Cement kiln flue gas NOx content prediction method and device based on mixed deep learning
By combining random forest, time convolution network and deep echo state network, the NOx concentration prediction model is constructed, which solves the problem of failure to effectively capture the timing dependence of NOx concentration in cement kiln flue gas in the prior art, and improves the prediction accuracy.
Patent Information
- Application Number
- CN202411262955.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-10
- Publication Date
- 2024-11-12
AI Technical Summary
The existing method of cement kiln flue gas NOx concentration prediction based on deep neural networks failed to effectively consider the continuous slow time-varying process and dynamic characteristics of cement clinker calcination, ignored the time dependence between working conditions data, and affected the prediction accuracy. .
Using a hybrid deep learning-based approach, the input sequence is reconstructed and constructed by selecting candidate variables related to the NOx generation mechanism, using a random forest to evaluate the importance of variables, combining time convolutional network (TCN) and deep echo state network (DESN). NOx concentration prediction model, capturing the timing dependence and dynamic characteristics of sequence data.
It improves the online prediction accuracy of NOx concentration of cement kiln flue gas, effectively captures the timing dependence and dynamic correlation of sequence data, and improves the prediction accuracy.
Smart Images

Figure CN118937599A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a method and a device for predicting the NOx content in cement kiln flue gas based on hybrid deep learning, and belongs to the technical field of NOx content prediction in cement production. Background Art
[0002] The cement industry is a basic industry of the national economy. With the increase in cement demand, a large amount of NOx is emitted during the cement clinker calcination process, causing serious pollution to the environment. Therefore, the state has issued relevant policies to limit nitrogen oxide emissions. Real-time prediction of the NOx concentration generated during the cement clinker calcination process is the key to controlling pollution.
[0003] At present, the existing implementation scheme that is most similar to the present invention is "a method for predicting NOx concentration in flue gas of cement production lines based on deep neural networks". This method first screens out the relevant variables required for predicting NOx according to the NOx generation mechanism and the process flow of the cement production line, downloads the variable data from the cement database and performs preprocessing; then, the variable data is formed into input data in the form of a sliding window, so that the input data contains the time delay characteristics of each variable; then, the input layer and hidden layer of the DNN network are used as DBN for unsupervised training to extract the time delay characteristics, and the DNN network is reversely supervised trained with the BP algorithm to extract the corresponding relationship characteristics; finally, the historical data is combined with the model prediction value to make a rolling prediction of NOx for a period of time in the future.
[0004] In the above implementation scheme, the method for predicting the NOx concentration in the flue gas of a cement production line based on a deep neural network does not take into account that the calcination of cement clinker is a continuous slow time-varying process and dynamic characteristics, and ignores the time dependence between the working condition data, which affects the prediction accuracy of the model. Therefore, the present invention proposes a method and device for predicting the NOx content in the flue gas of a cement kiln based on hybrid deep learning. Summary of the invention
[0005] In order to solve the above problems, the present invention proposes a method and device for predicting the NOx content in cement kiln flue gas based on hybrid deep learning, which can perform online prediction of the NOx concentration in cement kiln flue gas and improve the prediction accuracy.
[0006] The technical solution adopted by the present invention to solve the technical problem is:
[0007] In a first aspect, an embodiment of the present invention provides a method for predicting NOx content in cement kiln flue gas based on hybrid deep learning, comprising the following steps:
[0008] Step 1, according to the generation mechanism of nitrogen oxides in the flue gas of the rotary kiln of the cement production line, select candidate variables for the NOx concentration prediction model;
[0009] Step 2, online acquisition of candidate variables and output variable data of the NOx concentration prediction model and data preprocessing, wherein the output variable of the NOx concentration prediction model is the NOx concentration feedback value of the smoke chamber at the next moment;
[0010] Step 3, using random forest to evaluate the importance of each candidate variable, and selecting the candidate variables with higher importance as the input variables of the NOx concentration prediction model to obtain n groups of processed data;
[0011] Step 4, reconstructing the input sequence of the NOx concentration prediction model based on the TCN network (Temporal Convolutional Network);
[0012] Step 5, constructing a NOx concentration prediction model based on the DESN network (Deep Echo State Network);
[0013] Step 6: Use the NOx concentration prediction model to predict the NOx content in cement kiln flue gas.
[0014] As a possible implementation method of this embodiment, the candidate variables include: NOx concentration feedback value in the smoke chamber, decomposition furnace outlet temperature feedback, kiln main machine current feedback, kiln head coal feeding feedback value, kiln head coal injection negative pressure, suspension preheater C5A feed pipe temperature, suspension preheater C5B feed pipe temperature, suspension preheater C5A outlet temperature feedback, suspension preheater C5B outlet temperature feedback, smoke chamber temperature, head coal Roots blower current, high temperature blower current, cooler exhaust air temperature, cooler exhaust air pressure, coal mill head pulverized coal bin temperature, kiln head primary fan current and kiln raw material feeding amount feedback value.
[0015] As a possible implementation of this embodiment, in step 2, the sampling period for online acquisition of candidate variables and output variable data of the NOx concentration prediction model is 1 minute, and the candidate variables and output variable data of the NOx concentration prediction model are sampled k times continuously.
