A kind of thermal power plant NOx prediction model characteristics of variables extracting method based on method of comparison
Technical field
The present invention relates to thermal power plant NOx prediction model characteristics of variables extracting method technical field, it is specially a kind of based on pair
Than the thermal power plant NOx prediction model characteristics of variables extracting method of method.
Background technique
Nitrogen oxides (mainly NO and a small amount of NO2, it is referred to as NOX) it is that the main atmosphere that combustion of fossil fuel generates is dirty
One of object is contaminated, data show that Thermal Power Generation Industry is still NO according to investigationsXThe most industry of discharge amount.The weight protected with Environmental
Depending on requirement of the new environmental regulation to the power generation feature of environmental protection is also further severe, NOXDischarge amount become measure boiler combustion superiority and inferiority
Key index, be boiler whether green operation judging basis.Therefore NOXAccurately measuring for discharge amount is particularly important, however
NOXDischarge amount influence factor has the characteristics that coupling is strong, non-linear, faces these features, and traditional sensors have been unable to get
Effect is used.Hard measurement is more difficult in modern complex process industry or even can not be by the effective of hardware on-line checking parameter real-time estimation
Means are based on auxiliary variable and leading variable using the procedure parameter that can be measured or be capable of precise measurement as auxiliary variable
Mathematical model, realize the on-line prediction of procedure parameter that conventional instrument is unable to measure.Therefore flexible measurement method pair is generally utilized
NOXDischarge amount realization accurately measures.
Accurate Input variable selection is the basis that soft-sensing model is established.The selection of input variable will directly influence
Measurement accuracy, complexity and the arithmetic speed of soft-sensing model.Influence NOXThe factor of discharge amount is sufficiently complex: SOFA is opened
Degree, flue gas flow, flue gas temperature of hearth outlet, total coal amount, total blast volume etc. all can be to NOXConcentration of emission has an impact.But it is modeling
In the process if all considering these factors, it will increase information redundancy, keep model excessively complicated.Therefore, to input variable
Carrying out precisely screening is particularly important.
In recent years, to thermal power plant NOXAuxiliary variable screening technique in soft-sensing model has Principal Component Analysis and is believing
Cease the mutual information screening method etc. on the basis of entropy and Mutual Information Theory.Classical principal component analytical method is calculated between each feature
Correlation, but the nonlinear correlation between variable can not be assessed.And mutual information then can be used for measuring between two variables mutually according to
Bad degree of strength, is not limited to linear correlation.Due to influencing NO in boiler running processXThe factor of generation is very more,
Formation mechenism is also sufficiently complex, unobvious using a kind of method characteristic extraction effect.In consideration of it, the invention proposes one kind to be based on
The method that mutual information is combined with principal component analysis, screens input variable, can be more by the combination of two methods
Precisely comprehensively input variable feature is extracted.
Summary of the invention
The purpose of the present invention is to provide a kind of thermal power plant NOx prediction model characteristics of variables extraction side based on method of comparison
Method, to solve the problems mentioned in the above background technology.
To achieve the above object, the invention provides the following technical scheme:
A kind of thermal power plant NOx prediction model characteristics of variables extracting method based on method of comparison, comprising the following steps:
S1 the design data for) collecting target boiler carries out mechanism point in conjunction with combustion process of the actual motion environment to coal dust
Analysis, finding out influences NOXThe principal element of discharge amount, determines the input variable of model, and exports a certain amount of historical data, carries out
Denoising, filter preprocessing.Determining input variable includes: 1~5 layer of SOFA aperture, the escaping of ammonia, spray ammonia flow, ammonia flow valve
Aperture, pipeline flow regulating valve setting value, regulating valve regulating command, reactor inlet pressure, flue gas flow, furnace outlet flue gas
Temperature, total coal amount, total blast volume, unit meet air preheater outlet Secondary Air temperature, A~F feeder instantaneous flow and coal-grinding
Wind flow of machine entrance and outlet temperature, totally 37;Output variable: SCR inlet NOXConcentration, totally 1.
7 days varying load condition operation datas of the unit are exported from SIS, the sampling interval is 1 minute, totally 500000 groups of numbers
According to.
It follows Rye to denoise up to criterion, mathematical notation is as follows:
To measured carry out equal precision measurement, x is independently obtained1,x2,......,xnCalculate its arithmetic average x and residue
Error vi=xi- x (i=1,2 ... ... n), and calculates standard deviation according to Bayside formula:
If some measured value xbResidual error vb(1≤b≤n) meets | vb|=| xb- x | 3 σ of > then thinks xbBe containing
The bad value of gross error should be rejected.
