CN115017818A - Power plant flue gas oxygen content intelligent prediction method based on attention mechanism and multilayer LSTM - Google Patents
Power plant flue gas oxygen content intelligent prediction method based on attention mechanism and multilayer LSTM Download PDFInfo
- Publication number
- CN115017818A CN115017818A CN202210694964.5A CN202210694964A CN115017818A CN 115017818 A CN115017818 A CN 115017818A CN 202210694964 A CN202210694964 A CN 202210694964A CN 115017818 A CN115017818 A CN 115017818A
- Authority
- CN
- China
- Prior art keywords
- data
- network
- attention
- oxygen content
- flue gas
- 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
- QVGXLLKOCUKJST-UHFFFAOYSA-N atomic oxygen Chemical compound [O] QVGXLLKOCUKJST-UHFFFAOYSA-N 0.000 title claims abstract description 51
- 229910052760 oxygen Inorganic materials 0.000 title claims abstract description 51
- 239000001301 oxygen Substances 0.000 title claims abstract description 51
- 238000000034 method Methods 0.000 title claims abstract description 47
- UGFAIRIUMAVXCW-UHFFFAOYSA-N Carbon monoxide Chemical compound [O+]#[C-] UGFAIRIUMAVXCW-UHFFFAOYSA-N 0.000 title claims abstract description 42
- 239000003546 flue gas Substances 0.000 title claims abstract description 42
- 230000007246 mechanism Effects 0.000 title claims abstract description 30
- 238000012549 training Methods 0.000 claims abstract description 26
- 238000005070 sampling Methods 0.000 claims abstract description 21
- 238000004140 cleaning Methods 0.000 claims abstract description 16
- 230000009467 reduction Effects 0.000 claims abstract description 13
- 230000008569 process Effects 0.000 claims abstract description 12
- 239000000779 smoke Substances 0.000 claims abstract description 10
- 238000001914 filtration Methods 0.000 claims abstract description 5
- 239000003245 coal Substances 0.000 claims description 33
- 238000010248 power generation Methods 0.000 claims description 26
- 239000013598 vector Substances 0.000 claims description 24
- 230000006870 function Effects 0.000 claims description 18
- 238000012216 screening Methods 0.000 claims description 17
- 230000004927 fusion Effects 0.000 claims description 13
- 238000000513 principal component analysis Methods 0.000 claims description 13
- 239000011159 matrix material Substances 0.000 claims description 9
- 238000012545 processing Methods 0.000 claims description 8
- 238000002485 combustion reaction Methods 0.000 claims description 7
- 238000001514 detection method Methods 0.000 claims description 6
- 238000004458 analytical method Methods 0.000 claims description 5
- 238000012544 monitoring process Methods 0.000 claims description 5
- 239000000126 substance Substances 0.000 claims description 5
- 238000004364 calculation method Methods 0.000 claims description 4
- 238000003860 storage Methods 0.000 claims description 4
- 230000004913 activation Effects 0.000 claims description 3
- 230000001186 cumulative effect Effects 0.000 claims description 3
- 238000000354 decomposition reaction Methods 0.000 claims description 3
- 239000002994 raw material Substances 0.000 claims description 3
- 238000005259 measurement Methods 0.000 description 6
- 238000007781 pre-processing Methods 0.000 description 5
- OKTJSMMVPCPJKN-UHFFFAOYSA-N Carbon Chemical compound [C] OKTJSMMVPCPJKN-UHFFFAOYSA-N 0.000 description 4
- 229910052799 carbon Inorganic materials 0.000 description 4
- 238000000691 measurement method Methods 0.000 description 4
- 238000013528 artificial neural network Methods 0.000 description 2
- 230000008901 benefit Effects 0.000 description 2
- 238000010276 construction Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000006467 substitution reaction Methods 0.000 description 2
- QRSFFHRCBYCWBS-UHFFFAOYSA-N [O].[O] Chemical compound [O].[O] QRSFFHRCBYCWBS-UHFFFAOYSA-N 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000012512 characterization method Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000009434 installation Methods 0.000 description 1
- 238000012423 maintenance Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 239000010893 paper waste Substances 0.000 description 1
- 230000000704 physical effect Effects 0.000 description 1
- 238000004886 process control Methods 0.000 description 1
- 238000011946 reduction process Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
Images
Classifications
-
- 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
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/16—Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
-
- 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
- G06Q50/06—Energy or water supply
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
- G06F2119/02—Reliability analysis or reliability optimisation; Failure analysis, e.g. worst case scenario performance, failure mode and effects analysis [FMEA]
-
- 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
- Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
- Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
- Y04S10/00—Systems supporting electrical power generation, transmission or distribution
- Y04S10/50—Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Business, Economics & Management (AREA)
- Mathematical Physics (AREA)
- General Engineering & Computer Science (AREA)
- Software Systems (AREA)
- Data Mining & Analysis (AREA)
- Computational Mathematics (AREA)
- Pure & Applied Mathematics (AREA)
- Mathematical Optimization (AREA)
- Evolutionary Computation (AREA)
- Mathematical Analysis (AREA)
- Economics (AREA)
- Health & Medical Sciences (AREA)
- Marketing (AREA)
- General Health & Medical Sciences (AREA)
- General Business, Economics & Management (AREA)
- Tourism & Hospitality (AREA)
- Strategic Management (AREA)
- Primary Health Care (AREA)
- Algebra (AREA)
- Human Resources & Organizations (AREA)
- Databases & Information Systems (AREA)
- Computing Systems (AREA)
- Water Supply & Treatment (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Public Health (AREA)
- Medical Informatics (AREA)
- Computer Hardware Design (AREA)
- Geometry (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention relates to an attention mechanism and multilayer LSTM-based power plant flue gas oxygen content intelligent prediction method, which comprises the steps of collecting data through a sensor, and storing the collected data of the sensor into an original database; sequentially sampling data from an original database, cleaning and removing data with higher similarity through data cleaning and data dimension reduction, filtering secondary features in the data through the data dimension reduction, and storing the processed data in a residual database; randomly extracting small-batch data from the residual database, obtaining weight characteristics through an attention network, and inputting the weight characteristic data into a prediction network based on the LSTM to perform prediction network training; and finally, predicting the oxygen content of the smoke by using the trained prediction network. The invention reduces redundant information between data and in data, thereby improving the model training efficiency. And an attention mechanism is introduced to endow different weights to main characteristics, so that the prediction system is closer to the physical process of an actual system, and the prediction precision is improved.