[0016] As a possible implementation of this embodiment, in step 2, the data preprocessing includes:
[0017] The candidate variables and output variable data of the sampled NOx concentration prediction model are firstly eliminated using the 3σ criterion, and then the missing data are filled by linear interpolation method, and then the maximum and minimum normalization methods are used to normalize the data.
[0018] As a possible implementation of this embodiment, in step 3, the importance of each candidate variable is evaluated using random forest, and the candidate variables with higher importance are selected as input variables of the NOx concentration prediction model to obtain n groups of processed data:
[0019] Ω={(x(t)|t=1,2,…,n}
[0020] Among them, x(t) = [x1(t); x2(t); x3(t); x4(t); x5(t); x6(t); x7(t); x8(t)], x1 is the feedback value of NOx concentration in the smoke chamber, x2 is the feedback value of coal feeding at the kiln head, x3 is the temperature of the discharge pipe of suspension preheater C5A, x4 is the smoke chamber temperature, x5 is the current of the high-temperature fan, x6 is the temperature of the residual air at the outlet of the cooler, x7 is the residual air pressure at the outlet of the cooler, and x8 is the feedback value of the amount of raw material discharged into the kiln.
[0021] As a possible implementation of this embodiment, step 4 includes the following steps:
[0022] Step 41, in the causal convolutional layer, the sliding window covers the input sequence x at time t i (t-τ) to x i (t-1), in the sliding window, the convolution kernel and the covered input sequence perform a convolution operation, move the convolution kernel one position to the right, and repeat the above convolution operation until the entire sliding window is covered, and the output at time t is obtained
[0023]
[0024] in, is the output of the causal convolutional layer at time t, * is the convolution operator, w i ∈R 1×d is the weight of the convolution kernel of the input variable, and each input variable has a different convolution kernel size, w i =[w i,1 ,w i,2 ,…,w i,d ], i is the serial number of the input variable, i = 1, 2, ..., 8; τ represents the size of the sliding window;
[0025] Step 42, perform batch normalization at the batch normalization layer:
[0026]
[0027] Among them, μ and δ 2 They are The mean and variance of , ε is a constant, α is an offset, and γ is a scale factor;
[0028] Introducing the nonlinear activation function Relu in the Relu layer yields:
[0029]
[0030] Step 43, after two cycles in the above convolutional layer, batch normalization layer and Relu layer, the fully connected layer is introduced:
[0031]
[0032] Among them, u i (t) is the output of the TCN network, It represents the weight of the fully connected layer obtained by the back-propagation algorithm, and b is the bias term.
[0033] As a possible implementation of this embodiment, step 5 includes the following steps:
[0034] Step 51: Constructing the training sample set required for the NOx concentration prediction model
[0035]
[0036] Among them, u(t)=[u1(t); u2(t); u3(t); u4(t); u5(t); u6(t); u7(t); u8(t)], d(t)=u1(t+1);
[0037] Step 52, setting parameters of the deep echo state network;
[0038] Step 53, set the number of cleaning steps wp, construct the dynamic information matrix H(t) and the target output matrix D in the deep echo state network;
[0039] H(t)=[u(t); S (1) (t); S (2) (t); ...; S (m) (t)] (10)
[0040] D=[d(wp+1); d(wp+2);…;d(n-τ)] (11)
[0041] Among them, wp is the number of cleaning steps, S (m) (t) is the actual output of the mth reservoir in the deep echo state network at time t;
[0042] Step 54, calculate the output weight matrix of the deep echo state network:
[0043] W out =DH T (HH T +αE) -1 (12)
[0044] Among them, α is the regularization parameter, E is the T The identity matrix of the same shape, ()-1 Represents the inverse operation of a matrix;
[0045] Step 55, calculate the predicted value of NOx concentration at the next moment according to the deep echo state network prediction model at moment t:
[0046] o(t)=W out H(t) (13)
[0047] Among them, W out ∈R 1×(K+m×N) is the output weight matrix of the deep echo state network, and H(t) is the dynamic information matrix at time t.
[0048] As a possible implementation of this embodiment, in step 52, the number of input neurons is K=8, the number of output neurons is 1, and the number of reserve pools is m; and the parameters of the mth reserve pool in the deep echo state network are set: the number of neurons in the reserve pool is N, the spectral radius ρ (m) , sparsity sp (m) ;
[0049] The step 52 comprises the following steps:
[0050] Step 52.1, randomly generate the network input weight matrix with uniform distribution in the interval (-1,1)
[0051]
[0052] Step 52.2, randomly generate sparsity sp in the interval (-1,1) using uniform distribution (m) The matrix
[0053] Step 52.3, calculate the matrix W (m) The maximum eigenvalue ρ max (W (m) );
[0054] Step 52.4, calculate the weight matrix of the mth reserve pool in the deep echo state network
[0055]
[0056] Step 52.5, let t = 1;
[0057] Step 52.6, set the input signal of the mth reserve pool in the deep echo state network at time t:
[0058]
[0059] Among them, S (m-1)(t) is the actual output of the m-1th reserve pool in the deep echo state network at time t;
[0060] Step 52.7, setting the target output of the deep echo state network at time t to d(t)=u1(t+1);
[0061] Step 52.8, calculate the neuron state of the mth reserve pool in the deep echo state network at time t:
[0062]
[0063] Among them, S (m) (0) = 0;
[0064] Step 52.9, update t: t←t+1;
[0065] Step 52.10, repeat steps 52.6 to 52.10 until the updated t=n-τ.
