j, y _j) replace (x be on reposition _i, y _i).

In a first step, the eigenmatrix Φ of front and back is replaced in contrast, only has a line and row to there occurs change, and i-th row of the Φ namely in eigenmatrix and N-th row and i-th arrange and N row exchange.After exchange, obtain new eigenmatrix, for:

Eigenmatrix Φ to Φ ₁conversion, can premultiplication matrix be passed through a matrix is taken advantage of with the right side matrix that the i-th row and the N-th row of cross-over unit matrix obtains. that i-th of cross-over unit matrix arranges and N row obtain.Therefore, these two matrixes are all unit matrixs, and order then: Φ ₁=I Φ I.

Therefore, Φ ₁inverse matrix be:

Φ ₁ ^-1＝I·Φ ^-1·I (6)

By Φ ₁be divided into:

${\mathrm{\Φ}}_1=\left[\begin{array}{cc}P& P_i\\ P_i& c_i\end{array}\right]---\left(7\right)$

In formula, c _i=K (x _i, x _i)+1/2 γ,

P _i＝[K(x ₁,x _i),...,K(x _i-1,x _i),K(x _N,x _i),K(x _i+1,x _i),...,K(x _N-1,x _i)] ^T。

According to the Inversion Formula of partitioned matrix, matrix Φ ₁inverse matrix be:

$\begin{array}{c}{\mathrm{\Φ}}_1^{-1}={\left[\begin{array}{cc}P& P_i\\ P_i& c_i\end{array}\right]}^{-1}\\ =\left[\begin{array}{cc}{P}^{-1}& 0\\ 0& 0\end{array}\right]+\left[\begin{array}{cc}{P}^{-1}{P}_i{s}_i^{-1}{P}_i^T{P}^{-1}& -P^{-1}{P}_i{s}_i^{-1}{P}_i^T\\ s_i^{-1}{P}_i^T{P}^{-1}& s_i^{-1}\end{array}\right]\end{array}---\left(8\right)$

In formula: $s_i=c_i-P_i^T{P}^{-1}{P}_i.$

If ${\mathrm{\Φ}}_1^{-1}=\left[\begin{array}{cc}{\widetilde{\mathrm{\Ψ}}}_{11}& {\widetilde{\mathrm{\Ψ}}}_{12}\\ {\widetilde{\mathrm{\Ψ}}}_{21}& {\widetilde{\mathrm{\Ψ}}}_{22}\end{array}\right],$ The inverse matrix of the matrix P then in formula is:

$P^{-1}={\widetilde{\mathrm{\Ψ}}}_{11}-{\widetilde{\mathrm{\Ψ}}}_{12}{\widetilde{\mathrm{\Ψ}}}_{22}^{-1}{\widetilde{\mathrm{\Ψ}}}_{21}---\left(9\right)$

For initial model, in the process of modeling, the inverse matrix of initial eigenmatrix Φ has calculated.Therefore, through type (6) can obtain matrix Φ ₁inverse matrix, and then to obtain according to formula (9), the matrix P of matrix of matrix P can be obtained ^-1.

Second step, with new data (x _j,y _j) replace (x be on reposition _i, y _i).New eigenmatrix Φ ₂for:

${\mathrm{\Φ}}_2=\left[\begin{array}{cccc}K(x_1,x_1)+\frac{1}{2\mathrm{\γ}}& K(x_1,x_N)& \·\·\·& K(x_1,x_j)\\ K({x_N}^{},x_1)& K(x_N,x_N)+\frac{1}{2\mathrm{\γ}}& \·\·\·& K(x_N,x_j)\\ \·& \·& \·& \·\\ \·& \·& \·& \·\\ \·& \·& \·& \·\\ K(x_j,x_1)& K(x_j,x_N)& \·\·\·& K(x_j,x_j)+\frac{1}{2\mathrm{\γ}}\end{array}\right]=\left[\begin{array}{cc}P& P_j\\ P_j& c_j\end{array}\right]---\left(10\right)$

(10) the matrix P in formula is identical with the matrix P in (8).Matrix P _jwith constant c _jas follows:

P _j＝[K(x ₁,x _j),...,K(x _i-1,x _j),K(x _N,x _j),K(x _i+1,x _j),...,K(x _N-1,x _j)] ^T，

c _j＝K(x _j,x _j)+1/2γ。

According to formula (8), eigenmatrix Φ ₂inverse matrix be:

${\mathrm{\Φ}}_2^{-1}={\left[\begin{array}{cc}P& P_j\\ P_j& c_j\end{array}\right]}^{-1}=\left[\begin{array}{cc}{P}^{-1}+P^{-1}{P}_j{s}_j{P}_j^T{P}^{-1}& -P^{-1}{P}_j{s}_j^{-1}\\ -s_j^{-1}{P}_j^T{P}^{-1}& s_j^{-1}\end{array}\right]---\left(11\right)$

According to the matrix P that formula (9) obtains ^-1, new eigenmatrix Φ ₂inverse matrix can be calculated by (11).

