CN115018158A - SCR (Selective catalytic reduction) outlet NOx emission prediction method based on BWOA-BiGRU-LAM (lean-reactive inert gas) - Google Patents

SCR (Selective catalytic reduction) outlet NOx emission prediction method based on BWOA-BiGRU-LAM (lean-reactive inert gas) Download PDF

Info

Publication number
CN115018158A
CN115018158A CN202210637691.0A CN202210637691A CN115018158A CN 115018158 A CN115018158 A CN 115018158A CN 202210637691 A CN202210637691 A CN 202210637691A CN 115018158 A CN115018158 A CN 115018158A
Authority
CN
China
Prior art keywords
nox emission
bigru
lam
bwoa
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
Application number
CN202210637691.0A
Other languages
Chinese (zh)
Inventor
武松
马永光
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
North China Electric Power University
Original Assignee
North China Electric Power University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by North China Electric Power University filed Critical North China Electric Power University
Priority to CN202210637691.0A priority Critical patent/CN115018158A/en
Publication of CN115018158A publication Critical patent/CN115018158A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • 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
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • 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
    • 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
    • Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T10/00—Road transport of goods or passengers
    • Y02T10/10—Internal combustion engine [ICE] based vehicles
    • Y02T10/40—Engine management systems

Landscapes

  • Engineering & Computer Science (AREA)
  • Business, Economics & Management (AREA)
  • Economics (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Health & Medical Sciences (AREA)
  • Strategic Management (AREA)
  • Human Resources & Organizations (AREA)
  • General Physics & Mathematics (AREA)
  • General Business, Economics & Management (AREA)
  • Tourism & Hospitality (AREA)
  • Marketing (AREA)
  • General Health & Medical Sciences (AREA)
  • Software Systems (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Mathematical Physics (AREA)
  • Computing Systems (AREA)
  • Molecular Biology (AREA)
  • Evolutionary Computation (AREA)
  • Development Economics (AREA)
  • Data Mining & Analysis (AREA)
  • Game Theory and Decision Science (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • General Engineering & Computer Science (AREA)
  • Biomedical Technology (AREA)
  • Operations Research (AREA)
  • Quality & Reliability (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Public Health (AREA)
  • Water Supply & Treatment (AREA)
  • Primary Health Care (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses an SCR outlet NOx emission prediction method based on BWOA-BiGRU-LAM, which adopts a deep neural network algorithm based on data drive and can realize automatic optimization of network parameters, and comprises the following steps: preprocessing the NOx emission sensitive parameters and corresponding NOx emission historical data, estimating variable delay time, and reconstructing a data sequence; fully extracting the characteristics of a modeling data sequence by combining a BiGRU network with an LAM mechanism, automatically distributing hidden layer neuron output weights, and highlighting the influence of key information; a BWOA algorithm is used for automatically optimizing four hyper-parameters such as a network learning rate, the number of hidden layer neurons, a small batch size, a Dropout coefficient and the like in the BiGRU-LAM network training process, so that the prediction capability of the model is fully mined, and the prediction accuracy is improved.

