CN104715142A - NOx emission dynamic soft-sensing method for power station boiler - Google Patents

Abstract

The invention discloses a NOx emission dynamic soft-sensing method for a power station boiler. The method comprises the steps of acquisition and preprocessing of data, initialization of a self-adaptive particle swarm algorithm and the like. According to the method, related operating and state parameters of a boiler combustion system serve as the input of a model, the nitrogen oxide emission concentration serves as the output of the model, historical operating data are selected as training samples, a support vector regression machine serves as a soft-sensing modeling tool, the idea of a non-linear auto-regression moving average model is combined, the orders of input variables and output variables of the model are considered, and therefore the soft-sensing model has the capability of describing the dynamic change process. By means of the method, the change of NOx emission in the boiler combustion dynamic operating process can be effectively traced and predicted, and the method has important significance on safe and optimized operation of the power station boiler.

Description

A kind of station boiler NO xdischarge dynamic soft-measuring method

Technical field

The present invention relates to heat power engineering and artificial intelligence interleaving techniques field, be specifically related to a kind of station boiler NO _xdischarge dynamic soft-measuring method.

Background technology

Along with the fast development of China's economy and power industry, the oxides of nitrogen that coal fired boiler of power plant produces becomes the main source of Pollution Index in Air Nitric Oxides, for meeting the environmental requirement of increasingly stringent, to the NO of coal unit _xthe control of discharge is had higher requirement.Realize station boiler NO _xthe prerequisite that the Emission Optimization controls sets up effective NO _xdischarge soft-sensing model.

Due to NO _xgenerate complicated in burning of coal product, boiler combustion system is also part very complicated in electric power station system, and the input variable related to is numerous and have strong nonlinearity and strong coupling, is difficult to set up mechanism model accurately.In recent years, along with the universal history data of magnanimity that makes of power station DCS, SIS system is retained, for neural network, support vector machine etc. provide good applied environment based on the soft-measuring technique of data-driven.Wherein, the support vector regression (SVMR) of structure based principle of minimization risk has shown generalization ability more better than neural network, solves the difficulties such as the small sample existed in soft sensor modeling, non-linear, over-fitting preferably.

Prior art great majority adopt the method for Steady state modeling to station boiler NO _xstable state soft-sensing model is set up in discharge, and because steady state condition down-sampled data is relatively more accurate, therefore steady-state model precision is higher.But the station boiler of reality uses stable state soft-sensing model to there is following problem in running:

First: usually only consider the input and output as stable state soft-sensing model of data that each measuring point sequential mates, but the station boiler operational process great majority of reality are in dynamic change, certain state of current time system output quantity is that input state for the previous period determines thus, instead of the input state of some time points determines separately, namely have ignored the delay character of combustion process;

Second: because stable state soft-sensing model needs system to remain on certain hour under steady state condition, to ensure the accuracy obtaining steady state data, but under also making its applicability only be confined to steady state condition, lack the description to process dynamics variation characteristic, once boiler operatiopn occurs operating point change or produces disturbance, the tracking power of stable state soft-sensing model can significantly decline;

3rd: station boiler steady state operating conditions is often difficult to meet, and also limit the application of steady-state model.

Summary of the invention

Goal of the invention: the object of the invention is to solve the deficiencies in the prior art, provides a kind of flue gas in power station boiler NO realized based on support vector regression (SVMR) and nonlinear auto regressive moving average framework (NARMAX) _xdischarge dynamic soft-measuring method.

Technical scheme: a kind of station boiler NO of the present invention _xdischarge dynamic soft-measuring method, comprises the following steps:

(1) data acquisition and pre-service: Confirming model input variable and output variable, collect original sample from DCS system to original sample adopt normalization preprocess method, and the NO will recorded under each operating mode _xconcentration conversion is to the value under 6% oxygen amount; Wherein x _i∈ R ^mrepresent the i-th group model input amendment, m is the number of input variable, y _i∈ R represents the i-th group model output sample, and n is sample size;

Monitoring Data due to Environmental Protection Agency is all combustion conditions of a unified setting standard: O in residue flue gas ₂concentration be 6%, therefore in this patent, refer under each operating mode owing to remaining flue gas O in Actual combustion ₂concentration in differ be decided to be 6% situation, therefore need to convert by convert formula;

(2) initialization APSO algorithm: each particle p (i) comprises m+3 variable, comprising m input variable x _icorresponding order d _i, 1 order d that output variable y is corresponding _yand the penalty factor of support vector regression SVMR model and nuclear parameter σ, need two parameters of artificial setting when penalty factor and nuclear parameter σ are SVMR modeling, its value is comparatively large to model hard measurement Influence on test result, therefore needs together to be optimized;

(3) input data are become nonlinear auto regressive moving average framework NARMAX model structure with output data preparation: the input variable x comprised with particle p (i) _iorder d _iwith the order d of output variable y _y, input data are become NARMAX model structure with output data preparation, and input X (t) and output Y (t) that then calculate soft-sensing model are as follows:

$\left\{\begin{array}{c}X\left(t\right)=[x_1(t-T),...,x_1(t-d_{1}T),...,x_m(t-T),...,x_m(t-d_{m}T),y(t-T),...,y(t-d_{y}T)]\\ Y\left(t\right)=y\left(t\right)\end{array}\right.$

In above formula, T collects employing cycle of measuring point data, and X (t) is mode input amount corresponding to sampling instant t, comprises m input variable x _icontinuous d before sampling instant t _ithe quantity of state in individual sampling period, and the continuous d of measured y before sampling instant t _ythe quantity of state in individual sampling period; Y (t) is model output corresponding to sampling instant t, is the quantity of state of measured y when sampling instant t;

(4) SVMR model initialization: adopt gaussian kernel function K (x _i, x _j)=exp (-|| x _i-x _j|| ²/ (2 σ ²)) namely RBF is as the kernel function of SVMR modeling, the penalty factor comprised with particle p (i) and nuclear parameter σ are to arrange SVMR model:

$Y\left(t\right)=\underset{i=1}{\overset{l}{\mathrm{\Σ}}}({\mathrm{\α}}_i^{*}-{\mathrm{\α}}_i)\exp \{-\frac{{\left|\right|X\left(t\right)-X_i^{\mathrm{sv}}\left|\right|}^2}{{2\mathrm{\σ}}^2}\}+b$

In above formula, α _ibe support vector regression parameter with b, for support vector, l is the number of support vector, and σ is RBF nuclear parameter;

(5) carry out SVMR training and matching, calculate particle ideal adaptation degree f _i, retain optimum individual: using X (t) as SVMR mode input, Y (t) exports as SVMR model, using before the sample set after arranging 3/4 as training sample, rear 1/4 as forecast sample, carries out training and the matching of model; With forecast sample Deviation Indices for fitness function, calculate particle ideal adaptation degree f _i, and to obtain colony's optimal-adaptive degree be f _mand retain its particle position P _m, so the fitness function of particle p (i) is:

$f\left(p_i\right)=\frac{1}{n}\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{(y_j-{\hat{y}}_j)}^2$

In above formula, n is predicted time sequence length, y _jwith be respectively measured value and the model predication value in a jth moment;

(6) optimizing convergence judges: if f _m<f _e, wherein f _efor expecting convergence threshold, work as k=K _maxtime, then optimizing stops, and performs step (9); Otherwise, make k=k+1, perform step (7); Wherein, k is current iteration number of times, K _maxfor maximum iteration time, the Search Range of order is [d _min, d _max];

(7) calculate particle colony average fitness, particle colony is divided into three subgroups: in kth time iteration, the fitness value of particle p (i) is f _i ^k, the fitness value of colony's optimal particle is f _m; Population average fitness is fitness value is better than particle fitness be averaging and obtain according to particle ideal adaptation degree f _i ^kbe divided into local optimal searching, balance optimizing, global optimizing three subgroups, calculated inertia weight w and the Studying factors c of its correspondence respectively ₁, c ₂corresponding step (7.3) is classified, and calculates this three coefficients;

(8) the individual speed of particle and location updating: the inertia weight w obtained according to step (7) and Studying factors c ₁, c ₂, to the individual speed V of particle _iwith position P _icarry out calculating by following formula to upgrade

v _in＝wv _in+c ₁rand()(b _in-p _in)+c ₂rand()(b _gn-p _in)

p _in＝p _in+v _in

Wherein, v _infor particle rapidity, p _infor particle position, b _infor personal best particle, b _gnfor colony's optimal location;