Description
Technical Field
The invention belongs to the technical field of intelligent measurement and control of coal-fired power plants, and particularly relates to an attention mechanism and multilayer LSTM-based intelligent prediction method for oxygen content of flue gas in a power plant.
Background
With the rapid development of national economy, the layout scale of a domestic power grid is continuously enlarged, and the power generation of a coal-fired power plant is taken as a main power generation way, so that the carbon emission of the power plant is continuously increased. On one hand, the increasing exhaustion of non-renewable energy sources and the strategy of bidding on line lead the power generation cost of the coal-fired power plant to be higher and higher; on the other hand, the strict national requirements for carbon emissions from power plants also make coal-fired power plants compelled to reduce carbon emissions by means of technical upgrade iterations. Especially in recent years, the price of coal is increasing, and the pressure of daily operation of coal-fired power plants is increasing, so that the power plants have to be managed and planned in various aspects to improve the power generation efficiency and reduce the power generation cost and carbon emission.
The oxygen content of the flue gas of the power plant is taken as an important index for measuring the power generation efficiency, and the oxygen content is generally controlled to be 3-5 percent, so that the power generation efficiency can reach the highest. The accurate measurement of the oxygen content of the flue gas is the basis for realizing the control of high-quality fire coal, and most coal-fired power plants measure the oxygen content of the flue gas by means of a hardware sensor at present. Among them, the thermomagnetic oxygen sensor and the oxygen-oxygen sensor are two common ones. However, the traditional hardware sensor has higher investment and installation risks, and the working environment is very severe, so that the daily maintenance workload is quite heavy, and the service life is greatly reduced. In addition, problems such as oxygen meter faults, untimely feedback of certain operation control information and the like can occur in daily life, online monitoring is not facilitated, and the thermal efficiency of the power plant boiler can be affected.
With the wider application of neural network technology in the field of thermal process control, the flue gas oxygen content measurement method based on soft measurement is widely studied [1 ]. And (3) analyzing a large amount of historical data by using a deep neural network, and further performing characterization learning of the data, so that the prediction accuracy and the prediction stability of the oxygen content in the smoke can be improved [2 ]. And a large amount of complex work of manually processing input data is saved in the whole process, so that the prediction process is simpler, more convenient and more intelligent. On the basis, the problems of high cost and inaccurate measurement of hardware measurement are avoided, data obtained through soft measurement can be used for optimizing a combustion system on line, the combustion efficiency of a power plant boiler can be greatly improved, and the power generation cost of a power plant is reduced. However, the existing soft measurement method only analyzes and processes the acquired data, and the physical properties of the data are rarely concerned and expert experience is used, so that the deep research of the soft measurement method is still one of the hot spots and difficulties of intelligent optimization control of the power plant.
Disclosure of Invention
In order to solve the problems in the prior art, namely, the existing method for measuring the oxygen content of the flue gas only considers the characteristics of originally acquired data, and does not consider the physical significance and expert knowledge of the data; in addition, most soft measurement methods mostly rely on big data to carry out model training, the training efficiency of general models is low, and the prediction precision of the models is not high, the invention provides an attention mechanism and multilayer LSTM-based intelligent prediction method for the oxygen content of the flue gas of the power plant, which comprises the following steps:
(1) acquisition and storage of raw sensor data. And under different working conditions of the coal-fired power plant boiler, acquiring data of the sensor at a set frequency, and storing the data in a database.
(2) Raw sensor data preprocessing. The pretreatment mainly comprises two stages: data cleaning and data dimension reduction. The stored raw sensor data is denoted as a ═ a 1 ,a 2 ,…,a l ] T ∈R l×s And l and s respectively represent the number of data and the characteristic dimension of the data.
(3) And fusing data based on expert knowledge. Through expert knowledge and analysis of influence of collected historical coal amount of the A1-A6 coal mill on a boiler combustion system, the coal amount of the A1-A6 coal mill is subjected to weighted fusion as shown in the formula (3) to obtain the total coal amount.
Wherein T represents the total coal amount, T i Representing the amount of coal divided, gamma, for each coal mill i Representing a weight coefficient, which can be determined by expert knowledge analysis of historical data.