The method that filtering uses mean filter, is specifically expressed as follows:
N number of sampled value is continuously taken to carry out arithmetic average operation:xiData are acquired for variable.
N=3 in the present invention.
S2 the data that S1 has been handled) are based on mutual information to screen, NO is selected using S-MIFS algorithmXConcentration influence compared with
Big input variable;
S3) data for obtaining S2 carry out principal component analysis, select the ingredient that accumulative contribution reaches 95%;
Further, S2 the following steps are included:
The final evaluation function of S-MIFS algorithm is obtained first:
In above formula, fi∈ F is variable to be selected, and c is leading variable, Sj∈ S is to have selected variable.β is penalty factor, | S | it is S
Norm selected works have closed S and have selected variable number in other words.
S2-1, Initialize installation, F are variables set to be selected, and S is to have selected variables set (emptying),cFor leading variable;
S2-2, f is calculatediWith the mutual information between leading variable c, the corresponding variable f of maximum value is taken outiIt is stored in S;It is counted
Be expressed as follows:
Wherein p (fi) it is fiMarginal probability distribution, p (c) be c marginal probability distribution, p (fi, c) and it is fiWith the joint of c
Probability distribution;
S2-3, greed search, recycle following steps, until searching K variable of destination number.
(1) candidate variables and mutual information has been selected between variable to calculate, has calculated all combination (fi,Sj) between mutual information I (fi;
Sj), Sj∈ S, fi∈F;
(2) next variable is selected according to evaluation function, takes out fiIt is stored in S.
The input variable subset that S2-4, output have been selected.
Further, S3 the following steps are included:
S3-1, initial data is formed to matrix X by rows;
S3-2, data normalization is carried out to X, becomes zero its mean value;
S3-3, the covariance matrix C for seeking X;
S3-4, feature vector is pressed into the descending arrangement of characteristic value, k form matrix P by row before taking;
S3-5, pass through calculating Y=PX, obtain data Y after dimensionality reduction;
S3-6, the contribution rate for calculating each characteristic root.
Compared with prior art, invention can be used for thermal power plant NOx prediction model input variable feature extraction, energy
Enough dimensions that input variable is reduced under the premise of slightly reducing precision, it is time-consuming and applied widely to reduce calculating.
Detailed description of the invention
Fig. 1 is the flow chart of the method for the present invention;
Fig. 2 is the flow chart of S-MIFS algorithm in mutual information;
Fig. 3 is the flow chart of principal component analysis (PCA);
Fig. 4 is to carry out NO after mutual information screening and principal component analysisXPredicted value compared with actual value figure.
Specific embodiment
It should be noted that in the absence of conflict, the feature in embodiment and embodiment in the present invention can phase
Mutually combination.
The technical solution of the patent is explained in further detail With reference to embodiment.
It is extracted the present invention provides a kind of based on mutual information and the thermal power plant NOx prediction model characteristics of variables of principal component analysis
Method.This method extracts thermal power plant NO based on the algorithm that mutual information is combined with principal component analysisXPrediction model input variable
Feature.It is fast with speed, the advantages that precision is high, and generalization ability is strong.
The present invention is with SCR inlet NOXFor the feature extraction of concentration soft-sensing input variable.It is true according to Analysis on Mechanism first
It is fixed and rings NOXThen the principal element of discharge amount acquires a certain amount of historical data, pre-processes to data, according to processing
Data later are screened by mutual information, then the input variable after screening is carried out principal component analysis, retain accumulative tribute
Offer reach 95% ingredient, finally be retained as being allocated as input, NOXConcentration is sent into neural network and is trained as output,
NO after being screenedXThe predicted value of concentration and the root-mean-square error of actual value, verifying mutual information with this is slightly reducing precision
In the case where effectively extract the feature of input variable.
Embodiment 1
Please refer to Fig. 1~4, a kind of thermal power plant NOx prediction model characteristics of variables extracting method based on method of comparison, including with
Lower step:
S1 the design data for) collecting target boiler carries out mechanism point in conjunction with combustion process of the actual motion environment to coal dust
Analysis, finding out influences NOXThe principal element of discharge amount, determines the input variable of model, and exports a certain amount of historical data, carries out
Denoising, filter preprocessing.Determining input variable includes: 1~5 layer of SOFA aperture, the escaping of ammonia, spray ammonia flow, ammonia flow valve
Aperture, pipeline flow regulating valve setting value, regulating valve regulating command, reactor inlet pressure, flue gas flow, furnace outlet flue gas
Temperature, total coal amount, total blast volume, unit meet air preheater outlet Secondary Air temperature, A~F feeder instantaneous flow and coal-grinding
Wind flow of machine entrance and outlet temperature, totally 37;Output variable: SCR inlet NOXConcentration, totally 1.
7 days varying load condition operation datas of the unit are exported from SIS, the sampling interval is 1 minute, totally 500000 groups of numbers
According to.
It follows Rye to denoise up to criterion, mathematical notation is as follows:
To measured carry out equal precision measurement, x is independently obtained1,x2,......,xnCalculate its arithmetic average x and residue
Error vi=xi- x (i=1,2 ... ... n), and calculates standard deviation according to Bayside formula:
If some measured value xbResidual error vb(1≤b≤n) meets | vb|=| xb- x | 3 σ of > then thinks xbBe containing
The bad value of gross error should be rejected.
The method that filtering uses mean filter, is specifically expressed as follows:
N number of sampled value is continuously taken to carry out arithmetic average operation:xiData are acquired for variable.
N=3 in the present invention.
S2 the data that S1 has been handled) are based on mutual information to screen, NO is selected using S-MIFS algorithmXConcentration influence compared with
Big input variable;
S3) data for obtaining S2 carry out principal component analysis, select the ingredient that accumulative contribution reaches 95%;
S4) result for obtaining S3 is divided into two groups, is trained in one group of feeding neural network, another group of feeding nerve net
It is tested in network;
S5 the output valve of neural network and the root-mean-square error of former test value) are calculated, being verified mutual information with this can
Effective screening input variable.Root-mean-square error RMSE mathematical notation is as follows:
In formulaFor neural network output valve, yiFor actual value.
Embodiment 2
The present embodiment is being further described on the basis of embodiment 1, S2 the following steps are included:
The final evaluation function of S-MIFS algorithm is obtained first:
In above formula, fi∈ F is variable to be selected, and c is leading variable, Sj∈ S is to have selected variable.β is penalty factor, | S | it is S
Norm selected works have closed S and have selected variable number in other words.
S2-1, Initialize installation, F are variables set to be selected, and S is to have selected variables set (emptying), and c is leading variable;
S2-2, f is calculatediWith the mutual information between leading variable c, the corresponding variable f of maximum value is taken outiIt is stored in S;It is counted
Be expressed as follows:
Wherein p (fi) it is fiMarginal probability distribution, p (c) be c marginal probability distribution, p (fi, c) and it is fiWith the joint of c
Probability distribution;
S2-3, greed search, recycle following steps, until searching K variable of destination number.
(1) candidate variables and mutual information has been selected between variable to calculate, has calculated all combination (fi,Sj) between mutual information I (fi;
Sj), Sj∈ S, fi∈F;
(2) next variable is selected according to evaluation function, takes out fiIt is stored in S.
The input variable subset that S2-4, output have been selected.
Embodiment 3
The present embodiment is being further described on the basis of embodiment 1, S3 the following steps are included:
S3-1, initial data is formed to matrix X by rows;
S3-2, data normalization is carried out to X, becomes zero its mean value;
S3-3, the covariance matrix C for seeking X;
S3-4, feature vector is pressed into the descending arrangement of characteristic value, k form matrix P by row before taking;
S3-5, pass through calculating Y=PX, obtain data Y after dimensionality reduction;
S3-6, the contribution rate for calculating each characteristic root.
Pass through the NO after screening is calculated of embodiment 1- embodiment 3XThe root mean square of concentration actual value and test value misses
Difference is 3.3204, is slightly reduced than precision before screening, but the used time obviously compares reduction about 30% before.The method of the present invention is in thermoelectricity
Factory's SCR denitration system entrance NOx mode input characteristics of variables has good effect in extracting really.And the method can be applied to institute
Have in thermal power plant's NOx model.
The innovation of the invention consists in that can be used for thermal power plant NOx prediction model input variable feature extraction, it can be slightly
The dimension that input variable is reduced under the premise of reducing precision, reduces and calculates time-consuming, and is applied widely, and utilizes two methods
It compares cross-referenced, guarantees the accuracy of result.
The preferred embodiment of the patent is described in detail above, but this patent is not limited to above-mentioned embodiment party
Formula within the knowledge of one of ordinary skill in the art can also be under the premise of not departing from this patent objective
Various changes can be made.