(9) export optimal result, complete the foundation of dynamic soft sensor model: export f _mthe d that corresponding optimum individual comprises _i, d _yand penalty factor and nuclear parameter σ, and set up final dynamic soft measuring SVMR model using its parameter as setup parameter;

(10) realization of dynamic soft measuring: for the sample of new collection arbitrarily, after first carrying out pre-service to it, the dynamic soft measuring SVMR model that input step 9 obtains, just can obtain corresponding hard measurement and export.

Further, in described step (1), choose the NO before furnace of power-plant boilers afterbody SCR reactor _xinlet concentration is as output quantity y _i, choose in boiler controller system dynamic running process NO _xdischarge the combustion system had an impact and run variable as input variable x _i; And original sample for the dynamically operating condition data continuously of the large and variable load operation of coverage in DCS system.

Further, in described step (1) to original sample the pretreated detailed process of employing normalization is:

The original value of each input data is zoomed in designation area by following formula:

$x^{\&prime;}=\frac{x-\min \left(x\right)}{\max \left(x\right)-\min \left(x\right)}$

Wherein, x is the value before normalization, and x ' is the value after normalization, and max (x) and min (x) is respectively maximal value between normalization back zone and minimum value;

Because under each operating mode, flue gas oxygen content is inconsistent, for ease of comparing NO _xvalue, according to following formula, by the NO recorded under each operating mode _xconcentration conversion is to the value under 6% oxygen amount:

${\mathrm{\&rho;}}^{\&prime;}\left(\mathrm{NOx}\right)=\mathrm{\&rho;}\left(\mathrm{NOx}\right)*\frac{21-{\mathrm{\&rho;}}^{\&prime;}\left(O_2\right)}{21-\mathrm{\&rho;}\left(O_2\right)}$

Wherein, ρ ' (NOx) and ρ ' (O ₂) represent the NO after converting respectively _xconcentration and O ₂concentration, ρ (NOx) and ρ (O ₂) represent the NO of actual measurement respectively _xconcentration and O ₂concentration.

Further, in described step (7) according to particle ideal adaptation degree f _i ^kthe process of three subgroups population is divided into be:

(7.1) calculate particle colony average fitness, the fitness value then setting particle p (i) in kth time iteration is f _i ^k, the fitness value of colony's optimal particle is f _m, particle colony average fitness value is

(7.2) fitness value is better than particle fitness average and obtain definition be used for evaluating the prematurity convergence degree of population, △ less explanation population body is more tending towards Premature Convergence;

(7.3) last according to particle ideal adaptation degree f _i ^kbe divided into following three classes:

If 1. then show that this particle is good particle in colony, be divided into local optimal searching subgroup, get less inertia weight w by following formula, make particle carry out Local Search, get larger c simultaneously ₁less c ₂, to strengthen the individual learning ability of self of particle and the local optimal searching ability of enhanced particles:

$w=w-(w-w_{\min })\&CenterDot;\left|\frac{{f_i}^k-f_{\mathrm{avgb}}^k}{f_m-f_{\mathrm{avgb}}^k}\right|$

$c_1=c_{1\min }+(c_{1\max }-c_{1\min })\&CenterDot;\left|\frac{{f_i}^k-f_{\mathrm{avgb}}^k}{f_m-f_{\mathrm{avgb}}^k}\right|$

$c_2=c_{2\max }-(c_{2\max }-c_{2\min })\&CenterDot;\left|\frac{{f_i}^k-f_{\mathrm{avgb}}^k}{f_m-f_{\mathrm{avgb}}^k}\right|$

If 2. then show that this particle is particle general in colony, be divided into balance optimizing subgroup, there is good global optimizing ability and local optimal searching ability simultaneously, maintain inertia weight w constant, and get c ₁=c ₂=1, the learning ability of equilibrium particle self cognition and social recognition;

If 3. then show that this particle is particle poor in colony, be divided into global optimizing subgroup, get larger w, to increase amplitude of variation, the ability of searching optimum of enhanced particles colony, gets less c simultaneously ₁larger c ₂, social learning's ability of enhanced particles, adjustable strategies carries out according to following formula:

