CN110298142A - A kind of denitrating catalyst operational characteristic index prediction technique based on measured data - Google Patents

A kind of denitrating catalyst operational characteristic index prediction technique based on measured data Download PDF

Info

Publication number
CN110298142A
CN110298142A CN201910654885.XA CN201910654885A CN110298142A CN 110298142 A CN110298142 A CN 110298142A CN 201910654885 A CN201910654885 A CN 201910654885A CN 110298142 A CN110298142 A CN 110298142A
Authority
CN
China
Prior art keywords
sample
catalyst
matrix
operational characteristic
data
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN201910654885.XA
Other languages
Chinese (zh)
Other versions
CN110298142B (en
Inventor
庄柯
林正根
姚杰
金定强
吴碧君
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Guodian Environmental Protection Research Institute Co Ltd
Original Assignee
Guodian Environmental Protection Research Institute Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Guodian Environmental Protection Research Institute Co Ltd filed Critical Guodian Environmental Protection Research Institute Co Ltd
Priority to CN201910654885.XA priority Critical patent/CN110298142B/en
Publication of CN110298142A publication Critical patent/CN110298142A/en
Application granted granted Critical
Publication of CN110298142B publication Critical patent/CN110298142B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • G06F18/2135—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on approximation criteria, e.g. principal component analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
    • G06Q10/063—Operations research, analysis or management
    • G06Q10/0639—Performance analysis of employees; Performance analysis of enterprise or organisation operations
    • G06Q10/06393—Score-carding, benchmarking or key performance indicator [KPI] analysis

Abstract

The fresh honeycomb type denitrification catalyst denitration efficiency prediction technique based on measured data that the invention discloses a kind of.Mainly comprise the steps of: that step 1, sample collection collect the multiple batches of cellular catalyst of same producer and sample are made；Step 2, data collection, sample experiments room dimensional measurement, physical and chemical measurement, operational characteristic measurement obtain measurement data；Step 3, data processing carry out correlation analysis, principal component analysis and calculating to measurement data；Step 4, LMBP neural net model establishing；Step 5, similar sample operational characteristic index prediction.The present invention efficiently uses the existing a large amount of measured datas in laboratory, unknown operational characteristic achievement data is predicted using existing experimental data, it can be realized the look-ahead to sample operational characteristic before detecting, uncontrollable factor in the fresh honeycomb type denitrification catalyst operational characteristic assessment component in early warning laboratory, brings great convenience to laboratory process Characteristics Detection.

Description

A kind of denitrating catalyst operational characteristic index prediction technique based on measured data
Technical field
The present invention relates to field of denitration catalyst, and in particular to a kind of fresh cellular denitration based on measured data is urged Agent operational characteristic index prediction technique.
Background technique
Honeycomb type denitrification catalyst is widely used in the flue gas denitrification system of coal-burning power plant, and these catalyst exist Before use, requiring to carry out control to its quality, most of quality control is carried out by catalyst manufacturer man, But for power plant, obtained report is detected independently of the third party of catalyst manufacturer man, in contrast, is had more Good confidence level and acceptance.Catalyst is commented as the independent third party for carrying out quality evaluation to catalyst in laboratory Valence is geometrical property evaluation, Evaluation of Physiochemical Properties, operational characteristic evaluation in terms of three mainly in accordance with corresponding professional standard Come carry out.The evaluation of operational characteristic, mainly to the denitration efficiency of catalyst and active evaluation, which be unable to do without evaluation Long period of experiments on device (also referred to as detection device, platform), time and effort consuming, and do not pay attention to slightly to go out in experimentation It is wrong.
Publication No. CN 203479783 U, CN 203616297 U, CN 108426975 the Chinese patent of A give The design scheme of the device for detecting activity of denitrating catalyst, publication number CN 106647285 A, CN 107202857 A scheme Different algorithms, which is respectively adopted, realizes the calculating of catalyst activity and denitration efficiency perhaps detection but aforementioned schemes or side Method does not account for the geometrical characteristic of denitrating catalyst and physicochemical characteristics have a certain impact to activity and denitration efficiency.It is real Test test of the room referring to national standard GB/T 31587-2015 " honeycomb type flue gas denitration catalyst " 6.5.2.1 flue gas chapters and sections defined Flue gas condition detects the activity and denitration efficiency of catalyst.Since test flue gas condition immobilizes, on long terms, The geometrical property and physicochemical property of different catalyst samples influence the coupling of the activity of catalyst and denitration efficiency, must Certain statistical regularity can be so shown, the research of this regularity feature can be with guided laboratory to catalyst sample Detect work, there is important application value, but at present there is no relevant patent or document to this statistical rule into Row research and and then guiding experiment work.
Summary of the invention
The purpose of the present invention is to solve defects existing in the prior art, and catalyst work can be disclosed by providing one kind It is regular between skill characterisitic parameter and geometry, physical and chemical parameter, the new catalyst sample using the production of same production technology is carried out Operational characteristic index carries out quick, Accurate Prediction method.
In order to achieve the above object, the present invention provides a kind of fresh honeycomb type denitrification catalyst based on measured data Operational characteristic index prediction technique, comprising the following steps:
(1) sample collection: the catalyst to be measured being collected is the fresh honeycomb type denitrification catalyst of same producer production；
(2) data collection: carrying out geometrical property measurement, physicochemical property measurement, operational characteristic measurement to the sample of collection, Obtain geometrical property parameter, physicochemical property parameter, operational characteristic parameter；
(3) data processing: step (2) are collected into the sample set matrix M that data are constituted and are normalized, and to M square The submatrix that geometrical property parameter, physicochemical property parameter after battle array normalization are constituted carries out correlation analysis and principal component point Analysis；
(4) LMBP neural network model is established；
(5) model prediction:
For catalyst sample to be predicted, step (1)~(3) are repeated, the data obtained is applicable in the mind of step (4) foundation Through network model, the operational characteristic parameter of catalyst is obtained to the prediction result renormalization obtained.
Wherein, the specific method is as follows for data processing in step (3):
A, the data that step (2) is collected constitute sample set matrix M, and matrix M is normalized to matrix MZ, and MZ is divided into two Submatrix X_MZ, Y_MZ；Every a line of the matrix M is known as a sample, and each sample standard deviation is joined by catalyst geometrical property Number, physicochemical property parameter, operational characteristic parameter composition；Geometrical property after the submatrix X_MZ is normalized by Metzler matrix is joined Number, physicochemical property parameter are constituted, and the operational characteristic parameter after submatrix Y_MZ is normalized by Metzler matrix is constituted；
B, Pearson correlation coefficient analysis is carried out to X_MZ, and each column element of the correlation matrix obtained is asked Absolute value, takes the maximum value (note: being set to 01 in correlation matrix) of each column absolute value elements, and judges all column Whether maximum value is respectively less than 0.3, if so, data processing terminates, otherwise carries out principal component analysis in step c；
C, principal component analysis is carried out, number of principal components is chosen according to contribution rate of accumulative total, extracts principal component scores coefficient matrix, Restore principal component matrix X_RZ；
Wherein, the modeling process of LMBP neural network is as follows in step (4):
D, to gained submatrix X_RZ final after step (3) data processing or the submatrix of principal component analysis is needed not move through X_MZ and corresponding submatrix Y_MZ, is adopted in a like fashion, is divided into training set, verifying collection, test set；
E, three layers of LMBP neural network are established, input node of the element as network in matrix X_RZ or X_MZ, matrix Output node of the element of Y_MZ as network；
F, continue repeatedly to train network, and whether the coefficient R of training of judgement collection, verifying collection, test set is all larger than 0.995, if meeting the condition, network deconditioning, network modelling is completed, and otherwise continues to train network until meeting the condition.
More specifically, the present invention is based on the predictions of the fresh honeycomb type denitrification catalyst operational characteristic index of measured data Method the following steps are included:
Step 1, sample collection: the catalyst, for different catalysts manufacturer, used catalysis Agent production technology and technique are often not fully identical, though which results in the catalyst of different manufacturers in geometric parameter and Physical and chemical parameter is essentially identical or in the case where, is marked using " honeycomb type flue gas denitration catalyst " GB/T 31587-2015 Standard measures operational characteristic parameter (the specially denitration efficiency, activity) value obtained and also differs larger.And for same catalysis For agent manufacturer, what the production technology and technology of commercial catalyst were usually relatively fixed, it seldom will appear and hold Continuous the case where changing.So the enigmatic premise of sample collection is the different of the constant producer of same production technology or same a batch Secondary difference sample.
Step 2, data collection: the data are with " honeycomb type flue gas denitration catalyst " (GB/T 31587- 2015), " catalyst for denitrating flue gas chemical composition analysis method " (GB/T 31590-2015), " chemical reagent inductive coupling etc. from Daughter atom emission spectrometry general rule " (GB/T 23942-2009), " mercury injection method and gas determination of adsorption method solid discretion aperture Distribution and porosity part 1: mercury injection method " (GB/T 21650.1-2008), " gas absorption BET method measurement solid matter ratio Surface area " industries such as (GB/T 19587-2004) or national standard be detection foundation, it is examined in corresponding special equipment It measures out.Specifically, the data that measurement obtains include the aperture d of catalyst, interior wall thickness d1, outer wall thickness d2, section side length a, Main chemical compositions (TiO2、V2O5、Al2O3、WO3、SiO2, CaO etc.), microchemistry ingredient (K, Fe, Na, Mg, P, As etc.), Specific pore volume and most probable pore size, specific surface area, activity, denitration efficiency etc..The measurement data is constructed of in sample set matrix Every a line constitutes a sample, and each sample includes all metrical informations of corresponding single sample, several samples constitute sample This collection matrix.
Step 3, data processing: the sample set matrix is divided into submatrix X and submatrix Y1 and Y ', submatrix X by Geometric parameter, physical and chemical parameter are constituted, and submatrix Y1 is made of the denitration efficiency of catalyst, submatrix Y ' by catalyst activity It constitutes, the matrix composition of three can be written as such form:
In formula, j indicates the quantity of sample, and i indicates the sum of catalyst geometric parameter, physical and chemical parameter number, XijIndicate jth I-th of element of a catalyst sample.YjIndicate the denitration efficiency of j-th of catalyst sample.
It should be noted that sample set matrix is M, i.e. [sample set]=M.
Further, M is standardized and (is also called normalization), submatrix is also normalized simultaneously, the side of use Formula is zero-mean value standardization (also referred to as standard deviation standardization), and the matrix after X normalization is X_MZ, the matrix after Y1 normalization For Y_MZ.
Further, pearson correlation analysis is carried out to matrix X_MZ, the formula of foundation is:
Obtain correlation matrix CORR, (auto-correlation coefficient is 1 to the maximizings of each column of correlation matrix 0) value is set, all maximum values constitute row vector R_ROW and then show if there is value to fall in except (- 0.3 ,+0.3) range in R_ROW There is the information that partly overlaps in the element in submatrix X, principal component analysis need to be carried out to submatrix X_MZ, and choose accumulative contribution The principal component of rate > 0.85 calculates according to formula (NEW=ZX* principal component coefficient) and restores and obtain principal component matrix X_RZ；If Within the scope of all values all fall within (- 0.3 ,+0.3) in R_ROW, show the coherent element microfacies pass or low in submatrix X Correlation, then without principal component analysis.
Step 4 establishes LMBP network.If step (3) has carried out principal component analysis, use X_RZ and Y_MZ as LMBP nerve network input parameter and output parameter are modeled using three-layer neural network structure；If not carrying out principal component point Analysis, then select X_MZ and Y_MZ to establish three-layer neural network as the input and output parameter of LMBP neural network.
Further, in order to which the fitting effect to LMBP neural network controls, first X_RZ and corresponding Y_MZ (or: it X_MZ and corresponding Y_MZ) is divided according to a certain percentage, division mode presses training set: test set: verifying collection =X1:X2:X3It carries out, and X1:X2:X3=(60~70): (15~20): (15~20).During training neural network, judgement These three collection coefficient R it is whether equal > 0.995, if more than 0.995, be then no longer trained, if being not more than 0.995, then it is persistently trained, until three correlation coefficient value are all satisfied the requirement greater than 0.995, stopping network training.
Particularly, the selection of the node in hidden layer of LMBP neural network, according to formulaIt carries out, formula In, m indicates that the number of nodes of hidden layer, n indicate the number of nodes of input layer, and the value range of α is 1~10.
(5) model prediction.Catalyst sample is selected using step 1, geometric parameter, the reason of sample are obtained using step 2 Change parameter, using the data processing method tissue and processing data of step 3, obtains X_MZ ' or X_RZ ', input trains The step of the four LMBP networks obtained, obtain prediction denitration efficiency η, using prediction data η and press formula K=-AVln (1- The activity of catalyst η) is calculated, wherein K is activity, and AV is face velocity.
The present invention has the advantage that compared with prior art
1. the present invention is predicted from the statistical law of mass data using existing denitrating catalyst detection data Unknown denitrating catalyst data to be tested, compared to traditional catalyst activity and denitration efficiency measures or prediction technique, Quick and precisely, unnecessary working hour is eliminated.
2. the present invention can carry out operational characteristic survey in the geometrical property parameter for obtaining sample, after physicochemical property parameter Before examination, activity and denitration efficiency are predicted using established prediction model, effective early warning experiment test occurs uncontrollable Factor, prevent the control parameter in operational characteristic test process it is improper caused by greatly deviate exact value.
Detailed description of the invention
Fig. 1 is prediction technique flow chart of the present invention；
Fig. 2 is PCA-LMBP neural network structure figure in prediction technique of the present invention；
Fig. 3 is sample set A training related coefficient figure in prediction technique of the present invention；
Fig. 4 is sample set A denitration efficiency prediction result in prediction technique of the present invention.
Specific embodiment
As shown in Figure 1, the present invention is based on the fresh honeycomb type denitrification catalyst operational characteristic index prediction sides of measured data Specific step is as follows for method:
Step 1, sample collection: totally 70 groups of honeycomb type denitrification catalyst for collecting same producer's production.
Step 2, data collection:
The geometrical property and physicochemical property of each catalyst sample of collection are tested, following data is obtained:
1, main chemical compositions: 6 kinds of chemical component (TiO3、V2O5、WO3、Al2O3、 SiO2, CaO) content；
2, trace element detection: the content of 6 kinds of microelements (K, Na, Fe, Mg, P, As)；
3, specific pore volume and most probable pore size；
4, specific surface area.
And operational characteristic test is carried out to it, the denitration efficiency of catalyst sample is obtained, work is calculated by denitration efficiency Property.
In order to which the denitration efficiency (or activity) of the catalyst to same producer, different batches has a unified test Catalyst is cut into sectional dimension 150mm × 150mm, full-length 500mm and sample without obvious physical damnification by benchmark, Then it is detected by test flue gas condition shown in table 1.
Table 1 tests flue gas condition
Above data is summarized, a sample set matrix is constituted.
The sample set matrix is configured so that every a line of matrix represents the geometry, physics and chemistry and technique of single catalyst Feature, referred to as a sample.
Step 3, data processing:
By sample set matrix, it is divided into submatrix X and submatrix Y, submatrix X indicates geometry, physical and chemical test data, son Matrix Y indicates that the denitration efficiency detection data for corresponding to X, the matrix composition of three can be written as such form:
Y=[Y1 … Yj]T
Wherein, Xij(i=1,2 ..., 16) successively represent a sample of measured jth (j=1,2 ..., 70) geometry, Physical and chemical numerical value, Yj(j=1,2 ..., 70) represents the denitration efficiency value of j-th measured of sample.
Foundation after submatrix X is simplified with submatrix Y relationship, consideration are avoided establishing SCR denitration reaction mechanism and mathematical modulo The complexity of type can use BP neural network model and come after analog sample subset X simplifies to corresponding denitration efficiency subset Y's It influences, which has a small amount of application in denitration area research, but BP neural network has certain limitation, improved LMBP It is predictability that network (Levenberg-Marquardt BP neural network, LMBP), which then has more excellent simulation, Energy.
2 correlation matrix of table
In table 2, X1Represent aperture, X2Represent specific surface area, X3Represent specific pore volume, X4Represent most probable pore size, X5-X10According to It is secondary to represent main chemical compositions, X11-X16Successively represent trace chemical elements.
Then principal component analysis is carried out:
Principal component analysis is a kind of high dimensional data to be converted to several main components under the premise of losing less information Statistical analysis technique.In general, if research object includes multiple characteristic attributes, and there may be moulds between multiple characteristic attributes The correlation of paste, Principal Component Analysis can by matrixing will initial multiple characteristic attribute linear combinations at it is several mutually not Relevant assemblage characteristic index had not only disclosed the rule between object internal feature attribute in this way, but also the characteristic attribute of object had been added With simplification, to disclose most attributes of research object with least characteristic index.
For submatrix X, every a line XiDimension it is not fully identical, in order to eliminate the inconsistent influence of dimension, need handle Submatrix X is standardized as matrix ZX, and the standardized way taken herein is standard deviation standardization (also referred to as zero-mean value specification Change), the advantages of this method is can uniformly to convert the same magnitude for different magnitude of data, to realize between data Comparativity.Then principal component analysis is carried out through MATLAB software programming to ZX, obtains the accumulative variance of preceding 9 principal component of sample Contribution rate statistics is as shown in table 3 (since data volume is larger, normalized matrix ZX and its principal component coefficient matrix are simultaneously unlisted). Tie element of the contribution rate of accumulative total on 85% is usually chosen as ingredient is retained, therefore, preceding 8 components is selected and presses formula Main variables are calculated in (NEW=ZX* principal component coefficient), carry it into LMBP neural network later and carry out modeling and forecasting, The advantage of doing so is that the complexity of neural network can be substantially reduced, simplifies network structure, be conducive to network and quickly receive It holds back.
Principal component of 3 contribution rate of accumulative total of table greater than 85% counts
Step 4 establishes LMBP network.
Standard BP network algorithm, signal forward-propagating, error back propagation, theoretically, it can be forced with arbitrary accuracy Nearly Any Nonlinear Function can complete arbitrary n dimension to m with one 3 layers of BP network and tie up mapping, and structure is relatively simple.Behaviour On work, the main thought of the network is: input learning sample, is carried out using weight and deviation of the back-propagation algorithm to network Adjusting training repeatedly, make output vector (predicted vector) and Mean Vector as close as.But standard BP network exists easy The disadvantages of falling into minimal point and slow convergence rate, thus have more innovatory algorithm.It is improved compared to standard BP algorithm Gradient descent method and Gauss-Newton method are combined (Wu W, Wang J, Cheng M S er by LMBP algorithm al.Convergence analysis of online gradient method for BP neural networks[J] .Neural Networks, 2010,12 (24): 91-98), allowable error is searched for along the direction of deterioration, under normal gradients Modified weight amount is adjusted between drop method and Newton algorithm by adaptively, to make network fast convergence, improves network Generalization ability has the advantages that global convergence, speed are fast, capability of fitting is strong.
After being analyzed through PCA, new samples collection matrix XY (as shown in Equation 7) that sample set submatrix NEW and Y are collectively formed In, corresponding-a denitration efficiency Y of the total i main variables of jth rowj.Establish LMBP neural network, whole i main variables As the input node of neural network input layer, corresponding denitration efficiency YjOutput node as output layer.XY is by being randomly assigned Mode, 60 samples are divided into sample set A, remaining 10 sample is divided into sample set B.The NEW of sample set Aij(i=1, 2 ..., 8, j=1,2 ..., 60) input as LMBP network, corresponding YjAs output, network model is established (such as Fig. 2 institute Show)；Sample set B is used to examine the Generalization Ability of the LMBP network model of aforementioned foundation.
Wherein, i=8, j=70.
After sample set A imports MATLAB Neural Network Toolbox NNTOOL using batch change mode, Training set is taken (training set): Cross Validation set (verifying collection): Test set (test set)=70:15:15 ratio, into Row training.Then selection for node in hidden layer, the present embodiment rule of thumb formula (8) Lai Jinhang primary election use network Whether the selection that tool box carrys out the practice examining number of nodes is reasonable.It is when node in hidden layer is 10, i.e., refreshing by repeatedly training Structure through network is 8-10-1, and hidden layer transmission function is " TANSIG ", and output layer transmission function is " TANSIG ", other When parameter uses default setting, optimal modeling effect (as shown in Figure 3) can be obtained.From figure 3, it can be seen that The coefficient R of Training, Validation, Test are both greater than 0.999, and coefficient R illustrates sample set closer to 1 NEW in AijAfter the coupling of multiple features index and corresponding denitration efficiency YjCorrelation it is higher, the fitting performance of network is also better.
In formula (8), m is the number of nodes of hidden layer, and n is the number of nodes of input layer, and n=8 herein, α are between 1~10 Constant.
In order to examine the PCA-LMBP model of foundation to the fitting effect of itself modeling data, again the several of sample set A What, physical and chemical and denitration efficiency numerical value input the model and are simulated, and the result obtained is as shown in Figure 4.From fig. 4, it can be seen that making Itself is predicted with established model, the maximum deviation of denitration efficiency measured value and predicted value shows less than 0.7 Model is preferable to the applicability of itself, does not occur obvious poor fitting (under-fitting) or over-fitting (over- Fitting) phenomenon.
For the popularization generalization ability of established model before examining, in other words to the prediction energy of same type sample set B Power, totally 10 samples input network to the sample set B that hereinbefore random division is come out, and show that the results are shown in Table 4, in table 4 The activity index of catalyst is calculated simultaneously.
4 model prediction result of table and activity export result
In table 4, Y1 is actual measurement denitration efficiency, and Y2 is prediction denitration efficiency, and M is the absolute value of (Y1-Y2)/Y1*100%, A1 is actual measurement activity, and A2 is prediction activity, and N is the absolute value of (A1-A2)/A1*100%, and the unit of Y1, Y2 are %, A1, A2 Unit be m/h.
It can be seen that the maximum relative deviation absolute value of the prediction to sample set B, denitration efficiency measured value and predicted value is small In 0.4%, actual measurement export activity, less than 0.8%, shows the mould established with the active maximum relative deviation absolute value of prediction export Type is also preferable to the prediction applicability of 10 samples of sample set B.

Claims (10)

1. a kind of fresh honeycomb type denitrification catalyst operational characteristic index prediction technique based on measured data, which is characterized in that The following steps are included:
(1) sample collection: collecting catalyst to be predicted, for the fresh honeycomb type denitrification catalyst of same producer production；
(2) data collection: geometrical property measurement, physicochemical property measurement, operational characteristic measurement are carried out to the sample of collection, obtained several What characterisitic parameter, physicochemical property parameter, operational characteristic parameter；
(3) data processing: step (2) are collected into the sample set matrix M that data are constituted and are normalized, and Metzler matrix is returned The submatrix that geometrical property parameter, physicochemical property parameter after one change are constituted carries out correlation analysis and principal component analysis；
(4) LMBP neural network model is established；
(5) model prediction:
For catalyst sample to be predicted, step (1)~(3) are repeated, are obtained using the neural network model that step (4) is established Predicted value renormalization out obtains the operational characteristic parameter of catalyst.
2. prediction technique according to claim 1, which is characterized in that the specific method of step (3) data processing is such as Under:
A, step (2) collects data and constitutes sample set matrix M, and matrix M is normalized to matrix MZ, and MZ is divided into two submatrix X_ MZ, Y_MZ；Every a line of the matrix M is known as a sample, and each sample standard deviation is by catalyst geometrical property parameter, physicochemical property Parameter, operational characteristic parameter composition；The submatrix X_MZ normalized by Metzler matrix after geometrical property parameter, physicochemical property ginseng Number is constituted, and the operational characteristic parameter after submatrix Y_MZ is normalized by Metzler matrix is constituted；
B, Pearson correlation coefficient analysis is carried out to X_MZ, and taken absolute value to each column of the correlation matrix obtained, and Judge the maximum value of every column element absolute value, if maximum value is respectively less than 0.3, data processing terminates, and otherwise carries out main in step c Constituent analysis；
C, principal component analysis is carried out, number of principal components is chosen according to contribution rate of accumulative total, extracts principal component scores coefficient matrix, reduction master Component matrix X_RZ.
3. prediction technique according to claim 2, which is characterized in that in the step b most to every column element absolute value When big value is judged, it is set to 0 again when the value is 1.
4. prediction technique according to claim 3, which is characterized in that the modeling of LMBP neural network in the step (4) Process is as follows:
D, to gained submatrix X_RZ final after step (3) data processing or the submatrix X_MZ of principal component analysis is needed not move through, And corresponding submatrix Y_MZ, it adopts in a like fashion, is divided into training set, verifying collection, test set；
E, three layers of LMBP neural network are established, input node of the element as network in matrix X_RZ or X_MZ, matrix Y_MZ Output node of the element as network；
F, continue repeatedly to train network, and whether the coefficient R of training of judgement collection, verifying collection, test set is all larger than 0.995, If meeting the condition, network deconditioning, network modelling is completed, and otherwise continues to train network until meeting the condition.
5. prediction technique according to claim 4, which is characterized in that geometrical property parameter, physical and chemical spy in the step (2) Property parameter includes catalyst aperture, interior wall thickness, outer wall thickness, section side length, main chemical compositions percentage contents, microchemistry Ingredient percentage contents, specific pore volume and most probable pore size, specific surface area.
6. prediction technique according to claim 5, which is characterized in that operational characteristic parameter includes urging in the step (2) Agent denitration efficiency.
7. prediction technique according to claim 6, which is characterized in that obtained in the step (5) by renormalization Catalyst process characterisitic parameter is denitration efficiency, then calculates catalyst activity according to denitration efficiency.
8. prediction technique according to claim 7, which is characterized in that the selection of number of principal components in the step c, according to right Submatrix X_MZ carries out the principal component contribution rate of accumulative total after principal component analysis and is greater than 0.85 progress.
9. prediction technique according to claim 8, which is characterized in that divide training set in the step d, verifying collection, survey The ratio of examination collection meets condition (60~70): (15~20): (15~20).
10. prediction technique according to claim 9, which is characterized in that the sample number collected in the step (1) is greater than single Geometrical property parameter, the number of the physicochemical property parameter adduction of a catalyst sample.
CN201910654885.XA 2019-07-19 2019-07-19 Denitration catalyst process characteristic index prediction method based on measured data Active CN110298142B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910654885.XA CN110298142B (en) 2019-07-19 2019-07-19 Denitration catalyst process characteristic index prediction method based on measured data

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910654885.XA CN110298142B (en) 2019-07-19 2019-07-19 Denitration catalyst process characteristic index prediction method based on measured data

Publications (2)

Publication Number Publication Date
CN110298142A true CN110298142A (en) 2019-10-01
CN110298142B CN110298142B (en) 2023-01-03

Family

ID=68031512

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910654885.XA Active CN110298142B (en) 2019-07-19 2019-07-19 Denitration catalyst process characteristic index prediction method based on measured data

Country Status (1)

Country Link
CN (1) CN110298142B (en)

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107193273A (en) * 2017-06-21 2017-09-22 广东电网有限责任公司电力科学研究院 A kind of thermal power plant's SCR denitration system optimum management and fault diagnosis platform
CN108038561A (en) * 2017-09-21 2018-05-15 南京航空航天大学 A kind of Multipurpose Optimal Method of SCR denitration preformed catalyst
CN109499364A (en) * 2018-11-29 2019-03-22 东南大学 A kind of catalyst auxiliary design method based on digital mirror image
CN109521139A (en) * 2018-12-17 2019-03-26 内蒙古电力(集团)有限责任公司内蒙古电力科学研究院分公司 A kind of catalyst activity online test method

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107193273A (en) * 2017-06-21 2017-09-22 广东电网有限责任公司电力科学研究院 A kind of thermal power plant's SCR denitration system optimum management and fault diagnosis platform
CN108038561A (en) * 2017-09-21 2018-05-15 南京航空航天大学 A kind of Multipurpose Optimal Method of SCR denitration preformed catalyst
CN109499364A (en) * 2018-11-29 2019-03-22 东南大学 A kind of catalyst auxiliary design method based on digital mirror image
CN109521139A (en) * 2018-12-17 2019-03-26 内蒙古电力(集团)有限责任公司内蒙古电力科学研究院分公司 A kind of catalyst activity online test method

Also Published As

Publication number Publication date
CN110298142B (en) 2023-01-03

Similar Documents

Publication Publication Date Title
CN112069567A (en) Method for predicting compressive strength of concrete based on random forest and intelligent algorithm
US20220082545A1 (en) Total Nitrogen Intelligent Detection Method Based on Multi-objective Optimized Fuzzy Neural Network
CN100514060C (en) On-line virtual monitoring method for water chemical oxygen demand
CN111967721A (en) Comprehensive energy system greening level evaluation method and system
CN107045672A (en) A kind of evaluation method of community's low carbon levels
CN105205002B (en) A kind of software safety defect based on test job amount finds the modeling method of model
CN109658988A (en) A kind of Hydrobon catalyst performance prediction method
CN104915522A (en) Mixed modeling method and system based on combination of process priors and data-driven model
CN110827169A (en) Distributed power grid service monitoring method based on grading indexes
CN107247888A (en) Sewage disposal water outlet total phosphorus TP flexible measurement methods based on reserve pool network
CN102799717B (en) Sample learning-based porous media modeling method
CN111027257B (en) Method for predicting safe storage time of pulverized coal covered coal pile by using neural network
CN113297685A (en) Vehicle operation condition mode identification method
CN110298142A (en) A kind of denitrating catalyst operational characteristic index prediction technique based on measured data
CN112860973A (en) Organization health assessment method based on social network analysis
Arya et al. Identification and classification of key variables and their role in environmental impact assessment: Methodology and software package intra
CN104104389B (en) A kind of signal reconstruction method and apparatus
CN103793582A (en) Optimization method for cylindrical shell large opening connecting pipe structure
CN103164531B (en) Based on the two-stage instance layer data integrating method of Fuzzy priority level
Sevryukova et al. A Study on Design Principles of Automatic System for Environment Monitoring
CN114814092A (en) IP index measuring method based on BP neural network
CN109784568B (en) Method for predicting lake water quality model through multi-target uncertainty analysis
CN114254928A (en) Industrial Internet platform development index monitoring method and system
CN111062118B (en) Multilayer soft measurement modeling system and method based on neural network prediction layering
CN112599205A (en) Method for designing soft measurement model of total phosphorus in effluent in event-driven sewage treatment process

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant