CN112884213A - Coal-fired boiler NOx prediction method based on wavelet decomposition and dynamic hybrid deep learning - Google Patents
Coal-fired boiler NOx prediction method based on wavelet decomposition and dynamic hybrid deep learning Download PDFInfo
- Publication number
- CN112884213A CN112884213A CN202110143854.5A CN202110143854A CN112884213A CN 112884213 A CN112884213 A CN 112884213A CN 202110143854 A CN202110143854 A CN 202110143854A CN 112884213 A CN112884213 A CN 112884213A
- Authority
- CN
- China
- Prior art keywords
- coal
- nox emission
- nox
- model
- component
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 42
- 238000000354 decomposition reaction Methods 0.000 title claims abstract description 26
- 238000013135 deep learning Methods 0.000 title claims abstract description 20
- 230000009466 transformation Effects 0.000 claims abstract description 9
- 238000012549 training Methods 0.000 claims description 19
- 239000003245 coal Substances 0.000 claims description 9
- 230000008569 process Effects 0.000 abstract description 7
- 239000000306 component Substances 0.000 description 56
- 238000013527 convolutional neural network Methods 0.000 description 54
- 230000006870 function Effects 0.000 description 21
- 230000004913 activation Effects 0.000 description 11
- 238000013528 artificial neural network Methods 0.000 description 11
- 238000012360 testing method Methods 0.000 description 10
- 238000002485 combustion reaction Methods 0.000 description 9
- QGZKDVFQNNGYKY-UHFFFAOYSA-N Ammonia Chemical compound N QGZKDVFQNNGYKY-UHFFFAOYSA-N 0.000 description 8
- 230000000694 effects Effects 0.000 description 6
- 238000005457 optimization Methods 0.000 description 6
- 230000003068 static effect Effects 0.000 description 6
- 229910021529 ammonia Inorganic materials 0.000 description 4
- 238000002347 injection Methods 0.000 description 4
- 239000007924 injection Substances 0.000 description 4
- 238000005070 sampling Methods 0.000 description 4
- 238000001914 filtration Methods 0.000 description 3
- 230000009467 reduction Effects 0.000 description 3
- 238000006722 reduction reaction Methods 0.000 description 3
- 238000013519 translation Methods 0.000 description 3
- 238000004364 calculation method Methods 0.000 description 2
- 238000010531 catalytic reduction reaction Methods 0.000 description 2
- 230000000052 comparative effect Effects 0.000 description 2
- 150000001875 compounds Chemical class 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000011478 gradient descent method Methods 0.000 description 2
- 230000007774 longterm Effects 0.000 description 2
- 239000011159 matrix material Substances 0.000 description 2
- 238000005259 measurement Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000003062 neural network model Methods 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 239000000243 solution Substances 0.000 description 2
- 238000012795 verification Methods 0.000 description 2
- 238000012935 Averaging Methods 0.000 description 1
- 238000012356 Product development Methods 0.000 description 1
- 108010076504 Protein Sorting Signals Proteins 0.000 description 1
- 230000003213 activating effect Effects 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 239000008358 core component Substances 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000004134 energy conservation Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 239000003344 environmental pollutant Substances 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 239000000446 fuel Substances 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 231100000719 pollutant Toxicity 0.000 description 1
- 238000011176 pooling Methods 0.000 description 1
- 238000010248 power generation Methods 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
- 230000001902 propagating effect Effects 0.000 description 1
- 230000004044 response Effects 0.000 description 1
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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Business, Economics & Management (AREA)
- Health & Medical Sciences (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Economics (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- Data Mining & Analysis (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Computational Linguistics (AREA)
- Human Resources & Organizations (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Strategic Management (AREA)
- General Business, Economics & Management (AREA)
- Marketing (AREA)
- Tourism & Hospitality (AREA)
- Primary Health Care (AREA)
- Water Supply & Treatment (AREA)
- Public Health (AREA)
- Development Economics (AREA)
- Game Theory and Decision Science (AREA)
- Entrepreneurship & Innovation (AREA)
- Operations Research (AREA)
- Quality & Reliability (AREA)
- Testing And Monitoring For Control Systems (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention provides a coal-fired boiler NOx prediction method based on wavelet decomposition and dynamic hybrid deep learning, relates to the technical field of coal-fired boiler operation, can accurately and stably predict NOx emission, and has good generalization performance and repeatability; the method comprises the following steps: s1, collecting the value of the NOx emission sensitive parameter in a certain time period; s2, carrying out wavelet transformation on the collected value of the NOx emission sensitive parameter to obtain a trend component and a high-frequency fluctuation component of the collected value of the NOx emission sensitive parameter; s3, performing wavelet reconstruction on the trend component and the high-frequency fluctuation component; s4, dynamically predicting the reconstructed trend component by adopting an LSTM model, and dynamically predicting the reconstructed high-frequency fluctuation component by adopting a CNN model; and S5, fusing the prediction results in the S4 to obtain a final NOx emission prediction result. The technical scheme provided by the invention is suitable for the process of predicting the NOx emission of the coal-fired boiler.
Description
Technical Field
The invention relates to the technical field of coal-fired boiler operation, in particular to a coal-fired boiler NOx prediction method based on wavelet decomposition and dynamic hybrid deep learning.
Background
In recent years, with the rapid development of new energy power generation scale and the competition of long-term and continuous increase of the total installed capacity of electric power in China, the average annual utilization hours of thermal power generating units in China continuously decrease. In order to meet the requirements of reduction of the number of hours of utilization and frequent peak regulation, most thermal power generating units are in a variable-load or even low-load operation state for a long time. During low-load operation, various energy efficiency indexes of the unit are reduced, the NOx emission concentration is increased due to the increase of the air amount corresponding to unit fuel, and the energy conservation and emission reduction still face challenges.
The prediction of the NOx emission of a power station boiler is the basis of operations such as combustion optimization and ammonia injection of a Selective Catalytic Reduction (SCR) system, and is also an important technology for the boiler to achieve both economy and low emission, and has been paid attention to by researchers in recent years.
At present, most combustion optimization modeling methods and molded product development are steady-state models established based on steady-state data of a boiler combustion system, and combustion optimization model prediction under the condition of dynamic variable load is difficult to meet. In order to achieve the training and prediction of the current power plant frequent load fluctuation pollutant emission model, many researchers are beginning to try a dynamic prediction model. Dynamic models always have higher prediction accuracy than static models because dynamic prediction models take into account the time-varying nature of plant operation, adding useful information provided by the changing characteristics in the process history data. The dynamic prediction model can well solve the variability of the internal information of the frequent load of the power plant and has better advantages in the aspect of model prediction.
Although the above models add dynamic information to the data and all achieve better prediction effect than the static model, most models do not consider the existence of suspected noise of the original data in frequent fluctuation and complicated characteristic information which may be caused by the influence of various factors in the acquisition process, and even directly remove the so-called noise. The power station combustion data with frequent peak shaving usually has complex characteristics of non-stationarity, nonlinearity, high fluctuation and the like, and the characteristics are usually the most important components in the original data and should be considered. And considering the defects that machine learning algorithms such as a neural network are easy to generate overfitting, have high convergence rate and local minimum and the like, the accuracy of the prediction model is limited.
Accordingly, there is a need to develop a method for predicting NOx in a coal-fired boiler by wavelet decomposition and dynamic hybrid deep learning to address the deficiencies of the prior art and to solve or mitigate one or more of the problems set forth above.
Disclosure of Invention
In order to meet the peak regulation requirement of a power grid, most thermal power generating units are usually operated under variable load or even low load, so that the NOx emission is increased. Ammonia injection by Selective Catalytic Reduction (SCR) denitration systems is a major approach to reduce NOx concentrations in utility boilers. In order to improve the control quality and economy of the SCR denitration system, it is necessary to establish an accurate dynamic model of the SCR denitration reactor inlet NOx generation. In view of the above, the invention provides a coal-fired boiler NOx prediction method based on wavelet decomposition and dynamic hybrid deep learning, which has accurate and stable modeling and prediction effects, and compared with other typical prediction methods, the model has better generalization performance and higher repeatability and stability, and provides better choices for further realizing accurate ammonia injection and combustion optimization.
In one aspect, the invention provides a method for predicting NOx of a coal-fired boiler by wavelet decomposition and dynamic hybrid deep learning, which is characterized by comprising the following steps:
s1, collecting the value of the NOx emission sensitive parameter in a certain time period (the NOx emission sensitive parameter is the relevant measurement parameter for predicting NOx emission, such as load, air door baffle opening, primary air pressure, secondary air pressure, air quantity, coal quantity, air distribution mode, coal mill operation combination mode and the like);
s2, performing wavelet transformation on all collected values of each NOx emission sensitive parameter to obtain a trend component and a high-frequency fluctuation component of the NOx emission sensitive parameter in the time period;
s3, performing wavelet reconstruction on the trend component and the high-frequency fluctuation component;
s4, dynamically predicting the reconstructed trend component by adopting an LSTM model, and dynamically predicting the reconstructed high-frequency fluctuation component by adopting a CNN model;
and S5, fusing the prediction results obtained by the LSTM model and the CNN model to obtain the final NOx emission prediction result.
The above-described aspects and any possible implementations further provide an implementation in which each NOx emission sensitive parameter is decomposed into a trend component and a plurality of high frequency fluctuation components.
The above-mentioned aspect and any possible implementation manner further provide an implementation manner, and the content fused in step S5 includes: and compensating the prediction result of the trend component by using the prediction result of the high-frequency fluctuation component.
As for the above-mentioned aspect and any possible implementation manner, a specific content of the wavelet reconstruction of the trend component and the high-frequency fluctuation component in step S3 is sequence reconstruction.
The above-described aspect and any possible implementation further provide an implementation, wherein the basis for the sequence reconstruction of the high-frequency fluctuation component is a fluctuation frequency.
In the above aspect and any possible implementation manner, there is further provided an implementation manner that the length of the time period for acquiring data in step S1 is one week, and the acquisition frequency is 1 time/minute.
In another aspect, the present invention provides a coal-fired boiler NOx modeling method by wavelet decomposition and dynamic hybrid deep learning, wherein the method comprises the steps of:
s1, collecting historical data of the NOx emission sensitive parameters and corresponding NOx emission amount;
s2, performing wavelet transformation on the historical data of the NOx emission sensitive parameters and the corresponding NOx emission to obtain trend components and high-frequency fluctuation components of the NOx emission sensitive parameters and the NOx emission;
s3, performing wavelet reconstruction on the trend component and the high-frequency fluctuation component;
s4, training the LSTM model by using the reconstructed trend component, and training the CNN model by using the reconstructed high-frequency fluctuation component to obtain a trained LSTM model and a trained CNN model;
the steps S1-S4 are repeated to obtain the final model.
The above aspects and any possible implementation manners further provide an implementation manner, and the NOx emission sensitive parameters include load, opening degree of a damper, primary air pressure, secondary air pressure, air volume, coal volume, air distribution manner, and operation combination manner of a coal mill.
The above-described aspects and any possible implementation further provide an implementation in which the setting parameters of the original LSTM model and the CNN model are obtained from empirical data.
The above-described aspects and any possible implementation further provide an implementation, and the parameter setting of the trained LSTM model includes: the number of hidden layer nodes is 128, maxEpochs is 128, miniBatchSize is 8, InitialalLearnRate is 0.006, and Dropout is 0.7;
the structure of the final CNN model includes: 1 input layer, 2 convolutional layers (C1, C3), 2 downsampled layers (S2, S4),2 fully-connected layers (C5, C6), 1 output layer; the size of an input layer is 9 × 9, 4 feature surfaces with the size of 8 × 8 are obtained by using 4 convolution kernels with the sliding step length of 2 × 2 in the C1 layer, 4 feature surfaces with the size of 4 × 4 in the downsampling layer S2 are obtained after downsampling, 8 feature surfaces with the size of 2 × 2 are obtained by using convolution kernels with the sliding step length of 3 × 3 in the C3 layer, 8 feature surfaces with the size of 1 × 1 in the S4 are obtained after downsampling, finally, the two fully-connected layers are respectively C5 layers and C6 layers, the two fully-connected layers are fully connected with the previous layer, the two fully-connected layers integrate various local features extracted in the early stage, and finally, a predicted value is obtained through the output layer.
Compared with the prior art, the invention can obtain the following technical effects: the model based on wavelet decomposition and dynamic hybrid deep learning, namely the WT-CNN-LSTM model, has the capability of adapting to the change of working conditions, has higher prediction precision and stability compared with other prediction methods, and can be further used as the basis for the combustion optimization of a coal-fired boiler and the accurate ammonia injection optimization of an SCR denitration system; meanwhile, the model has enough high-quality operation data and necessary data preprocessing, and has good application potential on similar coal-fired power plant boilers.
Of course, it is not necessary for any one product in which the invention is practiced to achieve all of the above-described technical effects simultaneously.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings needed to be used in the embodiments will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings based on these drawings without creative efforts.
FIG. 1 is a schematic diagram of the basic structure of an LSTM neural network according to an embodiment of the present invention;
FIG. 2 is a flow chart of a dynamic prediction method for coal fired boiler NOx emission prediction incorporating wavelet transformation, LSTM and CNN provided by one embodiment of the present invention;
fig. 3 is a schematic structural diagram of a CNN module in a prediction model according to an embodiment of the present invention;
FIG. 4 is an LSTM-CNN provided by an embodiment of the present inventionAllAnd LSTM-CNNLowTraining a data set comparison graph;
FIG. 5 is an LSTM-CNN provided by an embodiment of the present inventionAllAnd LSTM-CNNLowTesting a data set comparison graph;
FIG. 6 is an LSTM-CNN provided by an embodiment of the present inventionAllAnd LSTM-CNNLowRMSE comparison of (a);
FIG. 7 is an LSTM-CNN provided by an embodiment of the present inventionAllAnd LSTM-CNNLowRMSE standard deviation S comparison of (a);
FIG. 8 is a comparison of the predictive performance of training data sets of models provided by one embodiment of the present invention;
FIG. 9 is a comparison of the predictive performance of test data sets for each model provided by one embodiment of the present invention;
FIG. 10 is a graph comparing RMSE for each model provided by one embodiment of the present invention;
FIG. 11 is a graph comparing the RMSE standard deviations for each model provided by one embodiment of the present invention.
Detailed Description
For better understanding of the technical solutions of the present invention, the following detailed descriptions of the embodiments of the present invention are provided with reference to the accompanying drawings.
It should be understood that the described embodiments are only some embodiments of the invention, and not all embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
The terminology used in the embodiments of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. As used in the examples of the present invention and the appended claims, the singular forms "a," "an," and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise.
In order to capture the complex information contained in the utility boiler combustion data, the present invention utilizes wavelet decomposition transform (WT) signal processing techniques to decompose the raw combustion data samples into a smooth approximation component and a series of detail components. The decomposed sub-series components can be predicted by using a proper time series deep neural network model according to the characteristics of the sub-series components, namely, a mixed deep learning algorithm is used for predicting NOx emission before the inlet of the SCR denitration system. Different modes in the decomposed data are captured by using the characteristics of the hybrid deep neural network model, so that the accuracy of the final prediction model can be effectively improved.
The invention provides a dynamic prediction method for mixed wavelet transformation, LSTM and CNN coal-fired boiler NOx emission prediction. First, the original operation data is decomposed into high frequency components and low frequency components using wavelet transform. The high-frequency component reflects the fluctuation of data, namely the high-frequency fluctuation component; the low-frequency component reflects the trend, namely the trend component; and performing wavelet reconstruction on the high-frequency component and the low-frequency component to obtain a reconstructed high-frequency component and reconstructed low-frequency component, and applying the reconstructed high-frequency component and reconstructed low-frequency component to subsequent training or prediction. Secondly, the low frequency component is dynamically modeled by a deep learning long-short term memory neural network (LSTM), and the high frequency component is dynamically modeled by a Convolutional Neural Network (CNN). And then fusing the two prediction models to obtain a final NOx emission model.