Like this, the parameter of model after upgrading, can be calculated by (4).New model parameter is as follows:

$\left\{\begin{array}{c}{\mathrm{\α}}^{\′}={\mathrm{\Φ}}_2^{-1}{y}^{\′}-{\mathrm{\Φ}}_2^{-1}e\·\frac{e^T{\mathrm{\Φ}}_2^{-1}{y}^{\′}}{e^T{\mathrm{\Φ}}_2^{-1}{y}^{\′}}\\ b^{\′}=\frac{e^T{\mathrm{\Φ}}_2^{-1}{y}^{\′}}{e^T{\mathrm{\Φ}}_2^{-1}{y}^{\′}}\end{array}\right.---\left(12\right)$

In formula, y' is the output of model after upgrading, and is y'=[y ₁..., y _i-1, y _n, y _i+1..., y _n-1, y _j] ^t.

Step 8: according to the submodel h after renewal _i(), re-establishes the CM-LSSVM-PLS model after renewal.Complete whole online updating process.

Upload interface by the Distributed Control System (DCS) of boiler, result is uploaded in Distributed Control System (DCS) and carry out cooperation control, and be transferred to by network line in the watch-dog of field control room, provide effective reference for operations staff to current operation.

More than describe preferred embodiment of the present invention in detail.Should be appreciated that those of ordinary skill in the art just design according to the present invention can make many modifications and variations without the need to creative work.Therefore, all technician in the art, all should by the determined protection domain of claims under this invention's idea on the basis of existing technology by logic analysis, reasoning and the available technical scheme of limited experiment.

Claims (7)

1. an online soft sensor method, is characterized in that, comprises the following steps:

(1) N group service data is obtained, as N group sample sequence

(2) according to described sample sequence set up CM-LSSVM-PLS online soft sensor forecast model, and use described sample sequence to train described CM-LSSVM-PLS online soft sensor forecast model;

(3) new sample (x is obtained by DCS system _j, y _j) in x _jinput described CM-LSSVM-PLS online soft sensor forecast model, obtain the measured value predicted computational prediction error wherein y _jfor actual measured value, if predicated error ε is greater than prediction error threshold ε ₀, dynamically update described CM-LSSVM-PLS online soft sensor forecast model.

2. online soft sensor method as claimed in claim 1, it is characterized in that, the method setting up CM-LSSVM-PLS online soft sensor forecast model in step (2) comprises the following steps:

(21) apply C means clustering method by described sample sequence cluster, be divided into T data class L ₁..., L _t, and acquisition is subordinate to angle value accordingly:

Wherein μ _klrepresent and be subordinate to angle value, k=1,2 ..., T, l=1,2 ..., T;

(22) for each described subdata class L ₁..., L _t, set up submodel h based on LSSVM ₁() ..., h _t();

(23) according to sample in each data sample, calculate all submodels respectively prediction output valve by described prediction output valve the input value of angle value μ as offset minimum binary is subordinate to, actual measured value y=[y with described ₁..., y _n] ^tas output valve, set up partial least square model, wherein regression function is F (), and described CM-LSSVM-PLS online soft sensor forecast model is:

y＝F(h ₁(x),...h _T(x),μ)。

3. online soft sensor method as claimed in claim 2, it is characterized in that, the method dynamically updating described CM-LSSVM-PLS online soft sensor forecast model in step (3) comprises the following steps:

(31) by C means clustering method, by Data Placement to certain subclass, this data class is exactly the data class needing to upgrade;

(32) according to similarity, select and new sample (x _j, y _j) immediate sample is as being replaced sample (x _i, y _i), and with described new sample (x _j, y _j) be replaced sample described in replacement;

(33) corresponding to the described subdata class needing to upgrade submodel parameter upgrades.

4. online soft sensor method as claimed in claim 3, is characterized in that, step (32), according to similarity, is selected and new sample (x _m, y _m) immediate sample is as being replaced sample (x _i, y _i), concrete replacement samples selection is as follows:

$i=\arg \left(\min {\left|\right|x_k-x_m\left|\right|}_{k=1}^N\right).$

5. online soft sensor method as claimed in claim 4, is characterized in that, with described new sample (x in step (32) _m, y _m) be replaced sample (x described in replacement _i, y _i) method comprise the following steps:

(321) sample (x will be replaced _i, y _i) with described sample sequence in last sample (x _n, y _n) switch;

(322) with new sample (x _m, y _m) replace the sample (x being in described last sample position of sample sequence _i, y _i).