$w=1.5-\frac{1}{1+k_1\&CenterDot;\exp (-k_2\&CenterDot;\mathrm{\&Delta;})}$

$c_1=c_{1\max }=-\frac{1}{2}(c_{1\max }-c_{1\min })\&CenterDot;\left|\frac{{f_i}^k-f_m}{f_{\mathrm{avg}}^k-f_m}\right|$

$c_2=c_{2\min }=+\frac{1}{2}(c_{2\max }-c_{2\min })\&CenterDot;\left|\frac{{f_i}^k-f_m}{f_{\mathrm{avg}}^k-f_m}\right|$

In this step, w is the inertia weight of particle cluster algorithm, and for balanced algorithm ability of searching optimum and local search ability, larger w can strengthen the ability of searching optimum of algorithm, and less w can strengthen the local convergence ability of algorithm and accelerate convergence of algorithm speed.W span is generally [0.4,1.5].

In formula, c ₁and c ₂for the Studying factors of algorithm, span is the [c of setting _1min, c _1max], comparatively large i.e. corresponding interval [(c _1min+ c _1max)/2, c _1max], less i.e. corresponding interval [c _1min, (c _1min+ c _1max)/2], its concrete value is all calculate by computing formula corresponding in the different particle classifying situation of this three class in step (7.3).

In formula, c _1maxand c _2maxbe respectively cognitive learning factor c ₁with social learning factor c ₂the maximal value of the span of setting, c _1minand c _2minfor the minimum value of its span; W is the initial set value of inertia weight, w _minfor the setting minimum value of inertia weight.

Beneficial effect: NARMAX model thought is combined with support vector regression modeling method by the present invention, utilizes APSO algorithm to achieve the global optimization of model parameter simultaneously, compared with prior art, specifically has a little following:

(1) contemplated by the invention the order of soft-sensing model input and output variable, set up the dynamic soft sensor model of oxides of nitrogen, the hard measurement to oxides of nitrogen can be realized in the dynamic changing processes such as unit varying load, effectively improve hard measurement precision;

(2) the present invention adopts APSO algorithm, has carried out overall optimizing to the order of input/output variable and model parameter;

(3) the present invention directly can utilize the measuring point history data read from Power Plant DCS System, and without the need to site test, be easy to engineer applied, cost is low, predicts the outcome reliable.

Accompanying drawing explanation

Fig. 1 is dynamic soft measuring principle schematic of the present invention;

Fig. 2 is the process flow diagram of embodiment;

Fig. 3 is the prediction effect figure of embodiment.

Embodiment

Below technical solution of the present invention is described in detail, but protection scope of the present invention is not limited to described embodiment.

Embodiment: using the quadrangle tangential circle supercritical once-through boiler of certain power plant 600MW Thermal generation unit as research object, adopts the dynamic soft-measuring method based on nonlinear auto-companding running mean and support vector regression of the present invention, realizes its NO _xthe dynamic soft measuring of discharge.

As shown in Figure 2, the concrete steps of the present embodiment are as follows:

(1) data acquisition and pre-service: for power boiler burning system NO _xdynamic soft measuring is carried out in discharge, and model output y (t) is chosen for the NO before burner hearth afterbody SCR reactor _xinlet concentration, mode input optimize indexes conditional parameter (load, coal-supplying amount), main combustion zone parameter (main air intake aperture, secondary air register aperture), burning-out zone parameter (after-flame throttle opening), state parameter (oxygen amount) etc. can to boiler NO _xdischarge the boiler combustion system operation variable had an impact, have four input variables; Choose data in certain varying load adjustment process as training sample, load variations scope is 364.7MW ~ 554.4MW, and the duration is 62min, and the sampling period is set to 20s, acquires 186 groups of data altogether; Adopt normalized, the original value of each input data is zoomed in interval [0,1] by following formula, and the NO will recorded under each operating mode _xconcentration conversion is to the value under 6% oxygen amount;