Description

SCR (Selective catalytic reduction) outlet NOx emission prediction method based on BWOA-BiGRU-LAM (lean-reactive inert gas)
Technical Field
The invention belongs to the technical field of SCR flue gas denitration of coal-fired power plants, and particularly relates to a BWOA-BiGRU-LAM-based SCR outlet NOx emission prediction method.
Background
The coal-fired power plant is the main body of the power production industry in China, and is not only an energy-consuming household but also a significant emission source of atmospheric pollutants such as nitrogen oxides (NOx). Especially, with the proposal of the target of 'double carbon', the state continuously improves the emission standard.At present, the ultra-low limit emission standard of NOx in coal-fired power plants in China is 50mg/m 3 Stringent to the united states and european union standards. The NOx control technology mainly includes two types of control during combustion and control after combustion. The method for controlling in the combustion, namely, the method of staged combustion or burner improvement and the like is adopted to reduce NOx generated in the combustion process, the implementation cost of the measure is low, certain negative effects exist on the boiler, the denitration efficiency is not high, and the strict emission standard of China cannot be met when the measure is used alone; the other is post-combustion control, i.e., treatment of the generated NOx to reduce the NOx emission. At present, the emission of NOx is reduced by widely using a flue gas denitration technology of a Selective Catalytic Reduction (SCR) method, ammonia gas is used as a reducing agent, and the ammonia gas and the NOx in the flue gas are subjected to a Reduction reaction under the Catalytic action of a catalyst to convert the NOx into harmless nitrogen and water. Although the SCR denitration efficiency can reach more than 90%, the SCR reaction process has the characteristics of large delay, large inertia and nonlinearity, and the NOx concentration measuring equipment has certain time delay, so that the ammonia injection amount cannot be quickly adjusted to match the change of the NOx concentration, and the insufficient catalytic reduction reaction is caused by too little ammonia injection amount, so that the smoke emission does not reach the standard; excessive ammonia spraying amount can cause the increase of ammonia escape amount, secondary pollution to the environment, corrosion to downstream equipment and increase the operation cost of enterprises. Especially, the unit is operated under the working conditions of low load and variable load, and the adverse effect is very prominent at the moment.
In order to accurately control the ammonia injection amount, improve the denitration efficiency and reduce the NOx emission, an accurate NOx emission prediction model is established as a basis. The NOx generation mechanism is quite complex, and the traditional mechanism modeling method is difficult to accurately predict the NOx emission. With the rapid development of data-driven modeling and artificial intelligence methods, more and more researchers are focusing on the field of intelligent modeling. Common intelligent modeling methods include Artificial Neural Network (ANN), Support Vector Machine (SVM), Deep Learning Network (DNN), and the like, where DNN includes Deep Neural Network algorithms such as Convolutional Neural Network (CNN), Long-Short Term Memory Network (LSTM), Deep Belief Network (DBN), and the like. The ANN and the SVM belong to a shallow learning method, the depth characteristics of variables cannot be effectively learned, the prediction accuracy of the built model is general, and the generalization capability is poor; although deep neural networks such as CNN and LSTM can fully extract deep level features of variables and improve prediction accuracy, when an input sequence is long, early information is easily lost in a model, overfitting is easily caused in prediction output, and network parameters such as learning rate and neuron number are often manually adjusted by depending on personal experience, so that the method has strong subjective randomness.
In summary, how to establish a more accurate prediction model of the NOx emission at the outlet of the SCR and realize automatic optimization of model parameters is an urgent problem to be solved in the denitration control process of the coal-fired power plant.
Disclosure of Invention
In order to solve the problems, the invention discloses a method and a device for predicting SCR outlet NOx emission based on BWOA-BiGRU-LAM. In order to achieve the purpose, the technical scheme of the invention is as follows:
an SCR outlet NOx emission prediction method based on BWOA-BiGRU-LAM comprises the following steps:
acquiring NOx emission sensitive parameters and corresponding NOx emission historical data samples to form a data set;
cleaning the data set, and removing abnormal values and noises to obtain a cleaned data set;
calibrating time delay between each NOx emission sensitive parameter and NOx emission amount to form a corresponding relation between each NOx emission sensitive parameter and a NOx emission amount historical data sample to construct a data set; dividing a data set into a training set and a verification set;
fourthly, constructing a BiGRU-LAM network, inputting a training set, setting the value ranges of four network parameters of learning rate, minimum batch quantity, hidden layer neuron quantity and regularization coefficient, training the network, automatically optimizing the learning rate, the minimum batch quantity, the hidden layer neuron quantity and the regularization coefficient in the given ranges by adopting BWOA, inputting the BiGRU-LAM network after obtaining optimal parameter values, and forming a final prediction model BWOA-BiGRU-LAM after verification of a verification set;
and step five, collecting and collecting NOx emission sensitive parameters in real time, inputting the parameters into a final prediction model BWOA-BiGRU-LAM after time delay correction between each NOx emission sensitive parameter and NOx emission, and predicting the NOx emission in real time.
In a further improvement, the sensitive parameters of the NOx emission comprise SCR inlet oxygen content, SCR inlet flue gas flow, SCR inlet NOx concentration, SCR inlet flue gas temperature, SCR outlet ammonia gas concentration, SCR outlet oxygen content and SCR mixed ammonia gas flow.
In a further improvement, in the second step, the method for cleaning the data set includes:
and screening abnormal values in the data set by using a boxplot, then correcting the abnormal values by using a mean filling method, removing noise in the data set by using a wavelet semi-soft threshold denoising algorithm, and then standardizing by using a Z-Score method.
In a further improvement, in the third step, the method for calibrating the time delay between each NOx emission sensitive parameter and the NOx emission amount is as follows:
assuming that X and Y are two discrete random variables, the mutual information I (X, Y) between X and Y is:
Figure BDA0003682259420000031
wherein p (X, Y) is the joint probability distribution of X and Y, and p (X), p (Y) is the edge probability distribution of X and Y; in order to calculate the mutual information value between input variable and output NOx formed by each NOx emission sensitive parameter, an input variable time sequence matrix is set as X (t), SCR outlet NOx measured value is set as an output time sequence matrix Y (t), and the method comprises the following steps:
X(f)＝[x 1 (t)，x 2 (t)，Λ，x n (t)] (18)
wherein n is X (t) dimension, namely the number of variables;
setting τ i (τ i ∈[τ min ,τ max ]) For the ith (i epsilon (1, n)) variable x i Of (t) and Y (t)The time lag between, x (t), can be reconstructed as:
Figure BDA0003682259420000032
wherein the ith variable x i (t) reconstruction as:
Figure BDA0003682259420000033
according to the actual operation condition of the SCR system, the variable delay time range is set between 0s and 300s, the sampling period is 1min, and therefore tau is selected min ＝0，τ max (ii) 5; to find each input variable x i (t) calculating the optimal delay time by the following steps:
(a) calculating the mutual information value between each row of the matrix on the right side of the medium sign in the formula (20) and the output Y (t), wherein the delay time tau corresponding to the row with the maximum mutual information value is the optimal delay time estimation of the variable;
(b) repeating step (a) for delay times for all NOx emission sensitive parameters.
In a further improvement, the method for constructing the BiGRU-LAM network in the fourth step is as follows:
constructing a BiGRU network, wherein the BiGRU network comprises two GRU networks with the same structure, and the data processing directions of the two GRU networks are opposite; and the outputs of the gate control circulation units corresponding to each other in the two GRU networks are spliced and then used as the input of the LAM network.
Further improvement, the specific calculation process of the GRU network is as follows:
r t ＝sigmoid(W r ·[h t-1 ，x t ]+b r ) (1)
z t ＝sigmoid(W z ·[h t-1 ，x t ]+b z ) (2)
Figure BDA0003682259420000041
Figure BDA0003682259420000042
in the formula, x t For input at the present moment, h t-1 Is a hidden state of the neuron at the previous moment, r t And z t Outputs of reset and update gates, W, respectively, at the present moment r 、W z And W h Respectively a reset gate weight matrix, an update gate weight matrix, a neuron output weight matrix, b r 、b z And b h Respectively a reset gate bias term matrix, an updated gate bias term matrix and a neuron output bias term matrix,
Figure BDA0003682259420000043
candidate hidden states for the neuron at the current time, h t For the final state output at the current time, represent a matrix multiplication,
Figure BDA0003682259420000044
representing multiplication of corresponding elements of the matrix, wherein tanh and sigmoid are respectively a tanh activation function and a sigmoid activation function;
the specific calculation process of LAM is as follows:
Figure BDA0003682259420000045
Figure BDA0003682259420000046
Figure BDA0003682259420000047
Figure BDA0003682259420000048
wherein, the first and the second end of the pipe are connected with each other,a ts attention weight of neuron output hidden state at historical time, namely s time, to neuron output at current time, namely t time, score is a scoring function,
Figure BDA0003682259420000049
is the hidden state of the historical output of the neuron, h t Is the hidden state of the neuron output at the current moment, c t Is the intermediate variable that is the variable between,
Figure BDA00036822594200000410
is the final output of the neuron at the current moment after attention weighted calculation, W c Is a weight matrix.
The invention has the advantages that:
1) compared with other prediction models, the method can realize accurate prediction of the NOx emission at the SCR outlet at the future moment, has good model generalization capability and small prediction result error, thereby compensating the inertia and delay of the SCR system, leading ammonia injection to act in advance and improving the denitration efficiency of the system.
2) The invention uses the BiGRU network and introduces LAM, the BiGRU can utilize the information of the two directions of the sequence, the LAM automatically distributes weight to the historical output of the hidden layer neuron, the influence of important information is highlighted, and the model prediction accuracy is improved.
3) The invention uses BWOA to automatically optimize the four parameters of the network in the process of BiGRU-LAM network training, overcomes the defect that the parameters are manually adjusted by depending on personal experience in the prior art, and improves the model prediction precision.
4) According to the method, before model training, abnormal values in data acquired by adopting a box-line graph method and mean filling processing are processed, the noise in the data is processed by Wavelet transform Semi-soft Threshold Denoising (WSTD), then time delay between an input variable and NOx is estimated by Mutual Information (MI) theory, and a data sequence is reconstructed, so that the data precision is improved, and the influence of the data on the model precision in data drive modeling is fully reduced.
Drawings