(4) Principal Component Analysis (PCA). Assuming that after data fusion based on expert knowledge, the data is recorded as phi ═ phi 1 ,φ 2 ,...,φ n ] T ∈R n×m And m respectively represents the characteristic dimension of the data after data fusion. Covariance matrix C of data Φ is calculated as follows
The covariance matrix C may be subjected to eigenvalue decomposition as follows
Q T CQ=Λ=diag(λ 1 ,λ 2 ,....,λ m ) (3)
Wherein λ is 1 >λ 2 ,>...,>λ m For the eigenvalues of the covariance matrix C, the corresponding eigenvector is denoted U ═ U 1 ,u 2 ,…,u m ]。
By setting the cumulative variance contribution rate threshold, d main features having a large influence on the output can be obtained. With the ith main feature z i For example, it is shown as follows
Wherein u is ij And phi ij Can be directly obtained from the feature vector U and the fused data phi.
(5) Attention is paid to the mechanism. Since different main features have different influences on the output, an attention network with a 3-layer structure is designed to calculate the attention weight of each feature vector and obtain a weighted feature vector. The 3-layer structure comprises a feature input layer, an attention weight calculation layer and a weight feature output layer.
The attention weight of the attention network is calculated as follows
Wherein F is an input vector, and F ═ F p ,F o ] T ,F p =[z 1 ,z 2 ,…,z d ]∈R 1×d As principal feature vector, F o =[F 1 ,F 2 ]Is the measured value of two smoke oxygen content measuring points, omega and beta are the super parameters of an attention network, and kappa i Indicating the ith attention weight.
Wherein the content of the first and second substances,meaning dot multiplication, i.e. multiplication of elements at corresponding positions of two matrices, k ═ k [ -k ] 1 ,κ 2 ,...,κ d ]∈R 1×d 。
(6) A predictive network of multi-layer LSTM. Considering that the LSTM is suitable for processing time series data, the invention builds a multi-layer LSTM network structure to predict the oxygen content of the smoke. For each LSTM structural unit, it contains 3 gate units: the input gate, the forgetting gate and the output gate, and the network hyper-parameter are updated as follows
Wherein i t Representing the time series of the inputs at time t, i.e. the weight feature vector, h, at time t t Indicating the hidden state of the LSTM structural unit at time t, f t Indicating a forgotten door state at time t, o t Denotes the output gate state at time t, W ═ W xi ,W hi ,W xf ,W hf ,W xo ,W ho ,W xu ,W hu ]And b ═ b i ,b f ,b o ]To predict the hyper-parameters of the network, which can be obtained by network training, σ (-) represents the sigmoid activation function.
(7) And (5) predicting network training. When the prediction network is trained, the following mean square error is adopted as a training loss function L of the network, and the training of the network is the minimum loss function.
Wherein the content of the first and second substances,to predict the output value of the network, N represents the number of data in batch data training in the network training process, and min (-) represents the minimization function.
(8) And updating the hyper-parameters. Based on the loss function definition (10) of the predicted network, the hyper-parameters of the predicted network are updated as follows
Where α is the parameter update rate, and ← represents the assignment.
Further, data cleaning, which is to perform data screening on the original sensor data by taking the power generation load as a reference standard of the data cleaning, specifically comprises: the filtering of adjacent data and the filtering of non-adjacent data,
the screening of the neighboring data is based on the following:
|a t,1 -a (t+1),1 |≤ε 1 (10)
where t and t +1 denote the sampling instants, a t,1 And a (t+1),1 Respectively representPower generation load, epsilon, at the respective sampling time 1 Representing adjacent data screening threshold, setting the threshold through analyzing historical collected data and expert experience, and if the difference value of the power generation loads of adjacent sampled data meets (1), then sampling data a at the moment of t +1 can be used t+1 The method comprises the following steps of removing the waste paper,
the screening of non-adjacent data is based on the following:
|a t,1 -a p,1 |≤ε 2 (11)
where t and p denote two non-adjacent sampling instants, a t,1 And a p,1 Respectively representing the power generation load, epsilon, at the respective sampling instants 2 Representing a screening threshold value of non-adjacent data, setting the threshold value by analyzing historical collected data and expert knowledge, and if the power generation load difference value of the non-adjacent sampled data meets (2), then sampling data a at the moment p can be used p The raw materials are removed,
and (4) reducing dimensions of the data, and assuming that the data is cleaned, expressing the data as phi ═ phi 1 ,φ 2 ,...,φ n ] T ∈R n×s And n and s respectively represent the number of data and the characteristic dimensionality of the data, and the data is subjected to expert knowledge-based data fusion and PCA principal component analysis processing to realize the dimensionality reduction of the data.
Further, the stored data has 32 quantities, including collecting time, power generation load, main steam pressure, main steam temperature, main steam flow, hearth negative pressure, A1-A6 coal pulverizer coal quantity, hearth total air quantity, blower air quantity of A and B sides, primary and secondary total air quantity, A side secondary air quantity, A side primary hot air, A side primary cold air, B side secondary air quantity, B side primary hot air, B side primary cold air, A side flue gas flow, B side flue gas flow, 3 detection point temperatures of the A side of the reactor, 3 detection point temperatures of the B side of the reactor, A side flue gas oxygen content and B side flue gas oxygen content, wherein the symbols A1-A6 are used for only identifying 6 different monitoring points, and A and B are used for identifying two main collecting points of the combustion boiler.