The model may be used for NOx emission prediction after modeling. The time and frequency patterns in historical operating data can be captured simultaneously by first decomposing the raw data into an approximation component and multiple detail components based on Wavelet Transform (WT) algorithms. And then dynamically capturing the trend characteristics of the approximate components by using an LSTM model, and dynamically capturing the influence factor characteristics in the detail components by using a CNN network to finally obtain a corresponding prediction result. And finally, fusing the prediction results obtained by each component to obtain a final NOx emission prediction result.
1. Wavelet transform
Wavelet Transform (WT) is similar to the fourier transform principle, which is the decomposition of the original data by a wavelet function (derived from the mother wave by translation and stretching). The wavelet transform has better signal filtering capability, and compared with an original signal sequence, the decomposed sub-signal sequence can better reflect the variation characteristics of variables and can more accurately predict the concentration of NOx. Wavelet transform can be divided into 2 types: continuous Wavelet Transform (CWT) and Discrete Wavelet Transform (DWT).
The expression for a continuous wavelet transform for an arbitrary function y (t) is:
in the formula, the main wavelet is a wavelet mother wave, the scale parameter is a wavelet component, and the translation parameter is a wavelet component.
The values of CWT are called wavelet coefficients
The expression for the discrete wavelet transform is:
m is a scale parameter (decomposition level); n is a translation constant and is an integer.
The present invention uses the fast discrete wavelet transform algorithm developed by Mallat, which replaces the parent and parent wavelets with low-pass and high-pass filters. The low-pass filter is called a scale function and is used for analyzing low-frequency components, the high-pass filter is called a wavelet function and is used for analyzing high-frequency components, and the original signal is decomposed into a plurality of groups of time sequences through the filter, wherein one group of the time sequences is a smooth time sequence reflecting trend characteristics, and the other groups of the time sequences are time sequences reflecting details of disturbance signals. The selection of the parent wave has an important influence on the result.
2. LSTM neural network
Long-Short Term Memory artificial Neural Network (LSTM) is an improved time-cycling Neural Network (RNN). Because the LSTM neural network comprises a time memory unit, the LSTM neural network can learn the long-term and short-term dependence information of the time sequence, and is suitable for processing and predicting interval and delay events in the time sequence. The internal state of the LSTM mainly comprises three control gate switches formed by four related activation functions with different functions (three of the activation functions are sigmoid functions, and one is tanh function), namely a forgetting gate, an input gate and an output gate, and the input and the output of information are controlled by the gate limit structure, so that the forward and backward propagation of model training errors is realized, and finally the purpose of model convergence is achieved, and the basic structure of the LSTM model is shown in figure 1.
As shown in FIG. 1, Ct–1And CtAre the old and new states of the cell, which run directly on the entire chain just like a conveyor belt. In LSTM, the state of the cell is critical. Cell state information is added and deleted through a gate structure, the gate structure is composed of a sigmoid neural network layer and a pair-wise multiplication operation (sigmoid is a nonlinear activation function), and whether the information passes or not can be selectively determined. For a given sample sequence X ═ X1,x2,…,xm) The forward calculation method can be expressed as:
ft＝sigmoid(wxfxt+whfht-1+bf) (3)
it＝sigmoid(wxixt+whiht-1+bi) (4)
gt＝tanh(wxgxt+whght-1+bg) (5)
ot＝sigmoid(wxo+whoht-1+bo) (6)
wherein i, f, c and o are respectively an input gate, a forgetting gate, a cell state and an output gate; w and b are respectively corresponding weight coefficient matrix and bias item; σ and tanh are sigmoid and hyperbolic tangent activation functions, respectively.
3. CNN network
The CNN network is a convolutional neural network, and the basic structure of the CNN network is composed of an input layer, a convolutional layer (convolutional layer), a pooling layer (also called a downsampling layer), a full connection layer, and an output layer. The essence of the CNN model is to extract features of the input data by building multiple filters. As the network grows deeper, higher level features are extracted, eventually resulting in robust features with shift invariance from the original data.
3.1 convolutional layers
Convolutional layers are the core components of convolutional neural networks. A typical convolutional neural network consists of a plurality of convolutional layers, each of which may have a plurality of different convolutional kernels. Each convolution kernel can be regarded as a filter, each filter corresponds to a new feature map mapped after filtering, and each data in the same new feature map comes from the same identical filter, namely weight sharing of the convolution kernels. In the convolutional layer, the Feature map (Feature map) of the previous layer is convolved by a learnable convolution kernel, and then an Activation function (Activation function) is used to obtain the output Feature map. The feature output corresponding to each input feature map in the convolutional layer can be expressed as:
in the formula (I), the compound is shown in the specification,
is the output of the jth channel of convolutional layer l,
net activation of the jth channel, called convolutional layer l, by outputting a profile for the previous layer
And performing convolution summation and offset.
Called activation function, the present invention uses sigmoid function as activation function. MjRepresentation calculation
Is used to generate a set of input feature maps,
is a matrix of convolution kernels, and is,
is the bias to the convolved feature map. For an output profile
Each input feature map
Corresponding convolution kernel
Possibly differently, the symbol ". x" denotes convolution.
3.2 downsampling layer
A downsampling layer, which generally follows the convolutional layer, performs dimensionality reduction and feature extraction on each input feature map by the following formula:
in the formula (I), the compound is shown in the specification,
net activation of the j-th channel, called down-sampled layer l, from the previous layer output profile
The weight coefficient of the down sampling layer is beta,
is a bias term for the downsampled layer. The symbol down (.) represents a down-sampling function by fitting the input feature map
The output feature map is reduced by a factor of n in both dimensions by dividing into a number of non-overlapping n x n region blocks by a sliding window method and then summing, averaging or maximizing the values within each region block.
3.3 full connection layer
In the fully-connected network, all two-dimensional feature maps are spliced into one-dimensional features to be used as the input of the fully-connected network. The output of the fully-connected layer/can be obtained by weighted summation of the inputs and by activating the response of the function:
xl＝f(ul) (13)
ul＝wlxl-1+bl (14)
wherein u islCalled the net activation of the fully-connected layer l, which outputs the profile x from the previous layerl-1And weighting and biasing. w is alIs the weight coefficient of the fully connected network, blIs the bias term for the fully connected layer l.
Convolutional neural network training is a back propagation process, back propagating through an error function. The back propagation algorithm is a common supervised learning method for neural networks, and aims to estimate network parameters according to training samples and expected outputs. For the convolutional neural network, a convolutional kernel parameter k, a downsampling layer network weight β, a full-connection layer network weight w, a bias parameter b of each layer and the like are mainly optimized.
The essence of the back propagation algorithm is to allow us to compute valid errors for each network layer and derive therefrom a learning rule for the network parameters, so that the actual network output is closer to the target. The back propagation algorithm is mainly based on a gradient descent method, network parameters are initialized to random values, and then the network parameters are adjusted in the direction of reducing the training error through a gradient descent method until the network converges or the maximum iteration number is reached.
4. Prediction method structure
The invention adopts db1 as the mother wave, which can balance the wavelength and the smoothness well, and the decomposition result can reflect the variation characteristics of the input variable well. The wavelet is used to perform a 3-layer decomposition of the input variables in order to more fully and more meaningful describe the NOx mass concentration model. The original data sample sequence is decomposed and reconstructed into an approximate sequence (general trend component) and three detail sequences (high-frequency components). The approximate component is the low frequency component of the original sample sequence, which follows the trend of the signal and is predicted using the dynamic LSTM model. The CNN model can capture the high-volatility characteristics of data and is suitable for predicting detailed sequences. This process flow is shown in figure 2.
The steps of the prediction method comprise the following steps:
step 1: decomposing all original operation data into an approximate series c3(t) and three detail series d1(t), d2(t) and d3(t) through wavelet transformation;
step 2: in order to reproduce the original data sequence, it is important to reconstruct the approximation sequence and the detail sequence. The wavelet reconstructed corresponding values of C3t, D1t, D2t and D3t are denoted as C3t, D1t, D2t and D3t, and the reconstruction process is the inverse operation of the corresponding decomposition process. The sequence reconstructed by the wavelet has less loss compared with the original data:
Xt＝C3t+D1t+D2t+D3t (15)
the meaning of the decomposition is that different filtering signals are decomposed into an approximate component and a series of detail components, and the effect of removing noise is achieved at the same time; and then, the original signal is constructed inversely by a method of reconstructing the coefficient, and finally the purpose of denoising is achieved on the basis of not deviating from the original signal, so that the precision is higher.
And step 3: in the aspect of establishing NOx emission prediction, the LSTM model is used for dynamically predicting C3t, and the detail series (including D1t, D2t and D3t) are used for dynamically predicting by using the CNN model.
And 4, step 4: the final NOx prediction is obtained by compensating and merging the prediction results of the series D1t, D2t and D3t with the result of the series A3t, and can be expressed as:
example 1:
in order to verify the validity of the model, the invention collects the operation data of about 7 days from a Distributed Control System (DCS) of the researched power plant. The load fluctuation of the selected data-covered unit is frequent, the sampling interval is 1 minute, the number of sampling points is 11700, and a dynamic mixed LSTM and CNN neural network NOx emission model based on wavelet transformation is trained and tested. The present invention utilizes [ X (t), X (t-1), X (t-2), X (t-3), y (t-1), y (t)]The extended architecture of (3) establishes a dynamic model of the SCR system inlet NOx. To verify the accuracy of the LSTM-CNN model, the separate low frequency-built model LSTM was used alonelowAnd built without processed raw dataThe LSTM and BPNN dynamic models were compared to the static KPLS model. The validity of the model was verified using the above-mentioned field actual data, where the first 70% of the data was set as the training data set and the remaining 30% was set as the test data set. In order to evaluate the prediction capability of the method, two error measures are taken as evaluation indexes and are compared with different models. Root Mean Square Error (RMSE) is defined as follows:
where n is the number of samples, yiIn the form of an actual value of the value,
is a predicted value.
The stability index of the model is an important basis for testing the reliability of the fitting model so as to avoid the good and bad performance of the model. The prediction model can be measured by estimating the standard deviation of the error term, and the smaller the standard deviation of the model error term is, the more accurate, the more stable and the more reliable the prediction model is. The standard deviation of the root mean square error is expressed as:
model establishment and result analysis comparison:
after debugging, the important hyper-parameters set by the LSTM model of the low-frequency part of the verification model are as follows: the number of hidden layer nodes is 128, maxEpochs is 128, miniBatchSize is 8, initialLearnRate is 0.006, and Dropout is 0.7. High frequency three parts according to the above selected variable forms, the CNN model structure of the verification model of the present invention is shown in fig. 3, i.e. 1 input layer, 2 convolutional layers (C1, C3), 2 downsampled layers (S2, S4),2 fully-connected layers (C5, C6), 1 output layer; the size of an input layer is 9 × 9, 4 feature surfaces with the size of 8 × 8 are obtained by using 4 convolution kernels with the sliding step length of 2 × 2 in the C1 layer, 4 feature surfaces with the size of 4 × 4 in the downsampling layer S2 are obtained after downsampling, 8 feature surfaces with the size of 2 × 2 are obtained by using convolution kernels with the sliding step length of 3 × 3 in the C3 layer, 8 feature surfaces with the size of 1 × 1 in the S4 are obtained after downsampling, finally, the two fully-connected layers are respectively C5 layers and C6 layers, the two fully-connected layers are fully connected with the previous layer, the two fully-connected layers integrate various local features extracted in the early stage, and finally, a predicted value is obtained through the output layer.
Integral part LSTM-CNN of built power station boiler SCR system inlet NOx emission prediction model LSTM-CNN modelAllAnd the low frequency part LSTM-CNNLowAs shown in fig. 4-5, the blue line in the graph is the actual measurement. As can be seen from the figure, LSTM-CNNAllAnd LSTM-CNNLowAll have better accuracy, but the LSTM-CNN compensated by the CNN model of the high-frequency part is addedAllCloser to the actual value, its accuracy is higher. To illustrate the stability of the proposed method, 15 replicates were performed under the given parameters. FIGS. 6-7 further illustrate LSTM-CNNAllAnd LSTM-CNNLowSpecific comparative values for RMSE and its standard deviation S for 15 trials. LSTM-CNNAllThe mean RMSE for the model training data was 3.48mg/Nm3The mean RMSE for the test data set was only 6.823.48mg/Nm3And LSTM-CNNLowThe RMSE for the training and testing of the model was 5.09mg/Nm3、8.61mg/Nm3。LSTM-CNNAllModel ratio LSTM-CNNLowThe training precision of the model is improved by 31.7 percent on average, and the testing precision is improved by 20.8 percent on average. Also, from FIG. 4, it can be seen that LSTM-CNNAllModel and LSTM-CNNLowThe standard deviation of the model RMSE is small, which shows that the stability of the model RMSE and the standard deviation of the model RMSE are good. Good prediction performance on test data shows that the LSTM-CNN NOx emission model based on wavelet decomposition has strong universality.
FIGS. 8 and 9 are graphs comparing the WT-LSTM-CNN model with the LSTM, BPNN model alone and the static KPLS model. It is noted that the LSTM alone without wavelet transform uses the same parameters as the LSTM for the low frequency portion of the LSTM-CNN model. The BPNN model also uses the same dynamic structure, and the main parameters are set to be 128 hidden layer nodes, 0.01 learning rate, 1000 maximum training steps, and 0.01 training required precision. The static KPLS model adopts a Gaussian kernel function. For the comparison graph of the training data sets of the models in fig. 8, each model can better fit the emission trend of NOx, but for the test data set in fig. 9, it can be obviously seen that the deviation of the KPLS model of the static model is large, and the generalization capability is not good, which may be that the KPLS algorithm is limited to small samples and steady-state data samples, and once the fluctuation of the data samples is large, the accuracy is obviously reduced. The same 15 trials were carried out and the specific comparative values for each model are shown in FIGS. 10-11, and the standard deviation of RMSE is not given due to the larger test mean RMSE for KPLS. The standard deviation of the average RMSE of the BPNN model is larger, which shows that the network has good and bad performance and poor stability. Both LSTM-CNN and LSTM show better prediction performance, but the accuracy and stability of the LSTM-CNN model are better.
The method for predicting the NOx of the coal-fired boiler by wavelet decomposition and dynamic hybrid deep learning provided by the embodiment of the application is described in detail above. The above description of the embodiments is only for the purpose of helping to understand the method of the present application and its core ideas; meanwhile, for a person skilled in the art, according to the idea of the present application, there may be variations in the specific embodiments and the application scope, and in summary, the content of the present specification should not be construed as a limitation to the present application.
As used in the specification and claims, certain terms are used to refer to particular components. As one skilled in the art will appreciate, manufacturers may refer to a component by different names. This specification and claims do not intend to distinguish between components that differ in name but not function. In the following description and in the claims, the terms "include" and "comprise" are used in an open-ended fashion, and thus should be interpreted to mean "include, but not limited to. "substantially" means within an acceptable error range, and a person skilled in the art can solve the technical problem within a certain error range to substantially achieve the technical effect. The description which follows is a preferred embodiment of the present application, but is made for the purpose of illustrating the general principles of the application and not for the purpose of limiting the scope of the application. The protection scope of the present application shall be subject to the definitions of the appended claims.
It is also noted that the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a good or system that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such good or system. Without further limitation, an element defined by the phrase "comprising an … …" does not exclude the presence of other like elements in a commodity or system that includes the element.
It should be understood that the term "and/or" as used herein is merely one type of association that describes an associated object, meaning that three relationships may exist, e.g., a and/or B may mean: a exists alone, A and B exist simultaneously, and B exists alone. In addition, the character "/" herein generally indicates that the former and latter related objects are in an "or" relationship.
The foregoing description shows and describes several preferred embodiments of the present application, but as aforementioned, it is to be understood that the application is not limited to the forms disclosed herein, but is not to be construed as excluding other embodiments and is capable of use in various other combinations, modifications, and environments and is capable of changes within the scope of the application as described herein, commensurate with the above teachings, or the skill or knowledge of the relevant art. And that modifications and variations may be effected by those skilled in the art without departing from the spirit and scope of the application, which is to be protected by the claims appended hereto.
Claims (10)
1. The method for predicting the NOx of the coal-fired boiler by wavelet decomposition and dynamic mixed deep learning is characterized by comprising the following steps of:
s1, collecting the value of the NOx emission sensitive parameter in a certain time period;
s2, carrying out wavelet transformation on the collected value of the NOx emission sensitive parameter to obtain a trend component and a high-frequency fluctuation component of the collected value of the NOx emission sensitive parameter;
s3, performing wavelet reconstruction on the trend component and the high-frequency fluctuation component;
s4, dynamically predicting the reconstructed trend component by adopting an LSTM model, and dynamically predicting the reconstructed high-frequency fluctuation component by adopting a CNN model;
and S5, fusing the prediction results obtained by the LSTM model and the CNN model to obtain the final NOx emission prediction result.
2. The method for predicting NOx in a coal-fired boiler through wavelet decomposition and dynamic hybrid deep learning according to claim 1, wherein each NOx emission sensitive parameter is decomposed into a trend component and a plurality of high-frequency fluctuation components.
3. The method for predicting NOx in a coal-fired boiler by wavelet decomposition and dynamic hybrid deep learning according to claim 1, wherein the contents fused in step S5 include: and compensating the prediction result of the trend component by using the prediction result of the high-frequency fluctuation component.
4. The method for predicting NOx in a coal-fired boiler through wavelet decomposition and dynamic deep learning according to claim 1, wherein the specific content of wavelet reconstruction of the trend component and the high-frequency fluctuation component in the step S3 is sequence reconstruction.
5. The method for predicting NOx of a coal-fired boiler through wavelet decomposition and dynamic hybrid deep learning according to claim 4, wherein the basis for sequence reconstruction of high-frequency fluctuation components is fluctuation frequency.
6. The method for predicting NOx of a coal-fired boiler through wavelet decomposition and dynamic hybrid deep learning according to claim 1, wherein the period of time for collecting data in the step S1 is one week, and the collection frequency is 1 time/minute.
7. The method for predicting NOx in a coal-fired boiler through wavelet decomposition and dynamic hybrid deep learning according to claim 1, wherein the NOx emission sensitive parameters comprise load, opening degree of a damper, primary air pressure, secondary air pressure, air volume, coal volume, air distribution mode and coal mill operation combination mode.
8. The coal-fired boiler NOx modeling method based on wavelet decomposition and dynamic hybrid deep learning is characterized by comprising the following steps of:
s1, collecting historical data of the NOx emission sensitive parameters and corresponding NOx emission amount;
s2, performing wavelet transformation on the historical data of the NOx emission sensitive parameters and the corresponding NOx emission to obtain trend components and high-frequency fluctuation components of the NOx emission sensitive parameters and the NOx emission;
s3, performing wavelet reconstruction on the trend component and the high-frequency fluctuation component;
s4, training the LSTM model by using the reconstructed trend component, and training the CNN model by using the reconstructed high-frequency fluctuation component to obtain a trained LSTM model and a trained CNN model;
the steps S1-S4 are repeated to obtain the final model.
9. The method of claim 8, wherein the NOx emission sensitive parameters include load, damper opening, primary air pressure, secondary air pressure, air volume, coal volume, air distribution mode and coal mill operation combination mode.
10. The method of claim 8, wherein the parameter settings of the trained LSTM model include: the number of hidden layer nodes is 128, maxEpochs is 128, miniBatchSize is 8, InitialalLearnRate is 0.006, and Dropout is 0.7;
the structure of the final CNN model includes: 1 input layer, 2 convolutional layers, 2 downsampling layers, 2 full-link layers, 1 output layer; the size of an input layer is 9 × 9, 4 feature surfaces with the size of 8 × 8 are obtained by using 4 convolution kernels with the sliding step length of 2 × 2 by the convolution layer C1, 4 feature surfaces with the size of 4 × 4 are obtained by the downsampling, 8 feature surfaces with the size of 2 × 2 are obtained by the downsampling layer S2 by the convolution kernel with the sliding step length of 3 × 3 by the convolution layer C3, 8 feature surfaces with the size of 1 × 1 are obtained by the downsampling layer S4, finally, full connection is carried out through a full connection layer C5 with the size of 32 layers and a full connection layer C6 with the size of 4 layers, and finally, a predicted value is obtained through an output layer.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110143854.5A CN112884213B (en) | 2021-02-02 | 2021-02-02 | Coal-fired boiler NOx prediction method based on wavelet decomposition and dynamic mixed deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110143854.5A CN112884213B (en) | 2021-02-02 | 2021-02-02 | Coal-fired boiler NOx prediction method based on wavelet decomposition and dynamic mixed deep learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112884213A true CN112884213A (en) | 2021-06-01 |
| CN112884213B CN112884213B (en) | 2024-04-19 |
Family
ID=76055767
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110143854.5A Active CN112884213B (en) | 2021-02-02 | 2021-02-02 | Coal-fired boiler NOx prediction method based on wavelet decomposition and dynamic mixed deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112884213B (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114266200A (en) * | 2022-02-24 | 2022-04-01 | 山东大学 | Nitrogen dioxide concentration prediction method and system |
| 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 |
| CN117369282A (en) * | 2023-11-17 | 2024-01-09 | 上海四方无锡锅炉工程有限公司 | Control method for adaptive hierarchical air supply and solid waste CFB boiler thereof |
| CN117977579A (en) * | 2024-03-28 | 2024-05-03 | 西华大学 | DWT-LSTM neural network-based adjustable load prediction method |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109165789A (en) * | 2018-09-04 | 2019-01-08 | 广东电网有限责任公司 | The modeling method and device of emission of NOx of boiler amount prediction model based on LSTM |
| US20190324441A1 (en) * | 2017-08-02 | 2019-10-24 | Strong Force Iot Portfolio 2016, Llc | Data collection systems with pattern analysis for an industrial environment |
| CN110807577A (en) * | 2019-10-15 | 2020-02-18 | 中国石油天然气集团有限公司 | Pollution emission prediction method and device |
| WO2020181685A1 (en) * | 2019-03-12 | 2020-09-17 | 南京邮电大学 | Vehicle-mounted video target detection method based on deep learning |
-
2021
- 2021-02-02 CN CN202110143854.5A patent/CN112884213B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20190324441A1 (en) * | 2017-08-02 | 2019-10-24 | Strong Force Iot Portfolio 2016, Llc | Data collection systems with pattern analysis for an industrial environment |
| CN109165789A (en) * | 2018-09-04 | 2019-01-08 | 广东电网有限责任公司 | The modeling method and device of emission of NOx of boiler amount prediction model based on LSTM |
| WO2020181685A1 (en) * | 2019-03-12 | 2020-09-17 | 南京邮电大学 | Vehicle-mounted video target detection method based on deep learning |
| CN110807577A (en) * | 2019-10-15 | 2020-02-18 | 中国石油天然气集团有限公司 | Pollution emission prediction method and device |
Non-Patent Citations (1)
| Title |
|---|
| 邹思琳等: "时间精度与空间信息对神经网络模型 预报 PM2.5 浓度的影响", 《北京大学学报(自然科学版)》, vol. 56, no. 3, 22 January 2020 (2020-01-22), pages 417 - 426 * |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114266200A (en) * | 2022-02-24 | 2022-04-01 | 山东大学 | Nitrogen dioxide concentration prediction method and system |
| CN114266200B (en) * | 2022-02-24 | 2022-07-05 | 山东大学 | Nitrogen dioxide concentration prediction method and system |
| 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 |
| CN117369282A (en) * | 2023-11-17 | 2024-01-09 | 上海四方无锡锅炉工程有限公司 | Control method for adaptive hierarchical air supply and solid waste CFB boiler thereof |
| CN117369282B (en) * | 2023-11-17 | 2024-04-19 | 上海四方无锡锅炉工程有限公司 | Control method for adaptive hierarchical air supply and solid waste CFB boiler thereof |
| CN117977579A (en) * | 2024-03-28 | 2024-05-03 | 西华大学 | DWT-LSTM neural network-based adjustable load prediction method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112884213B (en) | 2024-04-19 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112884213B (en) | Coal-fired boiler NOx prediction method based on wavelet decomposition and dynamic mixed deep learning | |
| CN108256697B (en) | Prediction method for short-term load of power system | |
| CN110610232A (en) | Long-term and short-term traffic flow prediction model construction method based on deep learning | |
| Chen et al. | Short-term wind speed forecasting with principle-subordinate predictor based on Conv-LSTM and improved BPNN | |
| CN111582551B (en) | Wind power plant short-term wind speed prediction method and system and electronic equipment | |
| CN109886464B (en) | Low-information-loss short-term wind speed prediction method based on optimized singular value decomposition generated feature set | |
| CN110826791A (en) | Hybrid wind power prediction method based on long-time and short-time memory neural network | |
| CN114363195B (en) | Network flow prediction and early warning method for time and frequency spectrum residual convolution network | |
| CN110956312A (en) | Photovoltaic power distribution network voltage prediction method based on EMD-CNN deep neural network | |
| Zhang et al. | Photovoltaic power prediction based on hybrid modeling of neural network and stochastic differential equation | |
| CN112945162A (en) | Accumulation layer landslide displacement prediction model and prediction method | |
| CN116258269A (en) | Ultra-short-term load dynamic prediction method based on load characteristic decomposition | |
| CN111008726A (en) | Class image conversion method in power load prediction | |
| CN114548532A (en) | VMD-based TGCN-GRU ultra-short-term load prediction method and device and electronic equipment | |
| CN113673768A (en) | Wind power probability prediction model based on VMD and CNN-LSTM | |
| CN110782013B (en) | Thermal power plant NOx emission prediction method based on generalized mutual entropy self-encoder | |
| CN116819423A (en) | Method and system for detecting abnormal running state of gateway electric energy metering device | |
| CN113361782B (en) | Photovoltaic power generation power short-term rolling prediction method based on improved MKPLS | |
| CN114386452A (en) | Method for detecting faults of sun wheel of nuclear power circulating water pump | |
| CN113111592A (en) | Short-term wind power prediction method based on EMD-LSTM | |
| Zhu et al. | Wind Speed Short-Term Prediction Based on Empirical Wavelet Transform, Recurrent Neural Network and Error Correction | |
| CN116167465A (en) | Solar irradiance prediction method based on multivariate time series ensemble learning | |
| CN114817847B (en) | Energy storage power station intelligent monitoring method based on multilayer feedforward neural network | |
| CN115330082A (en) | PM2.5 concentration prediction method of LSTM network based on attention mechanism | |
| CN114896895A (en) | Gate-controlled cyclic neural network-based denitration system inlet nitrogen oxide emission prediction method and system |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |