CN110263997A - A kind of cement producing line flue gas NO based on deep neural networkxConcentration prediction method - Google Patents
A kind of cement producing line flue gas NO based on deep neural networkxConcentration prediction method Download PDFInfo
- Publication number
- CN110263997A CN110263997A CN201910529701.7A CN201910529701A CN110263997A CN 110263997 A CN110263997 A CN 110263997A CN 201910529701 A CN201910529701 A CN 201910529701A CN 110263997 A CN110263997 A CN 110263997A
- Authority
- CN
- China
- Prior art keywords
- data
- variable
- input
- concentration
- network
- 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.)
- Pending
Links
- 239000004568 cement Substances 0.000 title claims abstract description 39
- 238000000034 method Methods 0.000 title claims abstract description 33
- UGFAIRIUMAVXCW-UHFFFAOYSA-N Carbon monoxide Chemical compound [O+]#[C-] UGFAIRIUMAVXCW-UHFFFAOYSA-N 0.000 title claims abstract description 31
- 239000003546 flue gas Substances 0.000 title claims abstract description 31
- 230000001537 neural effect Effects 0.000 title description 2
- 238000012549 training Methods 0.000 claims abstract description 25
- 238000004519 manufacturing process Methods 0.000 claims abstract description 21
- 238000013528 artificial neural network Methods 0.000 claims abstract description 16
- 230000008569 process Effects 0.000 claims abstract description 16
- 230000000875 corresponding effect Effects 0.000 claims abstract description 12
- 230000002596 correlated effect Effects 0.000 claims abstract description 7
- 238000005096 rolling process Methods 0.000 claims abstract description 6
- 210000002569 neuron Anatomy 0.000 claims description 18
- QGZKDVFQNNGYKY-UHFFFAOYSA-N Ammonia Chemical compound N QGZKDVFQNNGYKY-UHFFFAOYSA-N 0.000 claims description 10
- 238000000605 extraction Methods 0.000 claims description 8
- VHUUQVKOLVNVRT-UHFFFAOYSA-N Ammonium hydroxide Chemical compound [NH4+].[OH-] VHUUQVKOLVNVRT-UHFFFAOYSA-N 0.000 claims description 5
- 229910021529 ammonia Inorganic materials 0.000 claims description 5
- 239000000908 ammonium hydroxide Substances 0.000 claims description 5
- 239000003245 coal Substances 0.000 claims description 5
- 238000012216 screening Methods 0.000 claims description 4
- 238000000354 decomposition reaction Methods 0.000 claims description 3
- 230000003252 repetitive effect Effects 0.000 claims description 2
- 230000005611 electricity Effects 0.000 claims 1
- 238000001514 detection method Methods 0.000 abstract description 4
- 238000012360 testing method Methods 0.000 abstract description 4
- 230000006870 function Effects 0.000 description 8
- 238000005070 sampling Methods 0.000 description 7
- 230000004913 activation Effects 0.000 description 4
- 230000009467 reduction Effects 0.000 description 4
- 238000006722 reduction reaction Methods 0.000 description 4
- 238000006243 chemical reaction Methods 0.000 description 3
- 238000002485 combustion reaction Methods 0.000 description 3
- 238000012937 correction Methods 0.000 description 3
- 239000000284 extract Substances 0.000 description 3
- 239000000446 fuel Substances 0.000 description 3
- MWUXSHHQAYIFBG-UHFFFAOYSA-N nitrogen oxide Inorganic materials O=[N] MWUXSHHQAYIFBG-UHFFFAOYSA-N 0.000 description 3
- 238000012545 processing Methods 0.000 description 3
- 239000007921 spray Substances 0.000 description 3
- IJGRMHOSHXDMSA-UHFFFAOYSA-N Atomic nitrogen Chemical compound N#N IJGRMHOSHXDMSA-UHFFFAOYSA-N 0.000 description 2
- 230000015572 biosynthetic process Effects 0.000 description 2
- 238000004364 calculation method Methods 0.000 description 2
- 238000010531 catalytic reduction reaction Methods 0.000 description 2
- 230000008859 change Effects 0.000 description 2
- 235000019504 cigarettes Nutrition 0.000 description 2
- 238000010438 heat treatment Methods 0.000 description 2
- 229910052757 nitrogen Inorganic materials 0.000 description 2
- 238000004321 preservation Methods 0.000 description 2
- NAWXUBYGYWOOIX-SFHVURJKSA-N (2s)-2-[[4-[2-(2,4-diaminoquinazolin-6-yl)ethyl]benzoyl]amino]-4-methylidenepentanedioic acid Chemical compound C1=CC2=NC(N)=NC(N)=C2C=C1CCC1=CC=C(C(=O)N[C@@H](CC(=C)C(O)=O)C(O)=O)C=C1 NAWXUBYGYWOOIX-SFHVURJKSA-N 0.000 description 1
- 239000004215 Carbon black (E152) Substances 0.000 description 1
- -1 Nitrogenous compound Chemical class 0.000 description 1
- 238000003916 acid precipitation Methods 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 239000003054 catalyst Substances 0.000 description 1
- 230000003197 catalytic effect Effects 0.000 description 1
- 239000002817 coal dust Substances 0.000 description 1
- 230000003750 conditioning effect Effects 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000009826 distribution Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 229930195733 hydrocarbon Natural products 0.000 description 1
- 150000002430 hydrocarbons Chemical class 0.000 description 1
- 239000013067 intermediate product Substances 0.000 description 1
- 150000002500 ions Chemical class 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 239000000203 mixture Substances 0.000 description 1
- 210000005036 nerve Anatomy 0.000 description 1
- 231100000252 nontoxic Toxicity 0.000 description 1
- 230000003000 nontoxic effect Effects 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 238000010248 power generation Methods 0.000 description 1
- 239000012716 precipitator Substances 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 238000005507 spraying Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 230000004083 survival effect Effects 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- 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
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/04—Manufacturing
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
- Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
- Y02P90/30—Computing systems specially adapted for manufacturing
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Business, Economics & Management (AREA)
- General Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Economics (AREA)
- General Health & Medical Sciences (AREA)
- Human Resources & Organizations (AREA)
- Strategic Management (AREA)
- General Business, Economics & Management (AREA)
- Computing Systems (AREA)
- Data Mining & Analysis (AREA)
- Marketing (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- General Engineering & Computer Science (AREA)
- Tourism & Hospitality (AREA)
- Molecular Biology (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Primary Health Care (AREA)
- Quality & Reliability (AREA)
- Operations Research (AREA)
- Entrepreneurship & Innovation (AREA)
- Game Theory and Decision Science (AREA)
- Development Economics (AREA)
- Manufacturing & Machinery (AREA)
- Treating Waste Gases (AREA)
- Feedback Control In General (AREA)
Abstract
The invention discloses the cement producing line flue gas NO based on deep neural networkxConcentration prediction method belongs to manufacture of cement flue gas NOxConcentration Testing field.First according to NOxMechanism of production combination cement producing line process flow filters out prediction NOxRequired correlated variables is downloaded the data of variable from cement database and is pre-processed;Then variable data is formed into input data in a manner of sliding window, so that input data implies the delay characteristics of each variable;Unsupervised training is carried out to extract delay characteristics, with the reversed Training DNN network of BP algorithm to extract corresponding relationship feature using the input layer of DNN network and hidden layer as DBN again;Finally, historical data binding model predicted value rolling forecast goes out the NO of following a period of timex.The method of the present invention is preferably solved because of flue gas NOxTest point is set to chimney, leads to NOxConcentration delay detection, to be difficult to set up NOxThe problem of prediction model.
Description
Technical field
The invention belongs to manufacture of cement flue gas NOxConcentration Testing field, more particularly to a kind of based on deep neural network
Cement producing line flue gas NOxConcentration prediction method.
Background technique
Nitrogen oxides (the NO of China's cement industryx) discharge amount accounts for national NOx10% or so of discharge amount is thermal power generation
With the third-largest NO after vehicle exhaustxEmission source.A large amount of NOxDischarge can change the property of atmosphere, so as to cause acid rain, photochemical
The atmosphere polluting problems such as pollution are learned, have significant damage to natural environment and environment for human survival." cement industry " 13 " development
Planning " in explicitly point out, the year two thousand twenty NOxDischarge amount reduced 30% than 2015.Therefore, cement industry is as NOxThe weight of discharge
One of industry is wanted, control and reduces NOxDischarge is of great significance.
Cement denitrating flue gas is mainly using selective non-catalytic reduction (Selective No Catalytic at present
Reduction is abbreviated as SNCR, hereafter indicates selective non-catalytic reduction with " SNCR ") method.SNCR method is in suitable temperature
Area sprays into ammonium hydroxide, by the NO in flue gas under conditions of not using catalystxQuickly it is reduced into nontoxic water and nitrogen, Jin Ershi
Existing denitrating flue gas.The NO that cement industry generatesxIt is reduced in denitration reaction generating region, completes NOxThe reduction of concentration, later flue gas
NOxGreat variety does not occur for concentration through process procedures such as mulitistage cyclone, conditioning Tower, dust-precipitators, flows to chimney discharge.
Effectively detection flue gas NOxConcentration sensor setting chimney be discharged at atmospheric air port rather than practical denitration reaction generating region, cause not
The NO of evitable process flowxConcentration detection delay.Cement production system has large dead time, time-varying, strong nonlinearity simultaneously
The characteristics of, so that NOxThe time delay of Concentration Testing has uncertainty.Flue gas NOxThe Accurate Prediction of concentration can be NOxEmission reduction control
It makes (the present invention is based on SNCR denitration controls) and effective data support is provided, so Accurate Prediction postpones the flue gas NO detectedx
Concentration is of great significance.
Summary of the invention
For overcome the deficiencies in the prior art, the invention proposes one kind to be based on deep neural network (Deep Neural
Network, DNN) prediction flue gas NOxThe method of concentration, being capable of the NO of instant prediction for a period of timexConcentration preferably solves cigarette
Gas NOxThe delay issue of detection provides reliable data for the ammonia spraying amount calculating in SNCR denitration control process and supports.
