CN106934209A - A kind of coal fired power plant flue gas oxygen content on-line prediction method - Google Patents

Abstract

The invention provides a kind of coal fired power plant flue gas oxygen content on-line prediction method, including；Step 1), obtain coal fired power plant historical data sample as off-line model training set；Step 2), set up offline forecast model, and recognized based on training sample set pair off-line model parameter；Step 3), the online real time data for obtaining operation, calculate prediction output using offline forecast model；Step 4), calculate predicated error, do not dealt with if predicated error meets requirement, cumulative number plus 1 if predicated error is unsatisfactory for requiring, if cumulative number is not reaching to n, return to step 2), off-line model parameter is recognized, if cumulative number reaches n, on-line amending forecast model.The on-line prediction method of the coal fired power plant flue gas oxygen content that the present invention is provided, improves modeling accuracy, reduces computation complexity.

Description

A kind of coal fired power plant flue gas oxygen content on-line prediction method

Technical field

The present invention relates to a kind of on-line prediction method of coal fired power plant flue gas oxygen content, belong to boiler heat power engineering and calculate Machine monitoring interleaving techniques field, specific forecast model is a local model networks for the online updating based on fuzzy clustering, energy The flue gas oxygen content output of enough real-time estimate subsequent times under the conditions of full working scope.

Background technology

Coal fired power plant flue gas oxygen content is the oxygen content in the flue gas in boiler air preheater exit, is reactive combustion situation Important parameter.The basis that on-line prediction model is the predictive control algorithm for further realizing flue gas oxygen content is set up, also can be pre- The combustion position of future time instance is surveyed for production operation provides suggestion, it is real generally speaking to set up flue gas oxygen content on-line prediction model The important foundation of existing burning optimization.Flue gas oxygen content is performance of the ratio on the flue gas of end of blowing and deliver coal, and reflects boiler The size of middle excess air coefficient.Flue gas oxygen content is high, means that excess air coefficient is big, and the sufficient burning of air quantity is abundant but right The flue gas loss answered is big；Another aspect flue gas oxygen content is low then to represent that excess air coefficient is smaller, incomplete combustion, corresponding row Cigarette loss is small.Therefore the tendency of flue gas oxygen content is predicted, is conducive to grasping in real time the combustion position variation tendency of boiler internal；Build Vertical forecast model can be further used for the PREDICTIVE CONTROL of flue gas oxygen content, to realize burning optimization.

The combustion system of coal fired power plant is a high complexity system, and the factor and variable being directed to are more, time delay phenomenon Substantially.Combustion process is influenceed greatly by load change, the flue gas oxygen content performance under different operating modes (load) in actual production process Go out different properties.Therefore the model of flue gas oxygen content is nonlinear time-varying, and related to operating mode.

The predictive model algorithm both at home and abroad to flue gas oxygen content or other boiler output variables mainly includes three kinds at present： The first is the modeling of mechanism method, and second is the intelligent modeling based on black-box model, and the third is hybrid modeling, bonding mechanism point Analysis and the modeling method of intelligent algorithm.Traditional thermal parameter modeling is using classical mechanism method, and model is simple, but on the one hand pre- Survey error big, another aspect mechanism method is directed to a certain class or a certain boiler, rely on the judgement of engineer, model generalization It is not high.At present, the Modeling Research on flue gas oxygen content largely employs LSSVM, and the improved model of neutral net, this two class is calculated Method is learnt by a large amount of historical datas, and predict the outcome high precision, but it is slow to show computationally intensive on-line operation Problem.

The content of the invention

For the difficult point of the not enough and flue gas oxygen content modeling of existing on-line prediction algorithm, the invention provides a kind of coal-fired Generating plant flue gas oxygen content on-line prediction method, by on-line prediction modeling algorithm, overcome precision in the prior art it is low or meter The long problem of evaluation time, to ensure the effective prediction in the range of full working scope.

To reach above-mentioned purpose, the technical solution adopted in the present invention is as follows：

A kind of coal fired power plant flue gas oxygen content on-line prediction method, is realized by modeling, including step is as follows：

Step 1), obtain coal fired power plant historical data sample as off-line model training set；

Step 2), set up offline forecast model, and recognized based on training sample set pair off-line model parameter；

Step 3), the online real time data for obtaining operation, calculate prediction output using offline forecast model；

Step 4), calculate predicated error, if predicated error meet require if do not deal with, if predicated error is unsatisfactory for It is required that then cumulative number adds 1, and if cumulative number is not reaching to n, return to step 2), off-line model parameter is recognized, if When cumulative number reaches n, then on-line amending forecast model, n >=5.

Compared with prior art, the present invention has following beneficial effect：

1st, based under full load condition in actual motion data set up, to multi-state in the case of have universality；

2nd, being capable of online updating model, it is ensured that the precision of model in running；

3rd, make use of the method for local model networks to be modeled compared to mechanism method, can more approach actual nonlinear model；

4th, the forecast model set up is produced based on data-driven, it is characterized in that give the structure framework of model with Method, the model of different units can be set up according to different input datas, and the unit similar to mechanism structure has good Generalization；

5th, using real time data sample on-line amending cluster centre, center and simplified algorithm based on the cluster for updating are online Amendment Gaussian width.

Brief description of the drawings

The detailed description made to non-limiting example with reference to the following drawings by reading, further feature of the invention, Objects and advantages will become more apparent upon：

Fig. 1 is off-line model identification algorithm schematic diagram；

Fig. 2 is on-line prediction algorithm schematic diagram；

Fig. 3 is offline prediction effect comparison diagram；

Fig. 4 is on-line prediction effect contrast figure.

Specific embodiment

With reference to specific embodiment, the present invention is described in detail.Following examples will be helpful to the technology of this area Personnel further understand the present invention, but the invention is not limited in any way.It should be pointed out that to the ordinary skill of this area For personnel, without departing from the inventive concept of the premise, some changes and improvements can also be made.These belong to the present invention Protection domain.

Shown in Fig. 1, Fig. 2, the present invention proposes a kind of multi-model online updating algorithm based on fuzzy C-means clustering to predict Flue gas oxygen content, the main thought of the method is to be coupled multiple submodels to approach the non-thread of complexity by fuzzy C-means clustering Property model and online updating cluster centre to realize on-line prediction function, mainly include two parts：Off-line model parameter identification With online updating algorithm.Off-line model identification mainly has following four step：First using flue gas oxygen content as the burning for exporting Model is expressed as local model networks structure, then recognizes basic function parameter using Fuzzy C-Means Cluster Algorithm, next uses The method of Subspace Identification recognizes the model parameter of all part (son) models, is finally calculated using Fuzzy C-Means Cluster Algorithm Local (son) the model construction world model of subordinated-degree matrix connection for obtaining.Online updating part mainly includes following two steps Suddenly：Judge whether conditions present meets update condition and update cluster centre vector sum Gauss distance using new sample data.

It is specific as follows：

(1), the training data of screening off-line model identification.Continuous N groups history data is chosen as sample sequenceLoad data covering full working scope scope wherein corresponding to sample data group.

(2) model framework of flue gas oxygen content, is builtWherein y (k+1) is flue gas oxygen content One-step prediction value,Represent i-th weighting function of partial model, M_iK () represents local submodel, L is the individual of partial model Number.

(3) local (son) model, is represented by the method for state space

(4) local (son) Number of Models L, model accuracy requirement ε and on-line study constant η, are given.

(5) parameter of offline world model, is recognized using FCM-Subspace (fuzzy C-means clustering), original is calculated Part (son) the model center c of beginning_i(i=1,2 ..., L) and Gaussian width s_i。

(6) weighting function, is calculatedWherein s_iIt is Gaussian function distance, φ= φ (1), φ (2) ..., φ (N } represent characteristic vector, c_iRepresent i-th cluster centre of partial model.

(7) prediction output y (k+1), is finally calculated.

(8) predicated error, is calculatedExported using Error Compensated Prediction

(9) if, predicated error be more than ε, count=count+1, otherwise count=0；

(10) if, cumulative number count≤5 return to step 6；

(11) the cluster centre vector of local (son) model, is updated：

(12) Gaussian function distance, is updated

A=c_i(k)-c_l(k),

(13), calculateAnd update sytem matrix

(14), k=k+1 and step 7 is returned to.

Fig. 3, Fig. 4 represent training set output with reality output comparison diagram, on-line prediction output and reality output contrast respectively Figure.Fig. 3 training sets output comparison diagram in solid line be reality output, dotted line be model output, training set models fitting degree compared with Height, can track reality output change.Solid line is on-line prediction output in Fig. 4 test proof diagrams, and dotted line is reality output, prediction Value is smaller with actual value error, and trend is close, and it is effective that checking obtains on-line prediction model proposed by the invention.

The present invention：

(1), using the multi-model process based on fuzzy C-means clustering and Subspace Identification, by the oxygen under the conditions of full working scope Amount model is showed.

(2), online updating forecast model, is capable of the dynamic change of real-time tracking system, improves the online of flue gas oxygen content Modeling accuracy.

(3), online updating is based on original result of calculation and calculates simplified conclusion, and algorithm is relatively easy, can be quick Realize in line computation on ground, it is ensured that the efficiency of on-line prediction.

(4), on-line prediction system of the invention is a system based on real-time running data, and required data are DCS Existing measuring point data, it is easy to obtain and need not additionally increase hardware setting, the method universality is strong.

Flue gas oxygen content is really that the another of excess air coefficient in boiler characterizes variable, therefore with flue gas oxygen content most Related is the coal amount and air capacity of entrance, has selected secondary air flow, absorbing quantity, coal-supplying amount to contain as flue gas with reference to actual measuring point The input of oxygen amount forecast model.One embodiment of the invention have chosen Zhejiang Jiaxing Power Plant service data as modeling and check Data.Specific time window is 0 point to 2015 0 point of January 12 day of January 4 day in 2015, and data sampling time is 1 minute.Utilize Flue gas oxygen content predictive model algorithm proposed by the present invention demonstrates offline and on-line Algorithm the precision of forecast model.Adopt in practice With the data modeling of 2 days, the data verification of 1 day.And the method for updating once offline basic model in 2 days is taken in actual production Ensure model continuous and effective.

Specific embodiment of the invention is described above.It is to be appreciated that the invention is not limited in above-mentioned Particular implementation, those skilled in the art can within the scope of the claims make a variety of changes or change, this not shadow Sound substance of the invention.In the case where not conflicting, feature in embodiments herein and embodiment can any phase Mutually combination.

Claims (7)

1. a kind of coal fired power plant flue gas oxygen content on-line prediction method, it is characterised in that realized by modeling, including step is such as Under：

Step 1), obtain coal fired power plant historical data sample as off-line model training set；

Step 2), set up offline forecast model, and recognized based on training sample set pair off-line model parameter；

Step 3), the online real time data for obtaining operation, calculate prediction output using offline forecast model；

Step 4), calculate predicated error, if predicated error meet require if do not deal with, if predicated error be unsatisfactory for require Then cumulative number adds 1, if cumulative number is not reaching to n, return to step 2) and, off-line model parameter is recognized, if accumulative When number of times reaches n, then on-line amending forecast model, n >=5.

2. coal fired power plant flue gas oxygen content on-line prediction method according to claim 1, it is characterised in that step 1) in Historical data includes coal-supplying amount, secondary air flow, absorbing quantity and flue gas oxygen content.

3. coal fired power plant flue gas oxygen content on-line prediction method according to claim 1, it is characterised in that step 1) choose Continuous N group history datas are used as sample sequenceLoad data covering wherein corresponding to sample data set is complete Condition range.

4. coal fired power plant flue gas oxygen content on-line prediction method according to claim 1, it is characterised in that step 2) pass through Offline forecast model is set up with reference to Multi-model MPCA and subspace state space system identification.

5. coal fired power plant flue gas oxygen content on-line prediction method according to claim 4, it is characterised in that set up offline pre- Survey model and specifically include following steps：

Step 51), build world model framework using local model networks, and adoption status space form represents local submodule Type, obtains a global state spatial model framework for time-varying；

Step 52), modeling initialization prepare, artificial screening part submodel number learns constant and precision prescribed；

Step 53), training set data is clustered, cluster number is equal to local submodel number, is calculated partial model Cluster centre；

Step 54), using the Gauss distance of input sample and cluster centre as the weight of the corresponding partial model of the sample, will be input into Sample builds Mixed design with weighting vector；

Step 55), exported based on Mixed design and history, using the method off-line identification forecast model of Subspace Identification.

6. coal fired power plant flue gas oxygen content on-line prediction method according to claim 5, it is characterised in that specific method is such as Under：

Build the model framework of flue gas oxygen contentWherein y (k+1) is flue gas oxygen content one-step prediction Value,Represent i-th weighting function of partial model, M_iK () represents local submodel, L is the number of partial model, is passed through The method of state space represents local submodel

Artificial screening part submodel number L, model accuracy requirement ε and on-line study constant η；

The parameter of offline world model is recognized using the method for fuzzy C-means clustering, is calculated in original local submodel Heart c_i(i=1,2 ..., L) and Gaussian width s_i；

Calculate weighting functionWherein s_iIt is Gaussian function distance, φ={ φ (1), φ (2) ..., φ (N } represent characteristic vector, c_iRepresent i-th cluster centre of partial model.

7. coal fired power plant flue gas oxygen content on-line prediction method according to claim 6, it is characterised in that step 4) it is online Amendment forecast model specific method is as follows：

Update the cluster centre vector of local (son) model：

Update Gaussian function distance：

$\begin{array}{c}{s_i}^2(k+1)=\left(\frac{1}{p}\underset{l=1}{\overset{p}{\Σ}}{\left(c_i\right(k+1)-c_l(k+1\left)\right)}^T\right(c_i\left(k+1\right)-c_l\left(k+1\right))\\ =(\frac{1}{p}\underset{l=1}{\overset{p}{\Σ}}{\[a+b\]}^T\[a+b\])\\ =\left(\frac{1}{p}\underset{l=1}{\overset{p}{\Σ}}\right(\left|\right|a|{|}^2+a^{T}b+b^{T}a+b^2)\\ =\frac{1}{p}\underset{i=1}{\overset{p}{\Σ}}\left(a^2\right)+\frac{1}{p}\underset{i=1}{\overset{p}{\Σ}}(a^{T}b+b^{T}a+b^2)\end{array}$

A=c_i(k)-c_l(k),

CalculateAnd update sytem matrix：

K=k+1 and return to step 3), calculate prediction output y (k+1).