6. online soft sensor method as claimed in claim 5, is characterized in that, step (33) comprises the following steps the described method needing the submodel upgraded to upgrade:

(331) step (321) will be replaced sample (x _i, y _i) with described sample sequence in last sample (x _n, y _n) switch, fisrt feature matrix Φ ₁for:

Wherein, c _i=K (x _i, x _i)+1/2 γ,

P _i＝[K(x ₁,x _i),...,K(x _i-1,x _i),K(x _N,x _i),K(x _i+1,x _i),...,K(x _N-1,x _i)] ^T；

(332) fisrt feature matrix Φ is calculated ₁inverse matrix:

$\begin{array}{c}{\mathrm{\Φ}}_1^{-1}={\left[\begin{array}{cc}P& P_i\\ P_i& c_i\end{array}\right]}^{-1}\\ =\left[\begin{array}{cc}{P}^{-1}& 0\\ 0& 0\end{array}\right]+\left[\begin{array}{cc}{P}^{-1}{P}_i{s}_i^{-1}{P}_i^T{P}^{-1}& -P^{-1}{P}_i{s}_i^{-1}{P}_i^T\\ s_i^{-1}{P}_i^T{P}^{-1}& s_i^{-1}\end{array}\right]\end{array},$

Wherein $s_i=c_i-P_i^T{P}^{-1}{P}_i;$

(333) establish ${\mathrm{\Φ}}_1^{-1}=\left[\begin{array}{cc}{\widetilde{\mathrm{\Ψ}}}_{11}& {\widetilde{\mathrm{\Ψ}}}_{12}\\ {\widetilde{\mathrm{\Ψ}}}_{21}& {\widetilde{\mathrm{\Ψ}}}_{22}\end{array}\right],$ The inverse matrix of the matrix P then in formula is:

$P^{-1}={\widetilde{\mathrm{\Ψ}}}_{11}-{\widetilde{\mathrm{\Ψ}}}_{12}{\widetilde{\mathrm{\Ψ}}}_{22}^{-1}{\widetilde{\mathrm{\Ψ}}}_{21};$

(334) the described new sample (x of step (322) _m, y _m) replace the sample (x being in described last sample position of sample sequence _i, y _i), second characteristic matrix Φ ₂for:

${\mathrm{\Φ}}_2=\left[\begin{array}{cccc}K(x_1,x_1)+\frac{1}{2\mathrm{\γ}}& K(x_1,x_N)& \·\·\·& K(x_1,x_j)\\ K(x_N,x_1)& K(x_N,x_N)+\frac{1}{2\mathrm{\γ}}& \·\·\·& K(x_N,x_j)\\ \·& \·& \·& \·\\ \·& \·& \·& \·\\ \·& \·& \·& \·\\ K(x_j,x_1)& K(x_j,x_N)& \·\·\·& K(x_j,x_j)+\frac{1}{2\mathrm{\γ}}\end{array}\right]=\left[\begin{array}{cc}P& P_j\\ P_j& c_j\end{array}\right],$

Wherein P _j=[K (x ₁, x _j) ..., K (x _i-1, x _j), K (x _n, x _j), K (x _i+1, x _j) ..., K (x _n-1, x _j)] ^t,

C _j=K (x _j, x _j)+1/2 γ, γ represent punishment parameter;

(335) second characteristic matrix Φ is calculated ₂inverse matrix be:

${\mathrm{\Φ}}_2^{-1}={\left[\begin{array}{cc}P& P_j\\ P_j& c_j\end{array}\right]}^{-1}=\left[\begin{array}{cc}{P}^{-1}+P^{-1}{P}_j{s}_j{P}_j^T{P}^{-1}& -P^{-1}{P}_j{s}_j^{-1}\\ -s_j^{-1}{P}_j^T{P}^{-1}& s_j^{-1}\end{array}\right]\text{;}$

(336) model parameter of the submodel of needs renewal is:

$\left\{\begin{array}{c}{a}^{\′}={\mathrm{\Φ}}_2^{-1}{y}^{\′}-{\mathrm{\Φ}}_2^{-1}e\·\frac{e^T{\mathrm{\Φ}}_2^{-1}{y}^{\′}}{e^T{\mathrm{\Φ}}_2^{-1}{y}^{\′}}\\ b^{\′}=\frac{e^T{\mathrm{\Φ}}_2^{-1}{y}^{\′}}{e^T{\mathrm{\Φ}}_2^{-1}{y}^{\′}}\end{array}\right..$

7. online soft sensor method as claimed in claim 1, it is characterized in that, described measured value is coal-fired power station boiler furnace outlet gas temperature.