To achieve the above object, the present invention is realized according to following technical scheme:
A kind of cement producing line flue gas NO based on deep neural networkxConcentration prediction method, includes the following steps:
Step S1: according to NOxMechanism of production combination cement producing line process flow, screening prediction NOxThe related of concentration becomes
Amount;
Step S2: the variable data of screening is downloaded from cement production enterprise database and is pre-processed, sliding window is passed through
Mode makes each variable data form ordered series of numbers, and the delay characteristics that each variable data includes are lain in the input ordered series of numbers of model;
Step S3: in such a way that unsupervised training and Training combine, by delay characteristics of input data and defeated
The corresponding relationship feature extraction and preservation of data establish NO into the parameter of DNN network outxPrediction model;
Step S4: in conjunction with historical data and NOxPredicted value rolling forecast goes out the NO of following a period of timexConcentration value.
In above-mentioned technical proposal, correlated variables includes input variable and output variable, and wherein input variable includes feeding capacity
X1, ammonium hydroxide flow X2, kiln current average X3, Coaling of Decomposing Furnace X4, secondary air temperature X5, level-one cylinder outlet temperature X6, kiln tail temperature
Spend X7, level-one cylinder O2 content X8, denitration is for ammonia pump frequency X9, kiln hood hello coal amount X10, decomposition furnace outlet temperature X11, throat NOxIt is dense
Spend X12;Output variable includes throat NOxConcentration Y.
It is using sliding window that selected variable data is defeated by formula (1)~(4) formation in step S2 in above-mentioned technical proposal
Enter output data, each input variable ordered series of numbers indicates are as follows:
Xi(t)={ Xi(t-k-m, t-k) }, i=1,2,3 ... 11 (1)
X12(t)={ X12(t-n,t)} (2)
The input ordered series of numbers of network input layer are as follows:
X (t)={ X1(t),X2(t),X3(t),……,X12(t)} (3)
Network output layer are as follows:
Y (t+1)=X12(t+1) (4)
T is prediction time in formula (1)~(4), and X (t) and Y (t+1) respectively indicate the data of input layer, output layer, and k value is small
In the minimal time delay time, Xi(t-k-m, t-k) indicates XiFor variable from the t-k-m moment to the time series at t-k moment, m is that the time is long
Degree, X12(t-n, t) indicates X12Time series of the variable from t-n moment to t moment, n are time span, X12(t+1) it indicates under using
The NO at one momentxThe label data that concentration value is inputted as t moment.
In above-mentioned technical proposal, step S3 is specifically included: the input data for enabling DNN network training process is X (t)={ X1
(t),X2(t),X3(t),……,X12(t) }, output data is Y (t+1)=X12(t+1);The input layer of DNN and hidden layer are made
Unsupervised training is carried out for DBN, to extract the delay characteristics for each variable for including in input data, and the feature of extraction is protected
There are on the initial weight of DNN input layer and hidden layer;The DBN for saving input data delay characteristics is connect with output layer,
DNN network parameter is reversely finely tuned by BP algorithm, so that the network for making the corresponding relationship feature of output data be stored in model is joined
In number;Finally establish NOxPrediction model, the number of plies of DNN network and every layer of neuron according to the concrete condition of variable data into
Row setting.
In above-mentioned technical proposal, step S4 is specifically included:
Using formula (5)~(7) historical data and NOxPredicted value combines, the NO of rolling forecast a period of time in futurexConcentration
Value, each input variable ordered series of numbers are become from formula (1) (2):
Network input layer ordered series of numbers is become from formula (3):
In formula (6),For the NO at t+1 momentxPredicted value, by ordered series of numbers X1(t) NO is inputtedxPrediction model obtains t+2
The NO at momentxPredicted valueContinuous repetitive (5)~(7) process, obtains k+1 NOxPredicted value, to realize t to t+k+
The NO at 1 momentxConcentration prediction.
Compared with prior art, the present invention having the following beneficial effects:
1, DNN cement producing line flue gas NO proposed by the present inventionxConcentration prediction model, when can be extracted from input data
Time-varying delay feature extracts corresponding relationship feature from output data, and the time delay of extraction and relationship characteristic is stored in network mould
In the parameter of type.To predict following flue gas NO by the variable data of historyxValue.
2, the present invention passes through the transformation of prediction model input data, i.e. mode of the historical data in conjunction with prediction data, in advance
Measure the NO of following a period of timexValue reaches and eliminates time-vary delay system Accurate Prediction flue gas NOxThe purpose of concentration.
3, the present invention establishes the historical data that the data of model are cement production enterprise databases, without newly adding standby acquisition number
According to.Therefore, cost is relatively low and has well adapting to property to different cement producing lines for model.
Detailed description of the invention
In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, to embodiment or will show below
There is attached drawing needed in technical description to be briefly described, it should be apparent that, the accompanying drawings in the following description is only this
Some embodiments of invention for those of ordinary skill in the art without creative efforts, can be with
Other attached drawings are obtained according to these attached drawings.
Fig. 1 is DNN cement flue gas NO proposed by the present inventionxConcentration prediction model structure chart;
Fig. 2 is DBN network structure;
Fig. 3 is limited Boltzmann machine structure chart;
Fig. 4 is that input ordered series of numbers shifts gears schematic diagram;
Fig. 5 is cement flue gas NO proposed by the present inventionxForecasting system flow chart.