[0066] In a second aspect, an embodiment of the present invention provides a cement kiln flue gas NOx content prediction device based on hybrid deep learning, comprising:
[0067] A candidate variable selection module is used to select candidate variables for the NOx concentration prediction model according to the generation mechanism of nitrogen oxides in the flue gas of the rotary kiln of the cement production line;
[0068] A data acquisition and preprocessing module is used to obtain candidate variables and output variable data of the NOx concentration prediction model online and perform data preprocessing. The output variable of the NOx concentration prediction model is the NOx concentration feedback value of the smoke chamber at the next moment.
[0069] An input variable determination module is used to evaluate the importance of each candidate variable using random forest, and select candidate variables with higher importance as input variables of the NOx concentration prediction model to obtain n groups of processed data;
[0070] An input sequence reconstruction module is used to reconstruct the input sequence of the NOx concentration prediction model based on the TCN network;
[0071] Prediction model building module, used to build NOx concentration prediction model based on DESN network;
[0072] The NOx content prediction module is used to predict the NOx content in cement kiln flue gas using the NOx concentration prediction model.
[0073] As a possible implementation method of this embodiment, the candidate variables include: NOx concentration feedback value in the smoke chamber, decomposition furnace outlet temperature feedback, kiln main machine current feedback, kiln head coal feeding feedback value, kiln head coal injection negative pressure, suspension preheater C5A feed pipe temperature, suspension preheater C5B feed pipe temperature, suspension preheater C5A outlet temperature feedback, suspension preheater C5B outlet temperature feedback, smoke chamber temperature, head coal Roots blower current, high temperature blower current, cooler exhaust air temperature, cooler exhaust air pressure, coal mill head pulverized coal bin temperature, kiln head primary fan current and kiln raw material feeding amount feedback value.
[0074] The technical solution of the embodiment of the present invention may have the following beneficial effects:
[0075] A method for predicting NOx content in cement kiln flue gas based on hybrid deep learning according to the technical solution of an embodiment of the present invention comprises the following steps: Step 1, selecting candidate variables of a NOx concentration prediction model according to the generation mechanism of nitrogen oxides in the flue gas of a rotary kiln of a cement production line; Step 2, acquiring candidate variables and output variable data of the NOx concentration prediction model online and performing data preprocessing, wherein the output variable of the NOx concentration prediction model is the feedback value of NOx concentration in the smoke chamber at the next moment; Step 3, evaluating the importance of each candidate variable using a random forest, and selecting candidate variables with higher importance as input variables of the NOx concentration prediction model to obtain n groups of processed data; Step 4, reconstructing the input sequence of the NOx concentration prediction model based on a TCN network; Step 5, constructing a NOx concentration prediction model based on a DESN network; Step 6, predicting the NOx content in cement kiln flue gas using the NOx concentration prediction model. The present invention is based on the NOx generation mechanism and the input variables of the random forest selection hybrid model, reconstructs the input sequence based on the time convolutional neural network (TCN), and effectively captures the temporal dependence of the sequence data; uses the deep echo state network to establish the dynamic correlation between NOx concentration and temporal characteristics, and predicts the NOx concentration in cement kiln flue gas online, thereby improving the prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0076] Figure 1 is a flow chart of a method for predicting NOx content in cement kiln flue gas based on hybrid deep learning according to an exemplary embodiment;
[0077] Figure 2 It is a structural diagram of a method and device for predicting NOx content in cement kiln flue gas based on hybrid deep learning according to an exemplary embodiment;
[0078] Figure 3 is a flow chart showing a method of predicting the NOx content in cement kiln flue gas using the device of the present invention according to an exemplary embodiment;
[0079] Figure 4 is a topological structure diagram of a NOx concentration prediction model according to an exemplary embodiment;
[0080] Figure 5 It is a comparison chart of the prediction result of the present invention and the measured NOx concentration value;
[0081] Figure 6 It is an error diagram between the prediction result of the present invention and the actual measured NOx concentration. DETAILED DESCRIPTION
[0082] The present invention will be further described below in conjunction with the accompanying drawings and embodiments:
[0083] In order to clearly illustrate the technical features of the present solution, the present invention is described in detail below through specific implementation methods and in conjunction with the accompanying drawings. The disclosure below provides many different embodiments or examples for realizing different structures of the present invention. In order to simplify the disclosure of the present invention, the components and settings of specific examples are described below. In addition, the present invention may repeat reference numbers and / or letters in different examples. This repetition is for the purpose of simplification and clarity, and does not itself indicate the relationship between the various embodiments and / or settings discussed. It should be noted that the components illustrated in the accompanying drawings are not necessarily drawn to scale. The present invention omits the description of known components and processing techniques and processes to avoid unnecessary limitations on the present invention.