The invention has the beneficial effects that:
(1) the invention relates to an attention mechanism and multilayer LSTM-based power plant flue gas oxygen content intelligent prediction method. The invention firstly preprocesses the original sensor data, and the preprocessing process comprises the following steps: data cleaning and data dimension reduction, wherein the data cleaning mainly comprises the step of removing original sensor data with strong relevance, so that the redundancy among the data is reduced, and the representativeness of the data is improved; the data dimension reduction mainly reduces the degree of association between different features in the data, thereby reducing the interference of secondary features on network training and improving the representation capability of primary features. Then based on a designed attention mechanism network, distributing different attention weights to the main characteristics of the preprocessed data according to different degrees of influence on output to obtain weight characteristics, and further inputting the weight characteristics into a multilayer LSTM-based prediction network to predict the oxygen content of the smoke; with the continuous updating and upgrading of the control system of the coal-fired power plant, the application prospect and the social and economic benefits of the invention are considerable.
(2) The invention relates to an attention mechanism and multilayer LSTM-based power plant flue gas oxygen content intelligent prediction method. For the data cleaning process, the power generation load of a power plant is used as a reference standard, different cleaning threshold values are designed according to whether sampled data are adjacent or not, different data cleaning strategies are further executed on adjacent or non-adjacent data, and repeated or similar data in original data are removed.
(3) The invention relates to an attention mechanism and multilayer LSTM-based power plant flue gas oxygen content intelligent prediction method. For the data dimension reduction process, firstly, based on expert knowledge and the physical significance of different characteristics of collected data, fusing the coal quantity characteristics of a plurality of coal distributing ports, and establishing the total coal quantity characteristic influencing the oxygen content of flue gas; and further carrying out principal component analysis on the characteristics of the data by using a PCA (principal component analysis) technology to finally obtain main characteristics influencing the output of the prediction system. Through the designed data dimension reduction strategy, the dimension of the main features of the data is effectively reduced, and the representation capability of the main features of the data is improved.
(4) The invention relates to an attention mechanism and multilayer LSTM-based power plant flue gas oxygen content intelligent prediction method. For the attention mechanism, a 3-layer attention network structure is designed, wherein the first layer is a feature input layer, the second layer is an attention weight calculation layer, and the third layer is a weight feature output layer. According to the correlation degree of each main feature and the oxygen content of the smoke, different attention weights are distributed to each main feature, so that the prediction network is more inclined to select the main feature with higher correlation degree, and the training speed and the prediction accuracy of the model are improved.
(5) The invention relates to an attention mechanism and multilayer LSTM-based power plant flue gas oxygen content intelligent prediction method, and a prediction network based on an LSTM structure is built. Due to the unique network structure characteristics of the LSTM, the LSTM is more suitable for analyzing and processing data related to time series, and the built multilayer LSTM prediction network can greatly improve the precision of the prediction model while ensuring the training efficiency of the prediction model.
Drawings
Other features, objects and advantages of the present application will become more apparent upon reading of the following detailed description of non-limiting embodiments thereof, made with reference to the accompanying drawings in which:
FIG. 1 is a block diagram of a data processing flow of an embodiment of an intelligent prediction method for flue gas oxygen content in a power plant based on an attention mechanism and a multi-layer LSTM;
FIG. 2 is a schematic view of an attention network of an embodiment of the intelligent prediction method for the oxygen content of flue gas in a power plant based on an attention mechanism and multiple layers of LSTMs according to the invention;
FIG. 3 is a schematic diagram of an LSTM unit according to an embodiment of the intelligent prediction method for the oxygen content of the flue gas of the power plant based on an attention mechanism and multiple layers of LSTMs;
FIG. 4 is a graph of an accumulated variance ratio after dimension reduction of data according to an embodiment of the intelligent prediction method for oxygen content in flue gas of a power plant based on an attention mechanism and a multi-layer LSTM;
FIG. 5 is a graph of the variance/ratio of each characteristic quantity in the main characteristics of an embodiment of the intelligent prediction method for the oxygen content of the flue gas of the power plant based on the attention mechanism and the multi-layer LSTM.
FIG. 6 is a graph of a network training loss prediction change in an embodiment of the method for intelligently predicting the oxygen content of flue gas in a power plant based on an attention mechanism and multiple layers of LSTMs.
FIG. 7 is a graph of oxygen content prediction errors at A, B oxygen content monitoring points according to an embodiment of the method for intelligently predicting the oxygen content of power plant flue gas based on attention mechanism and multi-layer LSTM.
Detailed Description
The present application will be described in further detail with reference to the following drawings and examples. It is to be understood that the specific embodiments described herein are merely illustrative of the relevant invention and not restrictive of the invention. It should be noted that, for convenience of description, only the portions related to the related invention are shown in the drawings.
It should be noted that the embodiments and features of the embodiments in the present application may be combined with each other without conflict. The present application will be described in detail below with reference to the embodiments with reference to the attached drawings.
The invention provides an attention mechanism and multilayer LSTM-based power plant flue gas oxygen content intelligent prediction method which comprises two stages of data preprocessing and model prediction, wherein the data preprocessing comprises data cleaning and data dimension reduction, and the method is mainly used for reducing redundancy among originally acquired data and different characteristics in the data and improving the effectiveness and representation capability of the data. The model prediction comprises attention network construction and prediction network construction, wherein the attention network is mainly constructed according to the correlation degree between the main characteristics and the prediction model output, the prediction accuracy of the prediction network is further improved by giving different attention weights to different main characteristics, the prediction network based on the multi-layer LSTM is constructed on the basis that the LSTM has a strong prediction effect on time series data, and the prediction accuracy and reliability are ensured.
The invention relates to an attention mechanism and multilayer LSTM-based power plant flue gas oxygen content intelligent prediction method, originally acquired data of the method are from a 300-MW coal-fired power plant, and parameters of a computer processor are 12 th generation Gen Intel (R) core (TM) i9-12900K 3.20 GHz and Yingwei GeForce RTX 3090 GPU. The method comprises the following steps:
acquisition and storage of raw sensor data. Under different working conditions of the coal-fired boiler, data of the sensor are acquired at a set frequency and stored in a database, wherein the specifically stored data comprise 32 data including acquisition time, power generation load, main steam pressure, main steam temperature, main steam flow, hearth negative pressure, coal quantity of an A1-A6 coal mill, total hearth air quantity, fan air supply quantity of A side and B side, primary and secondary total air quantity, secondary air quantity of A side, primary hot air of A side, primary cold air of A side, secondary air quantity of B side, primary hot air of B side, primary cold air of B side, flue gas flow of A side, flue gas flow of B side, temperature of 3 detection points of A side of the reactor, temperature of 3 detection points of B side of the reactor, oxygen content of flue gas of A side and oxygen content of flue gas of B side. Wherein the symbols A1-A6 are used to identify only 6 monitoring points, and A and B are used to identify two main collection points for the power plant combustion boilers.
Raw sensor data preprocessing. The pretreatment mainly comprises two stages: data cleansing and data dimension reduction. The stored raw sensor data is denoted as a ═ a 1 ,a 2 ,…,a l ] T ∈R l×s And l and s respectively represent the number of data and the characteristic dimension of the data.
And (6) data cleaning. The method comprises the following steps of taking a power generation load as a reference standard for data cleaning, and screening data of an original sensor, wherein the method specifically comprises the following steps: the screening of adjacent data and the screening of non-adjacent data.
The screening of the neighboring data is based on the following:
|a t,1 -a (t+1),1 |≤ε 1 (12)
where t and t +1 denote the sampling instants, a t,1 And a (t+1),1 Respectively representing the power generation load, epsilon, at the respective sampling instants 1 0.1 represents an adjacent data screening threshold, and the threshold can be set by analyzing historical collected data and expert knowledge. If the power generation load difference value of adjacent sampling data satisfies (1), the sampling data a at the time point of t +1 can be used t+1 And (5) removing.
The screening of non-adjacent data is based on the following:
|a t,1 -a p,1 |≤ε 2 (13)
where t and p denote two non-adjacent sampling instants, a t,1 And a p,1 Respectively representing the power generation load, epsilon, at the respective sampling instants 2 0.05 represents a non-adjacent data screening threshold, ofThe settings may be made by analyzing historical collected data and expert knowledge. If the power generation load difference of the non-adjacent sampling data satisfies (2), the sampling data a at the time point p can be used p And (5) removing.
And (5) reducing the dimension of the data. Assuming that after data cleansing, the data is expressed as Φ ═ Φ 1 ,φ 2 ,...,φ n ] T ∈R n×s And n and s respectively represent the number of data and the characteristic dimension of the data. And carrying out data fusion based on expert knowledge and Principal Component Analysis (PCA) on the data to realize the dimensionality reduction of the data.
And fusing data based on expert knowledge. Through expert knowledge and analysis of influence of collected historical coal amount of the A1-A6 coal mill on a boiler combustion system, the coal amount of the A1-A6 coal mill is subjected to weighted fusion as shown in the formula (3) to obtain the total coal amount.
Wherein T represents the total coal amount, T i The weight of each coal mill is gamma 1 =0.16,γ 2 =0.14,γ 3 =0.14,γ 4 =0.13,γ 5 =0.16,γ 6 =0.17。
Principal Component Analysis (PCA). Assuming that data fusion based on expert knowledge is performed, the data is written as phi ═ phi 1 ,φ 2 ,...,φ n ] T ∈R n×m And m respectively represents the characteristic dimension of the data after data fusion. Covariance matrix C of data Φ is calculated as follows
The covariance matrix C may be subjected to eigenvalue decomposition as follows
Q T CQ=Λ=diag(λ 1 ,λ 2 ,....,λ m ) (16)
Wherein λ is 1 >λ 2 ,>...,>λ m For the eigenvalues of the covariance matrix C, the corresponding eigenvector is denoted U ═ U 1 ,u 2 ,…,u m ]。
By setting the cumulative variance contribution rate threshold, d main features having a large influence on the output can be obtained. With the ith main feature z i For example, it is shown as follows
Wherein u is ij And phi ij Can be directly obtained from the feature vector U and the fused data phi.
Attention is paid to the mechanism. Since different main features have different influences on the output, an attention network with a 3-layer structure is designed to calculate the attention weight of each feature vector and obtain a weighted feature vector. The 3-layer structure comprises a feature input layer, an attention weight calculation layer and a weight feature output layer.
The attention weight of the attention network is calculated as follows
Wherein F is an input vector, and F ═ F p ,F o ] T ,F p =[z 1 ,z 2 ,…,z d ]∈R 1×d As principal feature vector, F o =[F 1 ,F 2 ]Is the measured value of two smoke oxygen content measuring points, omega and beta are hyper-parameters of an attention network, and kappa i Indicating the ith attention weight.
Wherein the content of the first and second substances,meaning dot multiplication, i.e. multiplication of elements at corresponding positions of two matrices, k ═ k [ -k ] 1 ,κ 2 ,...,κ d ]∈R 1×d Is the attention weight vector.
A predictive network of multi-layer LSTM. Considering that the LSTM is suitable for processing time series data, the invention builds a multi-layer LSTM network structure to predict the oxygen content of the smoke. For each LSTM structural unit, it contains 3 gate units: the input gate, the forgetting gate and the output gate, and the network hyper-parameter are updated as follows
Wherein i t Representing the time series of inputs at time t, i.e. the feature vector with attention weights at time t, h t Indicating the hidden state of the LSTM structural unit at time t, f t Indicating the status of the forgotten door at time t, o t Indicates the state of the output gate at time t, b ═ b i ,b f ,b o ]And W ═ W xi ,W hi ,W xf ,W hf ,W xo ,W ho ,W xu ,W hu ]To predict the hyper-parameters of the network, which can be obtained by network training, σ (-) represents the sigmoid activation function.
And (5) predicting network training. When the prediction network is trained, the following mean square error is adopted as a training loss function L of the network, and the training of the network is the minimum loss function.
Wherein the content of the first and second substances,for predicting the output value of the network, N represents the batch data in the network training processThe number of data in the training, min (-) represents the minimization function.
And updating the hyper-parameters. Based on the loss function definition (10) of the predicted network, the hyper-parameters of the predicted network are updated as follows
Where α is the parameter update rate, and ← represents the assignment.
In one embodiment of the invention, the attention mechanism and multi-layer LSTM-based power plant flue gas oxygen content intelligent prediction system mainly comprises a prediction algorithm and each module function realization.
It can be clearly understood by those skilled in the art that, for convenience and brevity of description, the specific working process and related description of the system described above may refer to the corresponding process in the foregoing method embodiments, and will not be described herein again.
It should be noted that, the attention mechanism and multi-layer LSTM based power plant flue gas oxygen content intelligent prediction system provided in the above embodiment is only exemplified by the division of the above functional modules, and in practical applications, the above functions may be allocated to different functional modules according to needs, that is, the modules or steps in the embodiment of the present invention are further decomposed or combined, for example, the modules in the above embodiment may be combined into one module, or may be further split into multiple sub-modules, so as to complete all or part of the above described functions. The names of the modules and steps involved in the embodiments of the present invention are only for distinguishing the modules or steps, and are not to be construed as unduly limiting the present invention.
Those of skill in the art would appreciate that the various illustrative modules, method steps, and modules described in connection with the embodiments disclosed herein may be implemented as electronic hardware, computer software, or combinations of both, and that programs corresponding to the software modules, method steps may be located in Random Access Memory (RAM), memory, Read Only Memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art. To clearly illustrate this interchangeability of electronic hardware and software, various illustrative components and steps have been described above generally in terms of their functionality. Whether such functionality is implemented as electronic hardware or software depends upon the particular application and design constraints imposed on the solution. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present invention.
The symbols "side a", "side B", and "a 1-a 6" related to the present invention are only used to identify different acquisition points, and according to the difference of the embodiments, those skilled in the art may make modifications to the symbols, and no further description is given here.
The terms "comprises," "comprising," or any other similar term are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus.
So far, the technical solutions of the present invention have been described in connection with the preferred embodiments shown in the drawings, but it is easily understood by those skilled in the art that the scope of the present invention is obviously not limited to these specific embodiments. Equivalent changes or substitutions of related technical features can be made by those skilled in the art without departing from the principle of the invention, and the technical scheme after the changes or substitutions can fall into the protection scope of the invention.
Claims (3)
1. An attention mechanism and multilayer LSTM-based power plant flue gas oxygen content intelligent prediction method is characterized by comprising the following steps:
(1) the acquisition and storage of the original sensor data, under different working conditions of the coal-fired power plant boiler, the data of the sensor are acquired at a set frequency and stored in a database,
(2) raw sensor data preprocessingThe pretreatment mainly comprises two stages: data cleaning and data dimension reduction, and the stored original sensor data is marked as A ═ a 1 ,a 2 ,…,a l ] T ∈R l×s L and s respectively represent the number of data and the characteristic dimension of the data,
(3) based on the data fusion of expert knowledge, the coal quantity of the coal mill A1-A6 is subjected to weighted fusion as shown in the formula (3) to obtain the total coal quantity through the analysis of the influence of the expert knowledge and the collected history A1-A6 coal quantity on a boiler combustion system,
wherein T represents the total coal amount, T i Representing the amount of coal divided, gamma, for each coal mill i Representing a weight coefficient, which can be determined by expert knowledge analysis of historical data,
(4) PCA principal component analysis, assuming expert knowledge based data fusion, data is recorded as phi ═ phi 1 ,φ 2 ,...,φ n ] T ∈R n×m M represents the characteristic dimension of the data after data fusion, and the covariance matrix C of the data phi is calculated as follows
The covariance matrix C may be subjected to eigenvalue decomposition as follows
Q T CQ=Λ=diag(λ 1 ,λ 2 ,....,λ m ) (3)
Wherein λ is 1 >λ 2 ,>...,>λ m For the eigenvalues of the covariance matrix C, the corresponding eigenvector is denoted U ═ U 1 ,u 2 ,…,u m ],
D main features with large influence on output can be obtained by setting a threshold value of the cumulative variance contribution rate, and the ith main feature z is used i For example, it is shown as follows
Wherein u is ij And phi ij Can be directly obtained from the feature vector U and the fused data phi,
(5) the attention mechanism, because different main characteristics have different influences on the output, an attention network with a 3-layer structure is designed to calculate the attention weight of each characteristic vector and obtain the weight characteristic vector, the 3-layer structure comprises a characteristic input layer, an attention weight calculation layer and a weight characteristic output layer,
the attention weight of the attention network is calculated as follows
Wherein F is an input vector, and F ═ F p ,F o ] T ,F p =[z 1 ,z 2 ,…,z d ]∈R 1×d As principal feature vector, F o =[F 1 ,F 2 ]Is the measured value of two smoke oxygen content measuring points, omega and beta are the super parameters of an attention network, and kappa i The ith attention weight value is represented,
Wherein the content of the first and second substances,which represents a dot multiplication, i.e. a multiplication of elements at corresponding positions of the two matrices,
κ=[κ 1 ,κ 2 ,...,κ d ]∈R 1×d as a weight vector, the weight vector is,
(6) the invention discloses a multilayer LSTM prediction network, which takes the LSTM suitable for processing time series data into consideration, builds a multilayer LSTM network structure to predict the oxygen content of smoke, and comprises 3 gate units for each LSTM structure unit: the input gate, the forgetting gate and the output gate, and the network hyper-parameter are updated as follows
Wherein i t Representing the time series of inputs at time t, i.e. the feature vector with attention weights at time t, h t Indicating the hidden state of the LSTM structural unit at time t, f t Indicating the status of the forgotten door at time t, o t Indicates the state of the output gate at time t, b ═ b i ,b f ,b o ]And W ═ W xi ,W hi ,W xf ,W hf ,W xo ,W ho ,W xu ,W hu ]To predict the hyper-parameters of the network, which can be obtained by network training, σ (-) represents the sigmoid activation function,
(7) the prediction network training adopts the following mean square error as a training loss function L of the network when the network is predicted, the network training is the minimum loss function,
wherein, the first and the second end of the pipe are connected with each other,in order to predict the output value of the network, N represents the number of data in batch data training in the network training process, min (-) represents a minimization function,
(8) the hyper-parameter update of the prediction network is based on the loss function definition (10) of the prediction network as follows
Where α is the parameter update rate, and ← represents the assignment.
2. The intelligent prediction method for flue gas oxygen content of power plant based on attention mechanism and multi-layer LSTM as claimed in claim 1,
data cleaning, namely performing data screening on original sensor data by taking a power generation load as a reference standard of the data cleaning, and specifically comprising the following steps: the filtering of adjacent data and the filtering of non-adjacent data,
the screening of the neighboring data is based on the following:
|a t,1 -a (t+1),1 |≤ε 1 (10)
where t and t +1 denote the sampling instants, a t,1 And a (t+1),1 Respectively representing the power generation load, epsilon, at the respective sampling instants 1 Representing adjacent data screening threshold, setting the threshold through analyzing historical collected data and expert experience, and if the difference value of the power generation loads of adjacent sampled data meets (1), then sampling data a at the moment of t +1 can be used t+1 The raw materials are removed,
the screening of non-adjacent data is based on the following:
|a t,1 -a p,1 |≤ε 2 (11)
where t and p denote two non-adjacent sampling instants, a t,1 And a p,1 Respectively representing the power generation load, epsilon, at the respective sampling instants 2 Representing a screening threshold value of non-adjacent data, setting the threshold value by analyzing historical collected data and expert knowledge, and if the power generation load difference value of the non-adjacent sampled data meets (2), then sampling data a at the moment p can be used p The raw materials are removed,
and (4) reducing dimensions of the data, and assuming that the data is cleaned, expressing the data as phi ═ phi 1 ,φ 2 ,...,φ n ] T ∈R n×s N andand s respectively represents the number of data and the characteristic dimensionality of the data, and the data is subjected to expert knowledge-based data fusion and PCA principal component analysis processing to realize the dimensionality reduction of the data.
3. The method of claim 1 for intelligent prediction of flue gas oxygen content in power plants based on attention mechanism and multi-layer LSTM, the method is characterized in that 32 data are stored, including acquisition time, power generation load, main steam pressure, main steam temperature, main steam flow, hearth negative pressure, coal quantity of an A1-A6 coal pulverizer, total air quantity of a hearth, air quantity of fans at A side and B side, total primary air quantity and secondary air quantity, secondary air quantity at A side, primary hot air at A side, primary cold air at A side and secondary air quantity at B side, primary hot air at B side, primary cold air at B side, flue gas flow at A side, flue gas flow at B side, temperature of 3 detection points at A side of a reactor, temperature of 3 detection points at B side of the reactor, oxygen content of flue gas at A side and oxygen content of flue gas at B side, wherein the symbols A1-A6 are used only to identify 6 different monitoring points, A and B being for the purpose of identifying two main collection points of the fired boiler.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210694964.5A CN115017818A (en) | 2022-06-20 | 2022-06-20 | Power plant flue gas oxygen content intelligent prediction method based on attention mechanism and multilayer LSTM |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210694964.5A CN115017818A (en) | 2022-06-20 | 2022-06-20 | Power plant flue gas oxygen content intelligent prediction method based on attention mechanism and multilayer LSTM |
Publications (1)
Publication Number | Publication Date |
---|---|
CN115017818A true CN115017818A (en) | 2022-09-06 |
Family
ID=83075264
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202210694964.5A Pending CN115017818A (en) | 2022-06-20 | 2022-06-20 | Power plant flue gas oxygen content intelligent prediction method based on attention mechanism and multilayer LSTM |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN115017818A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117369263A (en) * | 2023-10-23 | 2024-01-09 | 苏州大学 | Intelligent combustion control method of hot blast stove based on reinforcement learning and attention mechanism |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112488145A (en) * | 2019-11-26 | 2021-03-12 | 大唐环境产业集团股份有限公司 | NO based on intelligent methodxOnline prediction method and system |
CN113487064A (en) * | 2021-06-10 | 2021-10-08 | 淮阴工学院 | Photovoltaic power prediction method and system based on principal component analysis and improved LSTM |
CN113902202A (en) * | 2021-10-15 | 2022-01-07 | 南京工程学院 | Short-term load prediction model and method based on double attention mechanism and LSTM |
CN113935230A (en) * | 2021-09-14 | 2022-01-14 | 国网河北省电力有限公司电力科学研究院 | Implementation of NO based on attention mechanism LSTM modelxEmission amount prediction method |
CN114219139A (en) * | 2021-12-07 | 2022-03-22 | 国网湖北省电力有限公司宜昌供电公司 | DWT-LSTM power load prediction method based on attention mechanism |
-
2022
- 2022-06-20 CN CN202210694964.5A patent/CN115017818A/en active Pending
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112488145A (en) * | 2019-11-26 | 2021-03-12 | 大唐环境产业集团股份有限公司 | NO based on intelligent methodxOnline prediction method and system |
CN113487064A (en) * | 2021-06-10 | 2021-10-08 | 淮阴工学院 | Photovoltaic power prediction method and system based on principal component analysis and improved LSTM |
CN113935230A (en) * | 2021-09-14 | 2022-01-14 | 国网河北省电力有限公司电力科学研究院 | Implementation of NO based on attention mechanism LSTM modelxEmission amount prediction method |
CN113902202A (en) * | 2021-10-15 | 2022-01-07 | 南京工程学院 | Short-term load prediction model and method based on double attention mechanism and LSTM |
CN114219139A (en) * | 2021-12-07 | 2022-03-22 | 国网湖北省电力有限公司宜昌供电公司 | DWT-LSTM power load prediction method based on attention mechanism |
Non-Patent Citations (2)
Title |
---|
JIAN SUN,等: "Prediction of Oxygen Content Using Weighted PCA and Improved LSTM Network in MSWI Process", IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, vol. 70, 31 December 2021 (2021-12-31), pages 1 - 12, XP011850233, DOI: 10.1109/TIM.2021.3058367 * |
刘慧敏,等: "基于PCA-Attention-LSTM网络的土壤氮含量监测", 中国农机化学报, vol. 41, no. 09, 15 September 2020 (2020-09-15), pages 190 - 197 * |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117369263A (en) * | 2023-10-23 | 2024-01-09 | 苏州大学 | Intelligent combustion control method of hot blast stove based on reinforcement learning and attention mechanism |
CN117369263B (en) * | 2023-10-23 | 2024-07-09 | 苏州大学 | Intelligent combustion control method of hot blast stove based on reinforcement learning and attention mechanism |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN113128781B (en) | Distributed industrial energy operation optimization platform for automatically constructing intelligent model and algorithm | |
Xue et al. | Multi-step ahead forecasting of heat load in district heating systems using machine learning algorithms | |
Tang et al. | The multi-objective optimization of combustion system operations based on deep data-driven models | |
CN112070109A (en) | Calla kiln energy consumption abnormity detection method based on improved density peak clustering | |
Peña et al. | Soft-computing models for soot-blowing optimization in coal-fired utility boilers | |
CN109948858B (en) | Working condition identification method based on data driving | |
CN116167657A (en) | Enterprise full life cycle carbon footprint evaluation method and system and storage medium | |
Dong et al. | The prediction trend of enterprise financial risk based on machine learning arima model | |
Yin et al. | An integrated computational intelligence technique based operating parameters optimization scheme for quality improvement oriented process-manufacturing system | |
Geng et al. | Energy optimization and analysis modeling based on extreme learning machine integrated index decomposition analysis: application to complex chemical processes | |
CN113627735A (en) | Early warning method and system for safety risk of engineering construction project | |
CN112801388A (en) | Power load prediction method and system based on nonlinear time series algorithm | |
Liu et al. | Model fusion and multiscale feature learning for fault diagnosis of industrial processes | |
CN115936061A (en) | Thermal power plant flue gas oxygen content soft measurement method and system based on data driving | |
CN115017818A (en) | Power plant flue gas oxygen content intelligent prediction method based on attention mechanism and multilayer LSTM | |
Yuan et al. | Spatiotemporal attention mechanism-based deep network for critical parameters prediction in chemical process | |
CN115688581A (en) | Oil gas gathering and transportation station equipment parameter early warning method, system, electronic equipment and medium | |
CN116258055A (en) | Online prediction method for unit load and dynamic adjustment capability in automatic power generation control | |
Lim et al. | Smart soft-sensing for the feedwater flowrate at PWRs using a GMDH algorithm | |
Rahim et al. | Cross-Validation and Validation Set Methods for Choosing K in KNN Algorithm for Healthcare Case Study | |
Panjapornpon et al. | Energy efficiency and savings analysis with multirate sampling for petrochemical process using convolutional neural network-based transfer learning | |
CN116821695B (en) | Semi-supervised neural network soft measurement modeling method | |
CN111563631A (en) | Wind power generation power prediction method and device based on variational modal decomposition | |
CN116401545A (en) | Multimode fusion type turbine runout analysis method | |
CN114565209A (en) | Process industry energy consumption state evaluation method based on clustering |
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 |