Specific embodiment
In order to make the object, technical scheme and advantages of the embodiment of the invention clearer, below in conjunction with the embodiment of the present invention
In attached drawing, technical scheme in the embodiment of the invention is clearly and completely described, it is clear that described embodiment is
A part of the embodiment of the present invention, instead of all the embodiments.
The invention proposes one kind to predict cement producing line flue gas NO based on deep neural networkxThe method of concentration.First
Variable is chosen, corresponding data are downloaded from cement production enterprise database, for training DNN network, another part is used for a part
Examine the forecasting accuracy of network;Then the cement flue gas of the implicit time-vary delay system information of input based on deep neural network is established
NOxConcentration prediction model, structure are shown in Fig. 1;The DNN cement flue gas NO proposed by the present invention for inputting implicit time-vary delay system informationxPrediction
System flow chart is shown in Fig. 5;Finally, carrying out the reversed fine tuning of DNN network parameter by BP algorithm, i.e., by error to neuron
Weight and biasing tuning, realize the foundation of DNN prediction model, the specific steps are as follows:
Step S1: according to NOxFormation mechanism combination cement production process chooses correlated variables and downloading data, to each variable
Data carry out outlier processing and normalized;In step S1, sample data set is chosen from the database of cement production enterprise, and
Outlier processing and normalized are carried out to the data of selection.Deep neural network can be by the feature and input of correlated variables
Relationship between output is included in network parameter, therefore does not need to carry out variable data complicated pretreatment, as long as by shadow
Ring the outlier processing of training.What input data was chosen is different variable, and the unit and variation range of each variable are respectively not
It is identical, therefore input data is normalized before training.Variable data is set to form input using sliding window mode
Data, it is therefore an objective to which the time response of variable is implicit in input data.It can extract according to deep neural network special between variable
Sign relationship simultaneously makes it lie in the feature in network parameter, using the actual production data in cement production enterprise database, establishes
Play the corresponding NO of a period of time variable dataxThe network structure of concentration value, to excavate premeasuring and each input variable
Relationship reaches and eliminates time-vary delay system influence and then Accurate Prediction NOxThe purpose of concentration.
Specifically, NO in manufacture of cementxThere are mainly three types of types: burning type, heating power type, momentary type.Fuel type NOxIt is combustion
Nitrogenous compound in material is first heated is cracked into the intermediate products such as N, CN, HCN in combustion, after be oxidized and generate NOx.?
Using coal dust as in the cement kiln of fuel, fuel type NOxProduction quantity account for about the total NO of cement kilnx60% or more of production quantity.Heating power type
NOxIt is that N2 and O2 in air reacts generation under the high temperature conditions, reaction temperature production quantity after being higher than 1800k can become rapidly
Greatly.Momentary type NOxIt is that N molecule and the ion cluster that hydrocarbon combustion generates in air collides and be oxidized after reacting
It is formed, momentary type NOxThe general very little of production quantity.
According to NOxMechanism of production it is found that NOxMainly generated in two production links of rotary kiln and dore furnace.Ammonium hydroxide is through dividing
The spray ammonia equipment for solving furnace top sprays into, and restores NO under the conditions of suitable temperature on dore furnace topxComplete denitrification process.In conjunction with upper
State analysis, following 12 input variables of DNN model selection: feeding capacity X1, ammonium hydroxide flow X2, kiln current average X3, dore furnace feed
Coal amount X4, secondary air temperature X5, level-one cylinder outlet temperature X6, kiln end temperature X7, level-one cylinder O2 content X8, denitration is for ammonia pump frequency X9,
Kiln hood feeds coal amount X10, decomposition furnace outlet temperature X11, throat NOxConcentration X12;The output variable of selection is throat NOxConcentration Y.
DNN network model can propose the time-vary delay system feature implied in input data and each data variable and the relationship characteristic of premeasuring
It takes and saves to the structure weight of network, therefore each variable data need to only be pre-processed, rear formation input and output come
Training DNN model.
Step S2: pretreated variable data is formed into input data by way of sliding window, by each variable
The delay characteristics that data include are lain in the input ordered series of numbers of model, while using the throat NO of subsequent timexConcentration normalizing
Change numerical value as output data;
The X of m (s) duration is used first1~X11The X of variable data and n (s) duration12Variable data is in a manner of sliding window
Input layer is constructed, input layer structure is as shown in Figure 1.Wherein, each row of data indicates different input variables, and such as feeding capacity feeds coal
Amount etc.;Each column data indicate different moments each variable value.
Each input variable ordered series of numbers indicates are as follows:
Xi(t)={ Xi(t-k-m, t-k) }, i=1,2,3 ... 11 (1)
X12(t)={ X12(t-n,t)} (2)
Network input layer ordered series of numbers are as follows:
X (t)={ X1(t),X2(t),X3(t),……,X12(t)} (3)
Network output layer are as follows:
Y (t+1)=X12(t+1) (4)
T is prediction time in above formula, and X (t) and Y (t+1) respectively indicate the data of input layer, output layer, and k value is slightly less than most
Small decay time, Xi(t-k-m, t-k) indicates XiFor variable from the t-k-m moment to the time series at t-k moment, m is time span,
X12(t-n, t) indicates X12Time series of the variable from t-n moment to t moment, n are time span, X12(t+1) it indicates with next
The NO at momentxThe label data that concentration value is inputted as t moment.
Step S3: in such a way that unsupervised training and Training combine, by delay characteristics of input data and defeated
The corresponding relationship feature extraction and preservation of data establish NO into the parameter of DNN network outxPrediction model;
In step s3, input layer, the output layer neuron number of DNN network are determined according to input data, output data;
According to NOxThe industrial data feature of concentration correlated variables determines the number of plies and every layer of neuron number of DBN network.The maximum instruction of setting
Practice number, learning rate, iteration sample size size.Using the input layer of DNN network and hidden layer as DBN network, greedy nothing is carried out
The forward direction training of supervision, the initial value for determining weight w, biasing a and b, so that can be realized data characteristics by DBN extracts sum number
The characteristics of according to dimensionality reduction, the delay characteristics that input data implies were extracted and preserved.Having the anti-of supervision is carried out to entire DNN network again
To training, the adjustment of whole network parameter is realized, the corresponding relationship feature with output data is stored in network parameter.Most
Eventually, a period of time variable data prediction single-point NO is establishedxThe DNN model of value.
DNN network architecture proposed by the present invention is set as 4 layers, and three first layers are by two limited Boltzmann machines (RBM)
It stacks, an independent neuron is as the 4th layer.That is, three first layers as DBN, then connect with the output layer of single neuron
It connects.Every layer of neuron number of DNN model is followed successively by 672,100,50,1, and learning rate is set as 0.1, and by the company of DBN network structure
It connects weight and biasing is set as 0.
The unsupervised training process of DBN network is that each RBM is successively trained since bottom.First RBM is with entirely
The input layer training of DNN network;RBM above uses the output layer of previous RBM to complete to train as input layer, and all RBM are complete
At obtaining entire DBN network parameter, DBN network structure such as Fig. 2 after training.It is illustrated, is tied with the training process of single RBM
Structure is shown in Fig. 3.
RBM model is a kind of energy model, and utilisable energy function and probability are described and calculate.If aobvious layer nerve
First v has n, and hidden neuron h has m, then the energy function of a RBM given two state of value may be defined as
In formula, wijTo show layer neuron viWith hidden neuron hjThe weight of connection, aiFor neuron viBiasing, bjFor
Neuron hjBiasing.θ=(w, a, b) indicates to constitute the parameter of energy function.The probability distribution of energy function is
In formula, Z is normalization factor.
According to above formula, the activation probability of hidden layer is calculated by showing layer:
Aobvious layer is calculated by hidden layer and activates probability:
In formula,For sigmiod function.The contrast divergence algorithm proposed according to professor Hinton
(Contrastive Divergence, CD) trains RBM, and the thought of the algorithm is to calculate hidden layer to the sampling of aobvious layerAobvious layer is updated to hidden layer sampling againθ=(w, a, b) each parameter calculation formula
It is as follows:
bj=P (hj=1 | v(0))-P(hj=1 | v(k)) (12)
In contrast divergence algorithm, k value, which generally takes, 1 can obtain preferably result.
So that input layer ordered series of numbers X is entered aobvious layer, the activation probability P (h of hidden layer is obtained by formula (8)I=|1V), then with formula (9)
Reverse goes out the activation probability of aobvious layer, obtains weight and biasing finally by formula (10)~(12), to complete primary unsupervised
Training.
It repeats the above process, training is completed to the RBM of top, obtains weight and the biasing of whole DBN.
On the basis of obtaining DBN network parameter, optimized entirely according to exemplar using the reversed error correction algorithms of BP
DNN network parameter w, a, b have the reversed fine tuning in supervision ground to network, complete the foundation of DNN model.
Composition DNN network is connect before completing entirely with output layer to the DBN of unsupervised training, using BP algorithm to DNN network
Parameter carries out reversed layer-by-layer correction, realizes the reversed fine tuning of DNN network model.The mesh of network parameter is reversely adjusted using BP algorithm
Be, the corresponding relationship feature extraction of exemplar and input data and save into the parameter of network.Due to the step
Exemplar is needed to carry out error calculation, therefore the step has supervision.
Using trained DBN network parameter as three first layers structure initial parameter, the output layer of single neuron is then connected,
Connection weight takes random value, and single neuron biasing takes 0.The reversed micro- of DNN network parameter is realized with the reversed error correction algorithms of BP
It adjusts.
Error function:
In formula, p is neural network forecast value, and y is exemplar value.
It enablesIndicate l layers of network of i neuron output, f () indicates that activation primitive, η indicate learning rate.Accidentally
Difference function is to j-th of neuron weight in output layer (L layers) k-th of neuron and L-1 layersLocal derviation:
To the local derviation of biasing:
To L-1 hidden layer, local derviation of the error to weight:
Local derviation of the error to biasing:
It enables
So
The then variable quantity of l layers of weight and biasing are as follows:
Finally, the more new formula of weight and biasing:
wl=wl-η*△wl (24)
bl=bl-η*△bl (25)
The reversed trim process of DNN network model proposed by the present invention is introduced according to above-mentioned formula.Pass through formula (22) first
(23) weight of output layer (L layers) and preceding layer (L-1 layers) and the variable quantity of biasing are calculated, then is completed by formula (24) (25)
The update of weight and biasing.Reversed successively realize to all weights of DNN network and biasing that repeat the above steps updates tuning, from
And the corresponding relationship feature extraction that exemplar includes is stored in whole network and is completed in parameter.After the completion of reverse train, cigarette
Gas NOxConcentration value prediction model just establishes.
Step S4: with historical data combination NOxPredicted value rolling forecast goes out the flue gas NO of following a period of timexConcentration value.
Make to enter data into prediction model and obtains next sampling instant NOxPredicted value, after by input data include it is each
Variable series moves back a sampling instant, in conjunction with NOxPredicted value form new input data, next sampling instant NO under completionxValue
Prediction.It repeats the above steps and predicts the NO of following a period of timexValue.
Specifically, the input variable ordered series of numbers in step S1 is made to enter step trained NO in S3xIn concentration prediction model,
Obtain the NO of subsequent timexConcentration prediction valueBy variable X in step S1i(t) data move back a sampling instant and are formed
New input ordered series of numbersData move back a sampling instant also to by NOxConcentration prediction valueInclude, it is defeated
Enter ordered series of numbers and shift gears to see Fig. 4, input ordered series of numbers expression formula is as follows:
Each input variable ordered series of numbers is become from formula (1) (2)
Network input layer ordered series of numbers is become from formula (3)
By X1(t) ordered series of numbers input prediction model obtains the NO at t+2 momentxConcentration prediction valueIt repeats the above process
To the NO at t+3 momentxConcentration prediction valueAnd so on can obtain total k+1 NOxConcentration prediction value, realization pass through depth
Neural network prediction t moment is to t+k+1 moment flue gas NOxThe function of concentration value.
The foregoing is only a preferred embodiment of the present invention, but scope of protection of the present invention is not limited thereto,
In the technical scope disclosed by the present invention, any changes or substitutions that can be easily thought of by anyone skilled in the art,
It should be covered by the protection scope of the present invention.Therefore, protection scope of the present invention should be with scope of protection of the claims
Subject to.
Claims (5)
1. a kind of cement producing line flue gas NO based on deep neural networkxConcentration prediction method, which is characterized in that including as follows
Step:
Step S1: according to NOxMechanism of production combination cement producing line process flow, screening prediction NOxThe correlated variables of concentration;
Step S2: the variable data of screening is downloaded from cement production enterprise database and is pre-processed, by way of sliding window
So that each variable data is formed ordered series of numbers, the delay characteristics that each variable data includes are lain in the input ordered series of numbers of model;
Step S3: in such a way that unsupervised training and Training combine, by the delay characteristics of input data and output number
According to corresponding relationship feature extraction and save into the parameter of DNN network, establish NOxPrediction model;
Step S4: in conjunction with historical data and NOxPredicted value rolling forecast goes out the NO of following a period of timexConcentration value.
2. the cement producing line flue gas NO according to claim 1 based on deep neural networkxConcentration prediction method, it is special
Sign is that correlated variables includes input variable and output variable, and wherein input variable includes feeding capacity X1, ammonium hydroxide flow X2, kiln electricity
Levelling mean value X3, Coaling of Decomposing Furnace X4, secondary air temperature X5, level-one cylinder outlet temperature X6, kiln end temperature X7, level-one cylinder O2 content
X8, denitration is for ammonia pump frequency X9, kiln hood hello coal amount X10, decomposition furnace outlet temperature X11, throat NOxConcentration X12;Output variable packet
Include throat NOxConcentration Y.
3. the cement producing line flue gas NO according to claim 2 based on deep neural networkxConcentration prediction method, it is special
Sign is, in step S2, selected variable data is formed inputoutput data by formula (1)~(4) using sliding window, each to input
Variable series indicates are as follows:
Xi(t)={ Xi(t-k-m, t-k) }, i=1,2,3 ... 11 (1)
X12(t)={ X12(t-n,t)} (2)
The input ordered series of numbers of network input layer are as follows:
X (t)={ X1(t),X2(t),X3(t),……,X12(t)} (3)
Network output layer are as follows:
Y (t+1)=X12(t+1) (4)
T is prediction time in formula (1)~(4), and X (t) and Y (t+1) respectively indicate the data of input layer, output layer, and k value is less than most
Small decay time, Xi(t-k-m, t-k) indicates XiFor variable from the t-k-m moment to the time series at t-k moment, m is time span,
X12(t-n, t) indicates X12Time series of the variable from t-n moment to t moment, n are time span, X12(t+1) it indicates with next
The NO at momentxThe label data that concentration value is inputted as t moment.
4. the cement producing line flue gas NO according to claim 3 based on deep neural networkxConcentration prediction method, it is special
Sign is that step S3 is specifically included: the input data for enabling DNN network training process is X (t)={ X1(t),X2(t),X3
(t),……,X12(t) }, output data is Y (t+1)=X12(t+1);Nothing is carried out using the input layer of DNN and hidden layer as DBN
The feature of extraction to extract the delay characteristics for each variable for including in input data, and is stored in DNN input by supervised training
On the initial weight of layer and hidden layer;The DBN for saving input data delay characteristics is connect with output layer, it is anti-by BP algorithm
To fine tuning DNN network parameter, so that the corresponding relationship feature of output data be made to be stored in the network parameter of model;It is final to establish
NOxPrediction model, the number of plies of DNN network are configured with every layer of neuron according to the concrete condition of variable data.
5. the cement producing line flue gas NO according to claim 4 based on deep neural networkxConcentration prediction method, it is special
Sign is that step S4 is specifically included:
Using formula (5)~(7) historical data and NOxPredicted value combines, the NO of rolling forecast a period of time in futurexConcentration value, respectively
Input variable ordered series of numbers is become from formula (1) (2):
Network input layer ordered series of numbers is become from formula (3):
In formula (6),For the NO at t+1 momentxPredicted value, by ordered series of numbers X1(t) NO is inputtedxPrediction model obtains the t+2 moment
NOxPredicted valueContinuous repetitive (5)~(7) process, obtains k+1 NOxPredicted value, to realize t to the t+k+1 moment
NOxConcentration prediction.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910529701.7A CN110263997A (en) | 2019-06-19 | 2019-06-19 | A kind of cement producing line flue gas NO based on deep neural networkxConcentration prediction method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910529701.7A CN110263997A (en) | 2019-06-19 | 2019-06-19 | A kind of cement producing line flue gas NO based on deep neural networkxConcentration prediction method |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110263997A true CN110263997A (en) | 2019-09-20 |
Family
ID=67919239
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910529701.7A Pending CN110263997A (en) | 2019-06-19 | 2019-06-19 | A kind of cement producing line flue gas NO based on deep neural networkxConcentration prediction method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110263997A (en) |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111174824A (en) * | 2019-12-27 | 2020-05-19 | 北京首钢自动化信息技术有限公司 | Control platform that acid mist discharged |
CN111266004A (en) * | 2020-03-20 | 2020-06-12 | 大唐环境产业集团股份有限公司 | SCR denitration reactor inlet NOx delay estimation system |
CN111898867A (en) * | 2020-07-05 | 2020-11-06 | 西北工业大学 | Airplane final assembly production line productivity prediction method based on deep neural network |
CN112163335A (en) * | 2020-09-25 | 2021-01-01 | 广东电科院能源技术有限责任公司 | Training method, prediction method and device of NOx concentration prediction model |
CN113268871A (en) * | 2021-05-21 | 2021-08-17 | 燕山大学 | Cement chimney NOX prediction method based on multivariable time sequence depth network model |
CN113470763A (en) * | 2021-07-14 | 2021-10-01 | 中国航发贵阳发动机设计研究所 | System for measuring and calculating components of thermal dissociation gas generated by combustion of hydrocarbon fuel |
CN115245735A (en) * | 2022-06-16 | 2022-10-28 | 浙江省生态环境科学设计研究院 | Cement kiln flue gas iSNCR control method based on predictive control model |
JP7471105B2 (en) | 2020-02-27 | 2024-04-19 | Ube三菱セメント株式会社 | CEMENT CLINKER MANUFACTURING SUPPORT SYSTEM, CEMENT CLINKER MANUFACTURING SUPPORT METHOD, AND MANUFACTURING SUPPORT PROGRAM |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104793651A (en) * | 2015-04-20 | 2015-07-22 | 四川大学 | Ammonia spraying amount controlling method for SNCR (selective no catalytic reduction) flue gas denitrating system for cement manufacturing |
CN107679671A (en) * | 2017-10-23 | 2018-02-09 | 燕山大学 | A kind of manufacture of cement power consumption Forecasting Methodology of implicit time series depth belief network |
US20180149094A1 (en) * | 2016-11-30 | 2018-05-31 | General Electric Company | Emissions modeling for gas turbine engines |
CN109165798A (en) * | 2018-10-19 | 2019-01-08 | 燕山大学 | A kind of Free Calcium Oxide Contents in Cement Clinker on-line prediction method and system |
-
2019
- 2019-06-19 CN CN201910529701.7A patent/CN110263997A/en active Pending
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104793651A (en) * | 2015-04-20 | 2015-07-22 | 四川大学 | Ammonia spraying amount controlling method for SNCR (selective no catalytic reduction) flue gas denitrating system for cement manufacturing |
US20180149094A1 (en) * | 2016-11-30 | 2018-05-31 | General Electric Company | Emissions modeling for gas turbine engines |
CN107679671A (en) * | 2017-10-23 | 2018-02-09 | 燕山大学 | A kind of manufacture of cement power consumption Forecasting Methodology of implicit time series depth belief network |
CN109165798A (en) * | 2018-10-19 | 2019-01-08 | 燕山大学 | A kind of Free Calcium Oxide Contents in Cement Clinker on-line prediction method and system |
Non-Patent Citations (3)
Title |
---|
刘博文: "燃煤电厂脱硝系统入口NOx浓度测量修正及喷氨量优化控制研究", 《中国优秀硕士学位论文全文数据库 工程科技Ⅰ辑》 * |
李剑波 等: "《大数据挖掘技术与应用》", 31 July 2018, 延边大学出版社 * |
谭思超 等: "《海洋条件下反应堆热工水力特性研究》", 31 August 2018, 国防工业出版社 * |
Cited By (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111174824A (en) * | 2019-12-27 | 2020-05-19 | 北京首钢自动化信息技术有限公司 | Control platform that acid mist discharged |
CN111174824B (en) * | 2019-12-27 | 2022-04-12 | 北京首钢自动化信息技术有限公司 | Control platform that acid mist discharged |
JP7471105B2 (en) | 2020-02-27 | 2024-04-19 | Ube三菱セメント株式会社 | CEMENT CLINKER MANUFACTURING SUPPORT SYSTEM, CEMENT CLINKER MANUFACTURING SUPPORT METHOD, AND MANUFACTURING SUPPORT PROGRAM |
CN111266004A (en) * | 2020-03-20 | 2020-06-12 | 大唐环境产业集团股份有限公司 | SCR denitration reactor inlet NOx delay estimation system |
CN111898867A (en) * | 2020-07-05 | 2020-11-06 | 西北工业大学 | Airplane final assembly production line productivity prediction method based on deep neural network |
CN111898867B (en) * | 2020-07-05 | 2023-04-07 | 西北工业大学 | Airplane final assembly production line productivity prediction method based on deep neural network |
CN112163335A (en) * | 2020-09-25 | 2021-01-01 | 广东电科院能源技术有限责任公司 | Training method, prediction method and device of NOx concentration prediction model |
CN113268871A (en) * | 2021-05-21 | 2021-08-17 | 燕山大学 | Cement chimney NOX prediction method based on multivariable time sequence depth network model |
CN113470763A (en) * | 2021-07-14 | 2021-10-01 | 中国航发贵阳发动机设计研究所 | System for measuring and calculating components of thermal dissociation gas generated by combustion of hydrocarbon fuel |
CN113470763B (en) * | 2021-07-14 | 2023-03-14 | 中国航发贵阳发动机设计研究所 | System for measuring and calculating components of thermal dissociation gas generated by combustion of hydrocarbon fuel |
CN115245735A (en) * | 2022-06-16 | 2022-10-28 | 浙江省生态环境科学设计研究院 | Cement kiln flue gas iSNCR control method based on predictive control model |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110263997A (en) | A kind of cement producing line flue gas NO based on deep neural networkxConcentration prediction method | |
Shin et al. | Development of NOx reduction system utilizing artificial neural network (ANN) and genetic algorithm (GA) | |
CN107243257B (en) | It is suitble to the intelligence spray ammonia control system of full load | |
CN104765350B (en) | Cement decomposing furnace control method and system based on Combined model forecast control technology | |
CN105629738A (en) | SCR (Selective Catalytic Reduction) flue gas denitration system control method and apparatus | |
CN104636823B (en) | A kind of wind power forecasting method | |
CN104793651A (en) | Ammonia spraying amount controlling method for SNCR (selective no catalytic reduction) flue gas denitrating system for cement manufacturing | |
CN113094986B (en) | Method for constructing prediction model of pollutant emission in flue gas of waste incinerator and application | |
CN105404145A (en) | Denitration novel cascade control method based on index prediction and time-lag pre-estimation compensation | |
CN110299188A (en) | SCR flue gas denitrification system GRNN modeling method based on GA variables choice | |
CN110188383A (en) | A kind of power station SCR denitration modeling method based on selective ensemble model | |
CN114191953A (en) | Flue gas desulfurization and denitrification control method based on convolutional neural network and XGboost | |
CN109670625A (en) | NOx emission concentration prediction method based on Unscented kalman filtering least square method supporting vector machine | |
CN109426144B (en) | Power station boiler flue gas denitration method based on random forest model | |
CN107464014A (en) | A kind of Forecasting Methodology of coal unit denitration control system inlet nitrogen oxides | |
CN110263452A (en) | Flue gas Annual distribution characteristic analysis method, system and denitrating system in a kind of flue | |
CN111562744A (en) | Boiler combustion implicit generalized predictive control method based on PSO algorithm | |
CN115049139A (en) | Multi-index model prediction control method for cement sintering denitration system | |
CN114558447A (en) | Ammonia injection amount optimization control device and method based on SCR denitration system | |
CN113112072A (en) | NOx emission content prediction method based on deep bidirectional LSTM | |
CN109046021A (en) | A kind of SCR system of strong adaptive ability accurately sprays ammonia control method | |
CN113268871B (en) | Cement chimney NOX prediction method based on multivariable time sequence depth network model | |
CN114958395B (en) | Waste heat utilization method and system for biomass carbonization furnace | |
CN116071210A (en) | Electric energy carbon emission calculation method and system | |
CN112231978B (en) | Boiler flue gas acid dew point testing method based on artificial neural network |
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 | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190920 |