Fig. 1 is a diagram of a GRU network neuron cell structure.
Fig. 2 is a diagram of a BiGRU network architecture.
Fig. 3 is a schematic view of LAM principle.
FIG. 4 is a BWOA flowchart.
FIG. 5 is a diagram of the BWOA-BiGRU-LAM model.
FIG. 6 is a box plot algorithm schematic.
FIG. 7 is a flow chart of wavelet transform semi-soft threshold denoising.
FIG. 8 is a schematic diagram of the BWOA-BiGRU-LAM model prediction results.
FIG. 9 is a comparison of the predicted results of the four models.
FIG. 10 is a comparison graph of the linear fit of the four model predictions.
FIG. 11 is a diagram showing absolute error comparison of four model prediction results.
Detailed Description
The invention is further explained with reference to the drawings and the embodiments.
Firstly, constructing and constructing a BiGRU-LAM network:
1.1 GRU and BiGRU networks
The Gated Recurrent Unit (GRU) is a Recurrent Neural Network (RNN) for processing time series, which is an improvement of LSTM, and has only two gate structures inside, namely Reset gate and Update gate, which not only maintains the effect of LSTM, but also has fewer parameters and better convergence than LSTM. GRU network neuron refinement
The cell structure is shown in FIG. 1.
The specific calculation process of the GRU network is as follows:
r t ＝sigmoid(W r ·[h t-1 ，x t ]+b r ) (1)
z t ＝sigmoid(W z ·[h t-1 ，x t ]+b z ) (2)
Figure BDA0003682259420000061
Figure BDA0003682259420000062
in the formula, x t For input at the present moment, h t-1 Is a hidden state of the neuron at the previous moment, r t And z t The outputs of the reset gate and the update gate, respectively, for the current time, W is the weight matrix, b is the bias term matrix,
Figure BDA0003682259420000063
candidate hidden states for the neuron at the current time, h t For the final state output at the current time, represent a matrix multiplication,
Figure BDA0003682259420000064
the multiplication of corresponding elements of the matrix is shown, and tanh and delta (sigmoid) are activation functions.
The BiGRU network is expanded on the basis of the GRU network and is formed by splicing two GRU networks with the same structure. The GRU can only use information of one direction of the time sequence, while the BiGRU is bidirectional and can use information of two directions before and after the sequence. The BiGRU network structure is shown in fig. 2.
1.2 Luong attention mechanism LAM
The LAM constructs an Attention Layer (Attention Layer) on the basis of a classic Seq2Seq Encoder-Decoder structure, and distributes more Attention weights to parts with large association degrees by calculating the association degree between neuron history hidden states and hidden states at the current moment so as to realize the self-adaptive Attention to data characteristics. The LAM structure is shown in fig. 3.
Figure BDA0003682259420000065
Figure BDA0003682259420000066
Figure BDA0003682259420000067
Figure BDA0003682259420000068
In the formula, a ts The attention weight of the hidden state output for the neuron at the historical moment (s moment) to the current moment (t moment), score is a scoring function,
Figure BDA0003682259420000071
is the hidden state of the historical output of the neuron, h t Is the hidden state of the neuron output at the current moment, c t Is the intermediate variable(s) of the,
Figure BDA0003682259420000072
and finally outputting the neuron at the current moment after attention weighting calculation.
2.3 Black widow optimization Algorithm BWOA
BWOA is inspired by the unique mating behavior of the black widow spider, and the algorithm simulates the life cycle of the black widow spider. Experiments prove that the BWOA has the characteristics of high convergence rate, high precision and the like. The BWOA algorithm flow chart is shown in fig. 4. The BWOA calculation steps are as follows:
(a) the method comprises the steps of setting the population quantity of black widow spiders (all female spiders), determining parameters to be optimized and a value range of each black widow spider, randomly initializing each black widow within the value range, and determining the maximum iteration times and a fitness function.
(b) And (4) updating the position. And determining the current optimal black widow spider position according to the fitness function value, and then updating the position. The black widow spider moves in a linear and spiral manner in the grid, and the position is updated as shown in formula (9).
Figure BDA0003682259420000073
In the formula (I), the compound is shown in the specification,X i (t +1) updated Black widow spider position, X best For the current optimal black oligogynic position, m is [0.4,0.9 ]]Beta is [ -1,1 ] or]Random number therebetween, rand is [0,1 ]]Random number between, X r1 (t) is the position of the r1 th black oligowoman in a randomly selected population, X i (t) is the current black widow spider position.
(c) And (4) calculating pheromones as shown in a formula (10). Pheromones play a very important role in the process of coupling of spiders. When the pheromone value is equal to or less than 0.3, the female black widow spider is represented as a hungry spider (the female spider eats the male spider), the female spider will not be selected by a male in the process of seeking a spouse, but will be replaced by another female spider, and the position of the black widow spider at the moment is updated according to the formula (11).
Figure BDA0003682259420000074
X i (t)＝X best +0.5[X r1 -(-1) σ X r2 (t)] (11)
Wherein pheromone (i) is the ith black oligogyne pheromone, fitness max And fitness min For the worst and best fitness function values, fitness (i) is the ith black widow spider fitness value. X i (t) the location of the black widow spider with low pheromone value, r1 and r2 the random numbers within the population number (r1 ≠ r2), X r1 And X r2 The r1 th and r2 th black widow positions are sigma is a random binary number of 0 or 1.
(d) And re-evaluating the fitness function value, and updating the position of the optimal black and wife and the optimal fitness function value.
(e) And (c) repeating the steps (b), (c) and (d), and carrying out iterative operation until a stopping condition (the maximum iteration number or the allowable error is reached) is met, and outputting the optimal position of the black and oligogynes, namely the optimal value of the parameter to be optimized.
1.4 model construction
The BWOA-BiGRU-LAM model was constructed as shown in FIG. 5. The method comprises the steps that BWOA automatically optimizes four parameters including a network learning rate, the number of hidden layer neurons, the minimum batch number and a Dropout coefficient, a BiGRU receives an input sequence and extracts features, meanwhile, output weight distribution of a BiGRU neuron hidden state is calculated through LAM, and a final output result is obtained through full connection. After the model is trained, test data is input for NOx prediction.
1.5 example analysis and comparison
Example data come from historical operating data of a Distributed Control System (DCS) of a certain 2X 300MW power station for 7 continuous days, covering low, medium and high loads, and a sampling period of 1min, wherein 10080 data samples are obtained.
1.5.1 data preprocessing
The data-driven modeling is very sensitive to data precision and has higher requirements on data quality. The SCR measuring equipment is in a severe production environment, signal transmission delay is large, and historical data stored by the DCS is prone to have a small part of outliers and various noises such as Gaussian noise, measurement noise and SCR system operation noise due to reasons such as measuring equipment faults, so that experimental data are preprocessed before modeling, and influences of the experimental data on model accuracy are eliminated.
For outliers, boxplots (Box-while Plot) were used for screening, and then mean-filling was used to correct outliers. The box plot configuration is shown in fig. 6.
The box plot method comprises the following calculation steps:
a. the upper quartile (Q3), the middle quartile, and the lower quartile (Q1) are calculated.
b. Calculating the difference between the upper quartile and the lower quartile, i.e., the quartile difference (IQR) Q3-Q1;
c. numbers between Q1-3 and Q3+3 IQR are acceptable values, and data outside the range are designated as outliers.
For noise processing, the invention adopts a Wavelet Semi-soft Threshold Denoising (WSTD) algorithm to improve the quality of data. The semi-soft threshold is between the soft threshold and the hard threshold, so that the problems of local signal jitter caused by the denoising of the hard threshold and low signal-to-noise ratio of the denoising peak value of the soft threshold can be balanced. The signals obtained after wavelet reconstruction are smooth and have small errors. The process is shown in fig. 7.
The wavelet transform is defined as:
Figure BDA0003682259420000091
in the formula, CWT is a continuous wavelet transform coefficient, ψ (t) is a wavelet mother wave, α is a scale parameter, and τ is a shift parameter.
The invention adopts db4 wavelet as mother wave, and Mallat algorithm to decompose original signal by 4 layers to obtain 1 approximate component and 4 detail components, the half-soft threshold function f (x) and threshold lambda are selected as follows:
Figure BDA0003682259420000092
Figure BDA0003682259420000093
wherein N is the signal length.
Because the variable dimensions are different, some variable values are very large, and some variable values are very small, which brings difficulty to model learning. Therefore, after removing abnormal values and noise, the data is normalized and converted into a dimensionless expression, so that the convergence speed of the model is improved. The present invention normalizes the data using the Z-Score method.
Figure BDA0003682259420000094
Figure BDA0003682259420000095
Where μ is the data sample mean and σ is the standard deviation. x is the raw data and x' is the normalized data.
1.5.2 variable screening and delay time estimation
The raw data was processed using the method in 1.5.1. Through the analysis of the operation mechanism of the SCR system and the combination of the actual experience of field personnel, 7 most main factors influencing the NOx emission quantity of an SCR outlet are found out as model input variables. The SCR has the characteristics of large delay and large inertia, and different time delays exist between the measurement of each input characteristic variable and the NOx measurement value, so that an accurate corresponding relation is not formed between all relevant parameters in the operation data recorded by the DCS at the same moment, and therefore the time delay between the input variable and the output NOx is required to be calibrated before modeling, an original data sequence is reconstructed, and the alignment of the modeling data in a time dimension is realized.
Mutual Information (MI) is a quantitative representation of Information that reflects the degree of correlation between variables. Defining mutual information between two discrete random variables X and Y as:
Figure BDA0003682259420000101
wherein p (X, Y) is the joint probability distribution of X and Y, and p (X), p (Y) is the edge probability distribution of X and Y. In order to calculate the mutual information value between input variable and output NOx formed by each NOx emission sensitive parameter, an input variable time sequence matrix is set as X (t), SCR outlet NOx measured value is set as an output time sequence matrix Y (t), and the method comprises the following steps:
X(t)＝[x 1 (t),x 2 (t),Λ,x n (t)] (18)
where n is X (t) dimension (number of variables).
Setting τ i (τ i ∈[τ min ,τ max ]) For the ith (i epsilon (1, n)) variable x i The time delay between (t) and y (t), x (t) can be reconstructed as:
Figure BDA0003682259420000102
wherein the ith variable x i (t) reconstructing into:
Figure BDA0003682259420000103
according to the actual operation condition of the SCR system, the variable delay time range is set between 0s and 300s, the sampling period is 1min, and therefore tau is selected min ＝0，τ max (ii) 5; to find each input variable x i (t) calculating the optimal delay time by the following steps:
(a) calculating the mutual information value between each row of the matrix on the right side of the medium sign in the formula (20) and the output Y (t), wherein the delay time tau corresponding to the row with the maximum mutual information value is the optimal delay time estimation of the variable;
(b) repeating step (a) for delay times for all NOx emission sensitive parameters.
The delay time estimates for the calculated variables are shown in table 1 below.
TABLE 1 variable delay time estimation
Figure BDA0003682259420000104
Figure BDA0003682259420000111
And reconstructing the data sequence after obtaining the delay time of the variable, and aligning in the time dimension. In order to further improve the accuracy of model prediction, historical moment data of SCR outlet NOx emission is added on the basis of the historical moment data of the selected 7 input variables, and 8 variable historical moment data are used as the final input of the model, so that the SCR outlet NOx emission in a future time step (namely 1 sampling period and 1min) is dynamically predicted. In addition, the former 10000 pieces of data are selected from the determined final historical data as final modeling data, the proportion of a training set, a verification set and a test set is 7:2:1, namely, the former 7000 groups are selected as model training set data, the latter 2000 groups are selected as verification set data to enable BWOA to optimize network parameters, and the last 1000 groups are selected as test set data to test the prediction capability of the model.
1.5.3 evaluation index
To quantify the predictive power of the model, a Symmetric Mean Absolute Percent Error (SMAPE), a Root Mean Square Error (RMSE), a determinant coefficient (R-Square ) were selected herein 2 ) As an evaluation index for checking the accuracy of the model. Wherein, the smaller the SMAPE and RMSE values are, the higher the model prediction precision is, and R 2 The closer to 1 (R) 2 E (0,1)) indicates that the higher the data tracking capability, the better the curve fitting.
The specific calculation formula is as follows:
Figure BDA0003682259420000112
Figure BDA0003682259420000113
Figure BDA0003682259420000114
in the formula, y i Is an actual measured value of the NOx at the outlet of the SCR,
Figure BDA0003682259420000115
in order to predict the value of the model,
Figure BDA0003682259420000116
is the mean measured NOx value and n is the total number of samples.
1.5.4 model parameter settings
Setting the value ranges of the four network parameters in 7.4, initializing the black and oligogynae population number, and determining the fitness function, which is shown in Table 2.
TABLE 2 parameter settings
Figure BDA0003682259420000121
The fitness function is the RMSE between the model output value corresponding to the verification set data and the actual NOx value (RMSE ═ 1 is an allowable error, and is one of the conditions for stopping the iteration), and the optimal parameter value is the parameter value corresponding to the minimum RMSE.
1.5.5 prediction and comparison
Using the final partitioned data set in 7.5.2 and using BWOA to optimize the range of network parameters set in 7.5.4, the model was trained and predictions made for SCR system outlet NOx. The optimum parameter values obtained with BWOA are shown in Table 3 below. In order to better illustrate the prediction effect of the proposed model, the same data are used, and three other typical prediction models, namely BPNN, LSTM and BiGRU, are selected for comparison experiments, and the prediction result of the BWOA-BiGRU-LAM model is shown in FIG. 8. The comparison of the predicted results of the four models is shown in fig. 9. The comparison of the evaluation indexes of the prediction results of the four models is shown in Table 4.
TABLE 3 optimal parameter values
Figure BDA0003682259420000122
TABLE 4 comparison of predicted results
Figure BDA0003682259420000131
In FIG. 8, the NOx prediction result of the BWOA-BiGRU-LAM model is very close to the actual value, and the overall curve tracking effect is good; as can be seen in FIG. 9, the prediction result of BWOA-BiGRU-LAM is closest to the actual value, and BPNN prediction deviation is the largest, which indicates that the BWOA-BiGRU-LAM model has the highest prediction accuracy. In Table 4, the SMAPE and RMSE values of the BWOA-BiGRU-LAM model are the smallest, which indicates that the prediction precision of the model is the highest and the R of the model is the highest 2 The coefficient is maximum, which indicates that the data tracking capability of the model is best; BPNN has the largest SMAPE and RMSE, and R 2 The coefficient is minimum and the prediction accuracy is minimum. The learning capability of the shallow neural network model is insufficient, the deep features of the object cannot be obtained, and accurate prediction is difficult to make for a complex nonlinear object; BiGRThe prediction accuracy of the U and LSTM models is still better than that of BPNN but lower than that of BWOA-BiGRU-LAM.
To further illustrate the generalization capability of the proposed model, the predicted results of the four prediction models are respectively subjected to linear fitting, and the result pair is shown in fig. 10, and the absolute error pair of the predicted results of the four models is shown in fig. 11. As can be seen from fig. 10, for the BWOA-BiGRU-LAM, LSTM, BiGRU models, the scatter points formed by the predicted values and the measured values are almost distributed on both sides of the fitted straight line, and the fitted straight line is relatively close to the ideal straight line (y ═ x), and the BWOA-BiGRU-LAM model is closest to the ideal straight line. The four model fit results pairs are shown in table 5. As can be seen in FIG. 11, the absolute error of the BWOA-BiGRU-LAM prediction result is closest to 0. And the BWOA-BiGRU-LAM model has 68.5% of test sample relative error within 2%, which is far higher than 11.1% of BPNN model, 33% of LSTM and 49.4% of BiGRU, which shows that the BWOA-BiGRU-LAM prediction result error is minimum.
TABLE 5 comparison of Linear fitting results
Figure BDA0003682259420000132
While embodiments of the invention have been described above, it is not intended to be limited to the details shown herein, and to the particular embodiments shown, but it is to be understood that all changes and modifications that come within the spirit and scope of the invention are desired to be protected by the teachings herein.