(2) initialization APSO algorithm: population population scale S=50 is set, maximum iteration time K _max=1000; Each particle p (i) comprises 7 variablees altogether, and wherein 4 is mode input order d _i, 1 export order d _ywith 2 SVMR model parameters, the Search Range that setup times postpones order is [1,20] (i.e. [20s, 400s]), penalty factor is got [1,2000], core width cs is got [0.01,1000], the initial population of stochastic generation particle in this interval;

(3) input data and become NARMAX model structure with output data preparation: the order value d that data comprise by p (i) with output data will be inputted _iand d _ybe organized into NARMAX model structure, input X (t) and output Y (t) that calculate soft-sensing model are as follows:

$\left\{\begin{array}{c}X\left(t\right)=[x_1(t-T),...,x_1(t-d_{1}T),...,x_m(t-T),...,x_m(t-d_{m}T),y(t-T),...,y(t-d_{y}T)]\\ Y\left(t\right)=y\left(t\right)\end{array}\right.$

Using X (t) as SVMR mode input, Y (t) as SVMR model export, the soft-sensing model structure finally obtained as shown in Figure 1, wherein, x _{l_c}(t-T), x _{l_c}(t-2T), K, x _{l_c}(t-d _{l_c}t) d is in the past represented _{l_c}load in the individual period and coal-supplying amount as the input quantity of "current" model, d _{l_c}represent that load and coal-supplying amount are as the order corresponding to soft-sensing model input variable; x _{a_f}(t-T), x _{a_f}(t-2T), K, x _{a_f}(t-d _{a_f}t), x _sofa(t-T), x _sofa(t-2T) ..., x _sofa(t-d _sofat) and one secondary air register aperture, after-flame throttle opening and the oxygen amount of being illustrated respectively in in several periods as the input quantity of model, d _{a_f}, d _sofawith represent that it is as order corresponding during soft-sensing model input quantity respectively, y _nOx(t-T), y _nOx(t-2T) ..., y _nOx(t-d _yt) be past d _ythe output y in individual moment as the input quantity of model, d _yfor the order of output variable;

(4) SVMR model initialization: the penalty factor comprised by p (i) and core width cs setting SVMR model, using X (t) and Y (t) as SVMR mode input and output, sets up particle p (i) corresponding

SVMR-NARMAX model; Its Kernel Function is chosen for gaussian kernel function (RBF), and model form is as follows:

$Y\left(t\right)=\underset{i=1}{\overset{l}{\mathrm{\&Sigma;}}}({\mathrm{\&alpha;}}_i^{*}-{\mathrm{\&alpha;}}_i)\exp \{-\frac{{\left|\right|X\left(t\right)-X_i^{\mathrm{sv}}\left|\right|}^2}{{2\mathrm{\&sigma;}}^2}\}+b$

(5) SVMR training and matching is carried out, calculate particle ideal adaptation degree, retain optimum individual: according to the model set up, carry out the matching of training sample and the prediction of test sample book,, be calculated as follows using square error (MSE) evaluation index as particle fitness function:

${f_i}^k=\frac{1}{n}\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{(y_j-{\hat{y}}_j)}^2---\left(5\right)$

Wherein, n is predicted time sequence length, y _jwith be respectively measured value and the model predication value in a jth moment; Choose min (f _i ^k) individuality that namely prediction deviation is minimum is as colony optimum individual f _mretain;

(6) optimizing convergence judges: if f _m<f _eor k=K _maxtime, wherein f _efor expecting convergence threshold, then optimizing stops, and performs step 9; Otherwise, make k=k+1, perform step 7;

(7) calculate particle colony average fitness, particle colony is divided into three subgroups: in kth time iteration, the fitness value of particle p (i) is f _i ^k, the fitness value of colony's optimal particle is f _m; The average adaptive value of population is fitness value is better than particle fitness value be averaging and obtain definition be used for evaluating the prematurity convergence degree of population, the less explanation population of △ is more tending towards Premature Convergence.Following three classes are divided into according to particle ideal adaptation angle value:

If 1. this particle is good particle in colony, is divided into local optimal searching subgroup, gets less w by following formula, makes it change among a small circle, gets larger c simultaneously ₁less c ₂, to strengthen the individual learning ability of self of particle and the local optimal searching ability of enhanced particles

$w=w-(w-w_{\min })\&CenterDot;\left|\frac{{f_i}^k-f_{\mathrm{avgb}}^k}{f_m-f_{\mathrm{avgb}}^k}\right|---\left(6\right)$

$c_1=c_{1\min }+(c_{1\max }-c_{1\min })\&CenterDot;\left|\frac{{f_i}^k-f_{\mathrm{avgb}}^k}{f_m-f_{\mathrm{avgb}}^k}\right|---\left(7\right)$

$c_2=c_{2\max }-(c_{2\max }-c_{2\min })\&CenterDot;\left|\frac{{f_i}^k-f_{\mathrm{avgb}}^k}{f_m-f_{\mathrm{avgb}}^k}\right|---\left(8\right)$

If 2. this particle is particle general in colony, is divided into balance optimizing subgroup, has good global optimizing ability and local optimal searching ability simultaneously, maintain w constant, and get c ₁=c ₂=1, the learning ability of equilibrium particle self cognition and social recognition;

If 3. this particle is particle poor in colony, is divided into global optimizing subgroup, gets larger w,

To increase amplitude of variation, the ability of searching optimum of enhanced particles colony, gets less c simultaneously ₁larger c ₂, social learning's ability of enhanced particles, adjustable strategies carries out according to following formula:

$w=1.5-\frac{1}{1+k_1\&CenterDot;\exp (-k_2\&CenterDot;\mathrm{\&Delta;})}---\left(9\right)$

$c_1=c_{1\max }=-\frac{1}{2}(c_{1\max }-c_{1\min })\&CenterDot;\left|\frac{{f_i}^k-f_m}{f_{\mathrm{avg}}^k-f_m}\right|---\left(10\right)$

$c_2=c_{2\min }=+\frac{1}{2}(c_{2\max }-c_{2\min })\&CenterDot;\left|\frac{{f_i}^k-f_m}{f_{\mathrm{avg}}^k-f_m}\right|---\left(11\right)$

Wherein, c is got _1max=c _2max=1.8, c _1min=c _2min=0.2, k ₁=1.5, k ₂=0.3;

(8) the individual speed of particle and location updating: obtain each particle corresponding inertia weight w and Studying factors c by the result of calculation of (7) ₁, c ₂, according to following formula, the speed of particle individuality and position are upgraded:

v _in＝wv _in+c ₁rand()(b _in-p _in)+c ₂rand()(b _gn-p _in) (12)

p _in＝p _in+v _in(13)

(9) export optimal result, complete the foundation of dynamic soft sensor model: export f _mthe d that corresponding optimum individual comprises _i, d _yand penalty factor and nuclear parameter σ, as shown in table 1 below, and final dynamic soft measuring SVMR model is set up using its parameter as setup parameter;

Table 1

(10) realization of dynamic soft measuring: for the sample of new collection arbitrarily, after first carrying out pre-service to it, the dynamic soft measuring SVMR model that input step (9) obtains, just can obtain corresponding hard measurement and export.

In the present embodiment, wherein first 140 groups is training sample, and latter 40 groups is test sample book, as shown in Figure 3, utilizes the present invention can to the NO in boiler dynamic running process _xdischarge change is followed the tracks of accurately and predicts, realizes NO _xthe dynamic soft measuring of discharge.

Claims (4)

1. a station boiler NO _xdischarge dynamic soft-measuring method, is characterized in that: comprise the following steps:

(1) data acquisition and pre-service: Confirming model input variable and output variable, collect original sample from DCS system to original sample adopt normalization preprocess method, and the NO will recorded under each operating mode _xconcentration conversion is to the value under 6% oxygen amount; Wherein x _i∈ R ^mrepresent the i-th group model input amendment, m is the number of input variable, y _i∈ R represents the i-th group model output sample, and n is sample size;

(2) initialization APSO algorithm: each particle p (i) comprises m+3 variable, comprising m input variable x _icorresponding order d _i, 1 order d that output variable y is corresponding _yand the penalty factor of support vector regression SVMR model and nuclear parameter σ;

(3) input data are become nonlinear auto regressive moving average framework NARMAX model structure with output data preparation: the input variable x comprised with particle p (i) _iorder d _iwith the order d of output variable y _y, input data are become NARMAX model structure with output data preparation, and input X (t) and output Y (t) that then calculate soft-sensing model are as follows:

$\left\{\begin{array}{c}X\left(t\right)=[x_1(t-T),...,x_1(t-d_{1}T),...,x_m(t-T),...,x_m(t-d_{m}T),y(t-T),...,y(t-d_{y}T)]\\ Y\left(t\right)=y\left(t\right)\end{array}\right.$

In above formula, T refers to the sampling period of collecting measuring point data; X (t) is mode input amount corresponding to sampling instant t, comprises m input variable x _icontinuous d before sampling instant t _ithe quantity of state in individual sampling period, and the continuous d of measured y before sampling instant t _ythe quantity of state in individual sampling period; Y (t) is model output corresponding to sampling instant t, is the quantity of state of measured y when sampling instant t;

(4) SVMR model initialization: the penalty factor comprised with particle p (i) and nuclear parameter σ adopt gaussian kernel function and RBF to arrange SVMR model:

$Y\left(t\right)=\underset{i=1}{\overset{l}{\mathrm{\&Sigma;}}}({\mathrm{\&alpha;}}_i^{*}-{\mathrm{\&alpha;}}_i)\exp \{-\frac{{\left|\right|X\left(t\right)-X_i^{\mathrm{sv}}\left|\right|}^2}{2{\mathrm{\&sigma;}}^2}\}+b$

In above formula, α _ibe support vector regression parameter with b, for support vector, l is the number of support vector, and σ is RBF nuclear parameter;

(5) carry out SVMR training and matching, calculate particle ideal adaptation degree f _i, retain optimum individual: using X (t) as SVMR mode input, Y (t) exports as SVMR model, using before the sample set after arranging 3/4 as training sample, rear 1/4 as forecast sample, carries out training and the matching of model; With forecast sample Deviation Indices for fitness function, calculate particle ideal adaptation degree f _i, and to obtain colony's optimal-adaptive degree be f _mand retain its particle position Pm, so the fitness function of particle p (i) is:

$f\left(p_i\right)=\frac{1}{n}\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{(y_j-{\hat{y}}_j)}^2$

In above formula, n is predicted time sequence length, y _jwith be respectively measured value and the model predication value in a jth moment;

(6) optimizing convergence judges: if f _m<f _e, wherein f _efor expecting convergence threshold, work as k=K _maxtime, then optimizing stops, and performs step (9); Otherwise, make k=k+1, perform step (7); Wherein, k is current iteration number of times, K _maxfor maximum iteration time, the Search Range of order is [d _min, d _max];

(7) calculate particle colony average fitness, particle colony is divided into three subgroups: in kth time iteration, the fitness value of particle p (i) is f _i ^k, the fitness value of colony's optimal particle is f _m; Population average fitness is fitness value is better than particle fitness be averaging and obtain according to particle ideal adaptation degree f _i ^kbe divided into local optimal searching, balance optimizing, global optimizing three subgroups, calculated inertia weight w and the Studying factors c of its correspondence respectively ₁, c ₂;

(8) the individual speed of particle and location updating: the inertia weight w obtained according to step (7) and Studying factors c ₁, c ₂, to the individual speed V of particle _iwith position P _icarry out calculating to upgrade;

(9) export optimal result, complete the foundation of dynamic soft sensor model: export f _mthe d that corresponding optimum individual comprises _i, d _yand penalty factor and nuclear parameter σ, and set up final dynamic soft measuring SVMR model using its parameter as setup parameter;

(10) realization of dynamic soft measuring: for the sample of new collection arbitrarily, after first carrying out pre-service to it, the dynamic soft measuring SVMR model that input step 9 obtains, just can obtain corresponding hard measurement and export.

2. station boiler NO according to claim 1 _xdischarge dynamic soft-measuring method, is characterized in that: in described step (1), chooses the NO before furnace of power-plant boilers afterbody SCR reactor _xinlet concentration is as output quantity y _i, choose in boiler controller system dynamic running process NO _xdischarge the combustion system had an impact and run variable as input variable x _i; And original sample for the dynamically operating condition data continuously of the large and variable load operation of coverage in DCS system.

3. station boiler NO according to claim 1 _xdischarge dynamic soft-measuring method, is characterized in that: to original sample in described step (1) the pretreated detailed process of employing normalization is:

The original value of each input data is zoomed in designation area by following formula:

$x^{\&prime;}=\frac{x-\min \left(x\right)}{\max \left(x\right)-\min \left(x\right)}$

Wherein, x is the value before normalization, and x ' is the value after normalization, and max (x) and min (x) is respectively maximal value between normalization back zone and minimum value;

Because under each operating mode, flue gas oxygen content is inconsistent, for ease of comparing NO _xvalue, according to following formula, by the NO recorded under each operating mode _xconcentration conversion is to the value under 6% oxygen amount:

${\mathrm{\&rho;}}^{\&prime;}\left(\mathrm{NOx}\right)=\mathrm{\&rho;}\left(\mathrm{NOx}\right)*\frac{21-{\mathrm{\&rho;}}^{\&prime;}\left(O_2\right)}{21-\mathrm{\&rho;}\left(O_2\right)}$

Wherein, ρ ' (NOx) and ρ ' (O ₂) represent the NO after converting respectively _xconcentration and O ₂concentration, ρ (NOx) and ρ (O ₂) represent the NO of actual measurement respectively _xconcentration and O ₂concentration.

4. station boiler NO according to claim 1 _xdischarge dynamic soft-measuring method, is characterized in that: according to particle ideal adaptation degree f in described step (7) _i ^kthe process of three subgroups population is divided into be:

(7.1) calculate particle colony average fitness, the fitness value then setting particle p (i) in kth time iteration is f _i ^k, the fitness value of colony's optimal particle is f _m, particle colony average fitness value is

(7.2) fitness value is better than particle fitness average and obtain definition be used for evaluating the prematurity convergence degree of population, △ less explanation population body is more tending towards Premature Convergence;

(7.3) last according to particle ideal adaptation degree f _i ^kbe divided into following three classes:

If 1. then show that this particle is good particle in colony, be divided into local optimal searching subgroup, get less inertia weight w by following formula, make it change among a small circle, get larger c simultaneously ₁less c ₂, to strengthen the individual learning ability of self of particle and the local optimal searching ability of enhanced particles:

$w=w-(w-w_{\min })\&CenterDot;\left|\frac{f_i^k-f_{\mathrm{avgb}}^k}{f_m-f_{\mathrm{avgb}}^k}\right|$

$c_1=c_{1\min }+{(c}_{1\max }-c_{1\min })\&CenterDot;\left|\frac{f_i^k-f_{\mathrm{avgb}}^k}{f_m-f_{\mathrm{avgb}}^k}\right|$

$c_2=c_{2\max }-{(c}_{2\max }-c_{2\min })\&CenterDot;\left|\frac{f_i^k-f_{\mathrm{avgv}}^k}{f_m-f_{\mathrm{avgb}}^k}\right|$

If 2. then show that this particle is particle general in colony, be divided into balance optimizing subgroup, there is good global optimizing ability and local optimal searching ability simultaneously, maintain inertia weight w constant, and get c ₁=c ₂=1, the learning ability of equilibrium particle self cognition and social recognition;

If 3. then show that this particle is particle poor in colony, be divided into global optimizing subgroup, get larger w, to increase amplitude of variation, the ability of searching optimum of enhanced particles colony, gets less c simultaneously ₁larger c ₂, social learning's ability of enhanced particles, adjustable strategies carries out according to following formula:

$w=1.5-\frac{1}{1+k_1\&CenterDot;\exp (-k_2\&CenterDot;\mathrm{\&Delta;})}$

$c_1=c_{1\max }-\frac{1}{2}{(c}_{1\max }-c_{1\min })\&CenterDot;\left|\frac{f_i^k-f_m}{f_{\mathrm{avg}}^k-f_m}\right|$

$c_2=c_{2\min }+\frac{1}{2}{(c}_{2\max }-c_{2\min })\&CenterDot;\left|\frac{f_i^k-f_m}{f_{\mathrm{avg}}^k-f_m}\right|$

Wherein, c _1maxand c _2maxbe respectively cognitive learning factor c ₁with social learning factor c ₂the maximal value of the span of setting, c _1minand c _2minfor the minimum value of its span; W is the initial set value of inertia weight, w _minfor the setting minimum value of inertia weight.