[0084] like Figure 1 As shown, a method for predicting NOx content in cement kiln flue gas based on hybrid deep learning provided by an embodiment of the present invention includes the following steps:
[0085] Step 1, according to the generation mechanism of nitrogen oxides in the flue gas of the rotary kiln of the cement production line, select candidate variables for the NOx concentration prediction model;
[0086] Step 2, online acquisition of candidate variables and output variable data of the NOx concentration prediction model and data preprocessing, wherein the output variable of the NOx concentration prediction model is the NOx concentration feedback value of the smoke chamber at the next moment;
[0087] Step 3, using random forest to evaluate the importance of each candidate variable, and selecting the candidate variables with higher importance as the input variables of the NOx concentration prediction model to obtain n groups of processed data;
[0088] Step 4, reconstructing the input sequence of the NOx concentration prediction model based on the TCN network;
[0089] Step 5, constructing a NOx concentration prediction model based on the DESN network;
[0090] Step 6: Use the NOx concentration prediction model to predict the NOx content in cement kiln flue gas.
[0091] As a possible implementation method of this embodiment, the candidate variables include: NOx concentration feedback value in the smoke chamber, decomposition furnace outlet temperature feedback, kiln main machine current feedback, kiln head coal feeding feedback value, kiln head coal injection negative pressure, suspension preheater C5A feed pipe temperature, suspension preheater C5B feed pipe temperature, suspension preheater C5A outlet temperature feedback, suspension preheater C5B outlet temperature feedback, smoke chamber temperature, head coal Roots blower current, high temperature blower current, cooler exhaust air temperature, cooler exhaust air pressure, coal mill head pulverized coal bin temperature, kiln head primary fan current and kiln raw material feeding amount feedback value.
[0092] As a possible implementation of this embodiment, in step 2, the sampling period for online acquisition of candidate variables and output variable data of the NOx concentration prediction model is 1 minute, and the candidate variables and output variable data of the NOx concentration prediction model are sampled k times continuously.
[0093] As a possible implementation of this embodiment, in step 2, the data preprocessing includes:
[0094] The candidate variables and output variable data of the sampled NOx concentration prediction model are firstly eliminated using the 3σ criterion, and then the missing data are filled by linear interpolation method, and then the maximum and minimum normalization methods are used to normalize the data.
[0095] As a possible implementation of this embodiment, in step 3, the importance of each candidate variable is evaluated using random forest, and the candidate variables with higher importance are selected as input variables of the NOx concentration prediction model to obtain n groups of processed data:
[0096] Ω={(x(t)|t=1,2,…,n}
[0097] Among them, x(t) = [x1(t); x2(t); x3(t); x4(t); x5(t); x6(t); x7(t); x8(t)], x1 is the feedback value of NOx concentration in the smoke chamber, x2 is the feedback value of coal feeding at the kiln head, x3 is the temperature of the discharge pipe of suspension preheater C5A, x4 is the smoke chamber temperature, x5 is the current of the high-temperature fan, x6 is the temperature of the residual air at the outlet of the cooler, x7 is the residual air pressure at the outlet of the cooler, and x8 is the feedback value of the amount of raw material discharged into the kiln.
[0098] As a possible implementation of this embodiment, step 4 includes the following steps:
[0099] Step 41, in the causal convolutional layer, the sliding window covers the input sequence x at time t i (t-τ) to x i(t-1), in the sliding window, the convolution kernel and the covered input sequence perform a convolution operation, move the convolution kernel one position to the right, and repeat the above convolution operation until the entire sliding window is covered, and the output at time t is obtained
[0100]
[0101] in, is the output of the causal convolutional layer at time t, * is the convolution operator, w i ∈R 1×d is the weight of the convolution kernel of the input variable, and each input variable has a different convolution kernel size, w i =[w i,1 ,w i,2 ,…,w i,d ], i is the serial number of the input variable, i = 1, 2, ..., 8; τ represents the size of the sliding window;
[0102] Step 42, perform batch normalization at the batch normalization layer:
[0103]
[0104] Among them, μ and δ 2 They are The mean and variance of , ε is a constant, α is an offset, and γ is a scale factor;
[0105] Introducing the nonlinear activation function Relu in the Relu layer yields:
[0106]
[0107] Step 43, after two cycles in the above convolutional layer, batch normalization layer and Relu layer, the fully connected layer is introduced:
[0108]
[0109] Among them, u i (t) is the output of the TCN network, It represents the weight of the fully connected layer obtained by the back-propagation algorithm, and b is the bias term.
[0110] As a possible implementation of this embodiment, step 5 includes the following steps:
[0111] Step 51: Constructing the training sample set required for the NOx concentration prediction model
[0112]
[0113] Among them, u(t)=[u1(t); u2(t); u3(t); u4(t); u5(t); u6(t); u7(t); u8(t)], d(t)=u1(t+1);
[0114] Step 52, setting parameters of the deep echo state network;
[0115] Step 53, set the number of cleaning steps wp, construct the dynamic information matrix H(t) and the target output matrix D in the deep echo state network;
[0116] H(t)=[u(t); S (1) (t); S (2) (t);…;S (m) (t)] (10)
[0117] D=[d(wp+1); d(wp+2);…;d(n-τ)] (11)
[0118] Among them, wp is the number of cleaning steps, S (m) (t) is the actual output of the mth reservoir in the deep echo state network at time t;
[0119] Step 54, calculate the output weight matrix of the deep echo state network:
[0120] W out =DH T (HH T +αE) -1 (12)
[0121] Among them, α is the regularization parameter, E is the T The identity matrix of the same shape, () -1 Represents the inverse operation of a matrix;
[0122] Step 55, calculate the predicted value of NOx concentration at the next moment according to the deep echo state network prediction model at moment t:
[0123] o(t)=W out H(t) (13)
[0124] Among them, W out ∈R 1×(K+m×N) is the output weight matrix of the deep echo state network, and H(t) is the dynamic information matrix at time t.
[0125] As a possible implementation of this embodiment, in step 52, the number of input neurons is K=8, the number of output neurons is 1, and the number of reserve pools is m; and the parameters of the mth reserve pool in the deep echo state network are set: the number of neurons in the reserve pool is N, the spectral radius ρ (m) , sparsity sp(m) ;
[0126] The step 52 comprises the following steps:
[0127] Step 52.1, randomly generate the network input weight matrix with uniform distribution in the interval (-1,1)
[0128]
[0129] Step 52.2, randomly generate sparsity sp in the interval (-1,1) using uniform distribution (m) The matrix
[0130] Step 52.3, calculate the matrix W (m) The maximum eigenvalue ρ max (W (m) );
[0131] Step 52.4, calculate the weight matrix of the mth reserve pool in the deep echo state network
[0132]
[0133] Step 52.5, let t = 1;
[0134] Step 52.6, set the input signal of the mth reserve pool in the deep echo state network at time t:
[0135]
[0136] Among them, S (m-1) (t) is the actual output of the m-1th reserve pool in the deep echo state network at time t;
[0137] Step 52.7, setting the target output of the deep echo state network at time t to d(t)=u1(t+1);
[0138] Step 52.8, calculate the neuron state of the mth reservoir in the deep echo state network at time t:
[0139]
[0140] Among them, S (m) (0) = 0;
[0141] Step 52.9, update t: t←t+1;
[0142] Step 52.10, repeat steps 52.6 to 52.10 until the updated t=n-τ.
[0143] like Figure 2 As shown, an embodiment of the present invention provides a cement kiln flue gas NOx content prediction device based on hybrid deep learning, comprising:
[0144] A candidate variable selection module is used to select candidate variables for the NOx concentration prediction model according to the generation mechanism of nitrogen oxides in the flue gas of the rotary kiln of the cement production line;
[0145] A data acquisition and preprocessing module is used to obtain candidate variables and output variable data of the NOx concentration prediction model online and perform data preprocessing. The output variable of the NOx concentration prediction model is the NOx concentration feedback value of the smoke chamber at the next moment.
[0146] An input variable determination module is used to evaluate the importance of each candidate variable using random forest, and select candidate variables with higher importance as input variables of the NOx concentration prediction model to obtain n groups of processed data;
[0147] An input sequence reconstruction module is used to reconstruct the input sequence of the NOx concentration prediction model based on the TCN network;
[0148] Prediction model building module, used to build NOx concentration prediction model based on DESN network;
[0149] The NOx content prediction module is used to predict the NOx content in cement kiln flue gas using the NOx concentration prediction model.
[0150] As a possible implementation method of this embodiment, the candidate variables include: NOx concentration feedback value in the smoke chamber, decomposition furnace outlet temperature feedback, kiln main machine current feedback, kiln head coal feeding feedback value, kiln head coal injection negative pressure, suspension preheater C5A feed pipe temperature, suspension preheater C5B feed pipe temperature, suspension preheater C5A outlet temperature feedback, suspension preheater C5B outlet temperature feedback, smoke chamber temperature, head coal Roots blower current, high temperature blower current, cooler exhaust air temperature, cooler exhaust air pressure, coal mill head pulverized coal bin temperature, kiln head primary fan current and kiln raw material feeding amount feedback value.
[0151] like Figure 3 As shown, the specific process of using the device of the present invention to predict the NOx content in cement kiln flue gas is as follows.
[0152] 1. Determine candidate variables based on the NOx generation mechanism.
[0153] According to the generation mechanism of nitrogen oxides in the flue gas of the rotary kiln of the cement production line, candidate variables related to NOx concentration were preliminarily selected; the candidate input variables of the model are as follows: the candidate variables of the NOx concentration prediction model include the NOx concentration feedback value in the smoke chamber, the decomposition furnace outlet temperature feedback, the kiln main machine current feedback, the kiln head coal feeding feedback value, the kiln head coal injection negative pressure, the suspension preheater C5A discharge pipe temperature, the suspension preheater C5B discharge pipe temperature, the suspension preheater C5A outlet temperature feedback, the suspension preheater C5B outlet temperature feedback, the smoke chamber temperature, the head coal Roots blower current, the high temperature blower current, the cooler exhaust air temperature, the cooler exhaust air pressure, the coal mill head pulverized coal bin temperature, the kiln head primary fan current and the raw material feeding amount feedback value into the kiln; the output variable of the NOx concentration prediction model is the NOx concentration feedback value in the smoke chamber at the next moment.
[0154] 2. Data collection and preprocessing.
[0155] The candidate variables and output variable data of the model are obtained online, with a sampling period of 1 minute, and the candidate variables and output variable data of the NOx concentration prediction model are sampled k times continuously;
[0156] The sampled data were normalized using the 3σ criterion to remove abnormal data, linear interpolation to fill missing data, and maximum and minimum normalization methods to prevent the influence of different dimensions on the prediction accuracy.
[0157] 3. Select input variables based on random forest.
[0158] Use random forest to evaluate the importance of each candidate variable, and select the candidate variables with higher importance as the input variables of the NOx concentration prediction model to obtain the processed n groups of data:
[0159] Ω={(x(t)|t=1,2,…,n}
[0160] Among them, x(t) = [x1(t); x2(t); x3(t); x4(t); x5(t); x6(t); x7(t); x8(t)], x1 is the feedback value of NOx concentration in the smoke chamber, x2 is the feedback value of coal feeding at the kiln head, x3 is the temperature of the discharge pipe of suspension preheater C5A, x4 is the smoke chamber temperature, x5 is the current of the high-temperature fan, x6 is the temperature of the residual air at the outlet of the cooler, x7 is the residual air pressure at the outlet of the cooler, and x8 is the feedback value of the amount of raw material discharged into the kiln.
[0161] 4. Reconstruct the input sequence based on TCN.
[0162] The structure of TCN is as follows: The causal convolutional layer is the first layer of TCN, which is designed to process time series data by enforcing causal constraints. The batch normalization in the second layer helps to improve the training efficiency of the network. Due to the complexity of the data, the nonlinear activation function Relu is introduced in the second layer. The third and fourth layers are consistent with the structure of the first and second layers. The fully connected layer is the last layer of the network to generate the output of TCN.
[0163] The i-th (i=1,2,…,8) TCN processes the i-th input variable independently through a sliding window. i (t-τ) to x i The specific steps of reconstructing the input sequence based on TCN are as follows:
[0164] S4.1: In the causal convolutional layer, the sliding window covers the input sequence x at time t i (t-τ) to x i (t-1), in the sliding window, the convolution kernel and the covered input sequence perform a convolution operation, move the convolution kernel one position to the right, and repeat the above convolution operation until the entire sliding window is covered, and the output at time t is obtained
[0165]
[0166] in, is the output of the causal convolutional layer at time t, * is the convolution operator, w i ∈R 1×d is the weight of the convolution kernel of the input variable, and each input variable has a different convolution kernel size, w i =[w i,1 ,w i,2 ,…,w i,d ], i is the serial number of the input variable, i = 1, 2, ..., 8; τ represents the size of the sliding window.
[0167] S4.2: After the causal convolution calculation process is completed, batch normalization is performed:
[0168]
[0169] Among them, μ and δ 2 They are The mean and variance of . ε is a default value of 10 -5 and 10 -9 The α offset and γ scaling factor are learnable parameters that are updated during network training via the back-propagation algorithm. The target signal is x i (t).
[0170] Due to the complexity of the data, the nonlinear activation function Relu is introduced to obtain:
[0171]
[0172] S4.3: In the above convolutional layer, batch normalization layer and ReLU layer, two cycles are performed. Subsequently, a fully connected layer is introduced, which is expressed as:
[0173]
[0174] Among them, u i (t) is the output of TCN. represents the weight of the fully connected layer obtained by the back propagation algorithm, and the target output is x i (t), b is the bias term.
[0175] 5. Build a prediction model based on DESN.
[0176] Prediction models such as Figure 4 As shown, the construction process is as follows:
[0177] S5.1: Training sample set required to build the prediction model:
[0178]
[0179] Among them, u(t)=[u1(t); u2(t); u3(t); u4(t); u5(t); u6(t); u7(t); u8(t)], d(t)=u1(t+1).
[0180] S5.2: Set the parameters of the deep echo state network: the number of input neurons K = 8, the number of output neurons 1, and the number of reserve pools m;
[0181] Set the parameters of the mth reservoir in the deep echo state network: the number of neurons in the reservoir N, the spectral radius ρ (m) , sparsity sp (m) ;
[0182] S5.2.1: Generate the network input weight matrix randomly with uniform distribution in the interval (-1,1)
[0183]
[0184] S5.2.2: Generate random numbers with sparsity sp using uniform distribution on the interval (-1,1). (m) The matrix
[0185] S5.2.3: Compute the matrix W(m) The maximum eigenvalue ρ max (W (m) );
[0186] S5.2.4: Compute the weight matrix of the mth reservoir in the deep echo state network
[0187]
[0188] S5.2.5: Let t = 1;
[0189] S5.2.6: Set the input signal of the mth reservoir in the deep echo state network at time t:
[0190]
[0191] Among them, S (m-1) (t) is the actual output of the m-1th reserve pool in the deep echo state network at time t;
[0192] S5.2.7: Set the target output of the deep echo state network at time t to d(t) = u1(t+1);
[0193] S5.2.8: Compute the neuron state of the mth reservoir in the deep echo state network at time t:
[0194]
[0195] Among them, S (m) (0) = 0;
[0196] S5.2.9: Update t: t←t+1;
[0197] S5.2.10: Repeat S5.1.6 to S5.1.10 until t = n - τ.
[0198] S5.3: Set the number of cleaning steps wp, construct the dynamic information matrix H(t) in the deep echo state network, and the target output matrix D;
[0199]
[0200] Among them, wp is the number of cleaning steps, S (m) (t) is the actual output of the mth reserve pool in the deep echo state network at time t.
[0201] S5.4: Calculate the output weight matrix of the deep echo state network:
[0202] W out =DH T (HH T +αE)-1 (12)
[0203] Among them, α is the regularization parameter, E is the T The identity matrix of the same shape, () -1 Represents the inverse operation of a matrix.
[0204] S5.5: Calculate the predicted value of NOx concentration at the next moment according to the deep echo state network prediction model at moment t:
[0205] o(t)=W out H(t) (13)
[0206] Among them, W out ∈R 1×(K+m×N) is the output weight matrix of the deep echo state network, and H(t) is the dynamic information matrix at time t.
[0207] Step 6: Predict the NOx content in cement kiln flue gas.
[0208] The present invention uses 3500 sets of data from a cement production line DCS system, of which the first 3000 sets are used for training, the last 500 sets are used for testing, and the first 100 steps are removed. The test results are as follows: Figure 5 As shown in the figure, the X-axis is the sample (samples / 1 minute), the Y-axis is the NOx concentration (mg / m3), the dotted line is the predicted result, and the solid line is the measured NOx concentration. The comparison and error between the model prediction result and the measured value of the NOx content are shown in the figure. Figure 6 As shown, X-axis: samples (pieces / 1 minute), Y-axis: the error between the model prediction result and the measured NOx concentration, the results prove the effectiveness of the present invention.
[0209] The present invention proposes a hybrid deep learning data-driven prediction method, which selects input variables based on mechanism analysis and random forest method to overcome noise interference; uses TCN to reconstruct the input sequence and capture the time series characteristics of the input sequence; uses DESN to establish the dynamic correlation between NOx concentration and the time series characteristics of the input sequence, and realizes the online prediction of cement kiln NOx concentration, which improves the prediction accuracy. Compared with the current measurement technology, the present invention effectively captures the time series dependence of sequence data and improves the prediction accuracy.
[0210] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.
Claims
1. A method for predicting NOx content in cement kiln flue gas based on hybrid deep learning, characterized in that: The following steps are involved: Step 1, according to the generation mechanism of nitrogen oxides in the flue gas of the rotary kiln of the cement production line, select candidate variables for the NOx concentration prediction model; Step 2, online acquisition of candidate variables and output variable data of the NOx concentration prediction model and data preprocessing, wherein the output variable of the NOx concentration prediction model is the NOx concentration feedback value of the smoke chamber at the next moment; Step 3, using random forest to evaluate the importance of each candidate variable, and selecting the candidate variables with higher importance as the input variables of the NOx concentration prediction model to obtain n groups of processed data; Step 4, reconstructing the input sequence of the NOx concentration prediction model based on the TCN network; Step 5, constructing a NOx concentration prediction model based on the DESN network; Step 6: Use the NOx concentration prediction model to predict the NOx content in cement kiln flue gas.
2. The method for predicting NOx content in cement kiln flue gas based on hybrid deep learning according to claim 1 is characterized in that: The candidate variables include: NOx concentration feedback value in smoke chamber, decomposition furnace outlet temperature feedback, kiln main machine current feedback, kiln head coal feeding feedback value, kiln head coal injection negative pressure, suspension preheater C5A discharge pipe temperature, suspension preheater C5B discharge pipe temperature, suspension preheater C5A outlet temperature feedback, suspension preheater C5B outlet temperature feedback, smoke chamber temperature, head coal Roots blower current, high temperature blower current, cooler exhaust air temperature, cooler exhaust air pressure, coal mill head pulverized coal bin temperature, kiln head primary fan current and kiln raw material feeding amount feedback value.
3. The method for predicting NOx content in cement kiln flue gas based on hybrid deep learning according to claim 1, characterized in that: In step 2, the sampling period for online acquisition of candidate variables and output variable data of the NOx concentration prediction model is 1 minute, and the candidate variables and output variable data of the NOx concentration prediction model are sampled k times continuously.
4. The method for predicting NOx content in cement kiln flue gas based on hybrid deep learning according to claim 1, characterized in that: In step 2, the data preprocessing includes: The candidate variables and output variable data of the sampled NOx concentration prediction model are firstly eliminated using the 3σ criterion, and then the missing data are filled by linear interpolation method, and then the maximum and minimum normalization methods are used to normalize the data.
5. The method for predicting NOx content in cement kiln flue gas based on hybrid deep learning according to claim 1, characterized in that: In step 3, random forest is used to evaluate the importance of each candidate variable, and the candidate variables with higher importance are selected as the input variables of the NOx concentration prediction model to obtain the processed n groups of data: Ω={(x(t)|t=1,2,…,n} Among them, x(t) = [x1(t); x2(t); x3(t); x4(t); x5(t); x6(t); x7(t); x8(t)], x1 is the feedback value of NOx concentration in the smoke chamber, x2 is the feedback value of coal feeding at the kiln head, x3 is the temperature of the discharge pipe of suspension preheater C5A, x4 is the smoke chamber temperature, x5 is the current of the high-temperature fan, x6 is the temperature of the residual air at the outlet of the cooler, x7 is the residual air pressure at the outlet of the cooler, and x8 is the feedback value of the amount of raw material discharged into the kiln.
6. The method for predicting NOx content in cement kiln flue gas based on hybrid deep learning according to claim 1, characterized in that: The step 4 comprises the following steps: Step 41, in the causal convolutional layer, the sliding window covers the input sequence x at time t i (t-τ) to x i (t-1), in the sliding window, the convolution kernel and the covered input sequence perform a convolution operation, move the convolution kernel one position to the right, and repeat the above convolution operation until the entire sliding window is covered, and the output at time t is obtained in, is the output of the causal convolutional layer at time t, * is the convolution operator, w i ∈R 1×d is the weight of the convolution kernel of the input variable, and each input variable has a different convolution kernel size, w i =[w i,1 ,w i,2 ,…,w i,d ], i is the serial number of the input variable, i = 1, 2, ..., 8; τ represents the size of the sliding window; Step 42, perform batch normalization at the batch normalization layer: Among them, μ and δ 2 They are The mean and variance of , ε is a constant, α is an offset, and γ is a scale factor; Introducing the nonlinear activation function Relu in the Relu layer yields: Step 43, after two cycles in the above convolutional layer, batch normalization layer and Relu layer, the fully connected layer is introduced: Among them, u i (t) is the output of the TCN network, It represents the weight of the fully connected layer obtained by the back-propagation algorithm, and b is the bias term.
7. The method for predicting NOx content in cement kiln flue gas based on hybrid deep learning according to any one of claims 1 to 6, characterized in that: The step 5 comprises the following steps: Step 51: Constructing the training sample set required for the NOx concentration prediction model Among them, u(t)=[u1(t); u2(t); u3(t); u4(t); u5(t); u6(t); u7(t); u8(t)], d(t)=u1(t+1); Step 52, setting parameters of the deep echo state network; Step 53, set the number of cleaning steps wp, construct the dynamic information matrix H(t) and the target output matrix D in the deep echo state network; H(t)=[u(t);S (1) (t);S (2) (t);...;S (m) (t)] (10) D=[d(wp+1);d(wp+2);…;d(n-τ)] (11) Among them, wp is the number of cleaning steps, S (m) (t) is the actual output of the mth reservoir in the deep echo state network at time t; Step 54, calculate the output weight matrix of the deep echo state network: W out =DH T (HH T +αE) -1 (12) Among them, α is the regularization parameter, E is related to HH T The identity matrix of the same shape, () -1 Represents the inverse operation of a matrix; Step 55, calculate the predicted value of NOx concentration at the next moment according to the deep echo state network prediction model at moment t: o(t)=W out H(t) (13) Among them, W out ∈R 1×(K+m×N) is the output weight matrix of the deep echo state network, and H(t) is the dynamic information matrix at time t.
8. The method for predicting NOx content in cement kiln flue gas based on hybrid deep learning according to claim 7, characterized in that: In step 52, the number of input neurons is K=8, the number of output neurons is 1, and the number of storage pools is m; and the parameters of the mth storage pool in the deep echo state network are set: the number of neurons in the storage pool is N, the spectral radius ρ (m) , sparsity sp (m) ; The step 52 comprises the following steps: Step 52.1, randomly generate the network input weight matrix with uniform distribution in the interval (-1,1) Step 52.2, randomly generate sparsity sp in the interval (-1,1) using uniform distribution (m) The matrix Step 52.3, calculate the matrix W (m) The maximum eigenvalue ρ max (W (m) ); Step 52.4, calculate the weight matrix of the mth reserve pool in the deep echo state network Step 52.5, let t = 1; Step 52.6, set the input signal of the mth reserve pool in the deep echo state network at time t: Among them, S (m-1) (t) is the actual output of the m-1th reserve pool in the deep echo state network at time t; Step 52.7, setting the target output of the deep echo state network at time t to d(t)=u1(t+1); Step 52.8, calculate the neuron state of the mth reserve pool in the deep echo state network at time t: Among them, S (m) (0) = 0; Step 52.9, update t: t←t+1; Step 52.10, repeat steps 52.6 to 52.10 until the updated t=n-τ.
9. A cement kiln flue gas NOx content prediction device based on hybrid deep learning, characterized in that: include: A candidate variable selection module is used to select candidate variables for the NOx concentration prediction model according to the generation mechanism of nitrogen oxides in the flue gas of the rotary kiln of the cement production line; A data acquisition and preprocessing module is used to obtain candidate variables and output variable data of the NOx concentration prediction model online and perform data preprocessing. The output variable of the NOx concentration prediction model is the feedback value of the NOx concentration in the smoke chamber at the next moment. An input variable determination module is used to evaluate the importance of each candidate variable using random forest, and select candidate variables with higher importance as input variables of the NOx concentration prediction model to obtain n groups of processed data; An input sequence reconstruction module is used to reconstruct the input sequence of the NOx concentration prediction model based on the TCN network; Prediction model building module, used to build NOx concentration prediction model based on DESN network; The NOx content prediction module is used to predict the NOx content in cement kiln flue gas using the NOx concentration prediction model.
10. The cement kiln flue gas NOx content prediction device based on hybrid deep learning according to claim 9, characterized in that: The candidate variables include: NOx concentration feedback value in smoke chamber, decomposition furnace outlet temperature feedback, kiln main machine current feedback, kiln head coal feeding feedback value, kiln head coal injection negative pressure, suspension preheater C5A discharge pipe temperature, suspension preheater C5B discharge pipe temperature, suspension preheater C5A outlet temperature feedback, suspension preheater C5B outlet temperature feedback, smoke chamber temperature, head coal Roots blower current, high temperature blower current, cooler exhaust air temperature, cooler exhaust air pressure, coal mill head pulverized coal bin temperature, kiln head primary fan current and kiln raw material feeding amount feedback value.