Claims (6)

1. An SCR outlet NOx emission prediction method based on BWOA-BiGRU-LAM is characterized by comprising the following steps:
acquiring NOx emission sensitive parameters and corresponding NOx emission historical data samples to form a data set;
cleaning the data set, and removing abnormal values and noise to obtain a cleaned data set;
calibrating time delay between each NOx emission sensitive parameter and NOx emission amount to form a corresponding relation between each NOx emission sensitive parameter and a NOx emission amount historical data sample to construct a data set; dividing a data set into a training set and a verification set;
fourthly, constructing a BiGRU-LAM network, inputting a training set, simultaneously setting the value ranges of four network parameters of a learning rate, a minimum batch number, a hidden layer neuron number and a regularization coefficient, training the network, automatically optimizing the learning rate, the minimum batch number, the hidden layer neuron number and the regularization coefficient in the given ranges by adopting BWOA, inputting the optimal parameter values into the BiGRU-LAM network, and forming a final prediction model BWOA-BiGRU-LAM after verification of a verification set;
and step five, collecting and collecting NOx emission sensitive parameters in real time, inputting the parameters into a final prediction model BWOA-BiGRU-LAM after time delay correction between each NOx emission sensitive parameter and NOx emission, and predicting the NOx emission in real time.
2. The BWOA-BiGRU-LAM based SCR outlet NOx emission prediction method of claim 1, wherein the NOx emission sensitive parameters include SCR inlet oxygen content, SCR inlet flue gas flow rate, SCR inlet NOx concentration, SCR inlet flue gas temperature, SCR outlet ammonia concentration, SCR outlet oxygen content and SCR mixed ammonia flow rate.
3. The BWOA-BiGRU-LAM-based SCR outlet NOx emission prediction method of claim 1, wherein in said second step, the method of scrubbing the data set is as follows:
and screening abnormal values in the data set by using a boxplot, then correcting the abnormal values by using a mean filling method, removing noise in the data set by using a wavelet semi-soft threshold denoising algorithm, and then standardizing by using a Z-Score method.
4. The BWOA-BiGRU-LAM based SCR outlet NOx emission prediction method of claim 1, wherein in the third step, the method of calibrating the time delay between each NOx emission sensitive parameter and the amount of NOx emission is as follows:
assuming that X and Y are two discrete random variables, the mutual information I (X, Y) between X and Y is:
Figure FDA0003682259410000021
wherein p (X, Y) is the joint probability distribution of X and Y, and p (X), p (Y) is the edge probability distribution of X and Y; in order to calculate the mutual information value between input variable and output NOx formed by each NOx emission sensitive parameter, an input variable time sequence matrix is set as X (t), SCR outlet NOx measured value is set as an output time sequence matrix Y (t), and the method comprises the following steps:
X(t)＝[x 1 (t),x 2 (t),Λ,x n (t)] (18)
wherein n is X (t) dimension, namely the number of variables;
setting τ i (τ i ∈[τ min ,τ max ]) For the ith (i epsilon (1, n)) variable x i A time delay between (t) and Y (t),
x (t) can be reconstructed as:
Figure FDA0003682259410000022
wherein the ith variable x i (t) reconstruction as:
Figure FDA0003682259410000023
according to the actual operation condition of the SCR system, the variable delay time range is set between 0s and 300s, the sampling period is 1min, and therefore tau is selected min ＝0，τ max (ii) 5; to find each input variable x i (t) calculating the optimal delay time by the following steps:
(a) calculating the mutual information value between each row of the matrix on the right side of the medium sign in the formula (20) and the output Y (t), wherein the delay time tau corresponding to the row with the maximum mutual information value is the optimal delay time estimation of the variable;
(b) repeating step (a) for delay times for all NOx emission sensitive parameters.
5. The BWOA-BiGRU-LAM based SCR outlet NOx emission prediction method of claim 1, wherein the method of constructing the BiGRU-LAM network in step four is as follows:
constructing a BiGRU network, wherein the BiGRU network comprises two GRU networks with the same structure, and the data processing directions of the two GRU networks are opposite; and the outputs of the gate control circulation units corresponding to each other in the two GRU networks are spliced and then used as the input of the LAM network.
6. The BWOA-BiGRU-LAM-based SCR outlet NOx emission prediction method of claim 5,
the method is characterized in that the specific calculation process of the GRU network is as follows:
r t ＝sigmoid(W r ·[h t-1 ,x t ]+b r ) (1)
z t ＝sigmoid(W z ·[h t-1 ,x t ]+b z ) (2)
Figure FDA0003682259410000031
Figure FDA0003682259410000032
in the formula, x t For input at the present moment, h t-1 Is a hidden state of the neuron at the previous moment r t And z t Outputs of reset and update gates, W, respectively, at the present moment r 、W z And W h Respectively a reset gate weight matrix, an update gate weight matrix, a neuron output weight matrix, b r 、b z And b h Respectively reset gate bias termsA matrix, an updated gate bias term matrix and a neuron output bias term matrix,
Figure FDA0003682259410000033
candidate hidden states for the neuron at the current time, h t For the final state output at the current time, represent a matrix multiplication,
Figure FDA0003682259410000034
representing multiplication of corresponding elements of the matrix, wherein tanh and sigmoid are respectively a tanh activation function and a sigmoid activation function;
the specific calculation process of LAM is as follows:
Figure FDA0003682259410000035
Figure FDA0003682259410000036
Figure FDA0003682259410000037
Figure FDA0003682259410000038
wherein, a ts Attention weight of neuron output hidden state at historical time, namely s time, to neuron output at current time, namely t time, score is a scoring function,
Figure FDA0003682259410000039
is the hidden state of the historical output of the neuron, h t Is the hidden state of the neuron output at the current moment, c t Is the intermediate variable that is the variable between,
Figure FDA00036822594100000310
is the final output of the neuron at the current moment after attention weighted calculation, W c Is a weight matrix.
CN202210637691.0A 2022-06-07 2022-06-07 SCR (Selective catalytic reduction) outlet NOx emission prediction method based on BWOA-BiGRU-LAM (lean-reactive inert gas) Pending CN115018158A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210637691.0A CN115018158A (en) 2022-06-07 2022-06-07 SCR (Selective catalytic reduction) outlet NOx emission prediction method based on BWOA-BiGRU-LAM (lean-reactive inert gas)

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210637691.0A CN115018158A (en) 2022-06-07 2022-06-07 SCR (Selective catalytic reduction) outlet NOx emission prediction method based on BWOA-BiGRU-LAM (lean-reactive inert gas)

Publications (1)

Publication Number Publication Date
CN115018158A true CN115018158A (en) 2022-09-06

Family

ID=83072265

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210637691.0A Pending CN115018158A (en) 2022-06-07 2022-06-07 SCR (Selective catalytic reduction) outlet NOx emission prediction method based on BWOA-BiGRU-LAM (lean-reactive inert gas)

Country Status (1)

Country Link
CN (1) CN115018158A (en)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115587294A (en) * 2022-12-08 2023-01-10 中国空气动力研究与发展中心设备设计与测试技术研究所 Compressor fault diagnosis method based on blade tip clearance and bearing vibration
CN116609672A (en) * 2023-05-16 2023-08-18 国网江苏省电力有限公司淮安供电分公司 Energy storage battery SOC estimation method based on improved BWOA-FNN algorithm
CN117056731A (en) * 2023-09-11 2023-11-14 重庆理工大学 Ammonia nitrogen prediction method based on jet impact-negative pressure reactor flow signal
CN117313936A (en) * 2023-09-30 2023-12-29 江南大学 Clean flue gas SO in flue gas desulfurization process of coal-fired power plant 2 Concentration prediction method

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115587294A (en) * 2022-12-08 2023-01-10 中国空气动力研究与发展中心设备设计与测试技术研究所 Compressor fault diagnosis method based on blade tip clearance and bearing vibration
CN116609672A (en) * 2023-05-16 2023-08-18 国网江苏省电力有限公司淮安供电分公司 Energy storage battery SOC estimation method based on improved BWOA-FNN algorithm
CN116609672B (en) * 2023-05-16 2024-05-07 国网江苏省电力有限公司淮安供电分公司 Energy storage battery SOC estimation method based on improved BWOA-FNN algorithm
CN117056731A (en) * 2023-09-11 2023-11-14 重庆理工大学 Ammonia nitrogen prediction method based on jet impact-negative pressure reactor flow signal
CN117313936A (en) * 2023-09-30 2023-12-29 江南大学 Clean flue gas SO in flue gas desulfurization process of coal-fired power plant 2 Concentration prediction method
CN117313936B (en) * 2023-09-30 2024-03-29 江南大学 Clean flue gas SO in flue gas desulfurization process of coal-fired power plant 2 Concentration prediction method

Similar Documents

Publication Publication Date Title
CN115018158A (en) SCR (Selective catalytic reduction) outlet NOx emission prediction method based on BWOA-BiGRU-LAM (lean-reactive inert gas)
Tang et al. Auto-encoder-extreme learning machine model for boiler NOx emission concentration prediction
Hu et al. A novel wind power probabilistic forecasting approach based on joint quantile regression and multi-objective optimization
CN110598929B (en) Wind power nonparametric probability interval ultrashort term prediction method
CN112884056A (en) Optimized LSTM neural network-based sewage quality prediction method
CN111144552B (en) Multi-index grain quality prediction method and device
CN111815053B (en) Prediction method and system for industrial time sequence data
CN113031555B (en) Intelligent purification system for harmful gas in environment of livestock and poultry house
CN113301127B (en) Livestock feed detection system
CN110956309A (en) Flow activity prediction method based on CRF and LSTM
CN112257847A (en) Method for predicting geomagnetic Kp index based on CNN and LSTM
CN116960978A (en) Offshore wind power prediction method based on wind speed-power combination decomposition reconstruction
CN115640874A (en) Transformer state prediction method based on improved grey model theory
CN113128666A (en) Mo-S-LSTMs model-based time series multi-step prediction method
CN115660167A (en) Short-term air quality prediction method based on sparrow search algorithm and decomposition error correction
CN114896895A (en) Gate-controlled cyclic neural network-based denitration system inlet nitrogen oxide emission prediction method and system
CN113488111B (en) Ammonia injection amount optimization modeling method for SCR denitration system
Liu et al. A prediction method of NOx in thermal power plants using GC-LSTM neural network
Fan Entrepreneurial Mentality Analysis for College Students Based on the Improved BP Neural Network
CN117190173B (en) Optimal control method and control system for flue gas recirculation and boiler coupling system
CN117420757A (en) Denitration prediction control method, system and equipment based on multivariate fusion model
CN116933926A (en) AE-GAN-based SCR inlet NOx emission prediction method
CN109472070A (en) A kind of flexible measurement method of the ESN furnace operation variable based on PLS
Sun et al. Modeling method of boiler combustion system based on empirical mode decomposition
CN112488145B (en) NO based on intelligent methodxOnline prediction method and system

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication