Channel-selective convolutional neural network method for predicting boiler full-load NOx emissions

Abstract

The present invention discloses a method for predicting boiler full-load NOx emissions using a channel-selective convolutional neural network, and belongs to the technical field of NOx emissions from coal-fired boilers. The historical data of the boiler combustion operation during the selected historical time period is extracted as needed, and then the boiler operation data in the historical data is preprocessed; the preprocessed boiler historical data is divided into a 4-D tensor by a sliding window method, and then the boiler historical data is input into a channel-selective convolutional neural network CS-CNN model in the form of a tensor for training. The CS-CNN model will automatically identify the channels ranked in the top 80% of importance according to the operating condition changes of the input boiler, release the hyperparameters of the channels with low importance, and maximize the use of the limited parameter resources of the CS-CNN model, so as to achieve the purpose of improving the prediction accuracy of the CS-CNN model. The method is simple, has high prediction accuracy, and fast prediction speed.

Description

Technical Field

[0001] The present invention belongs to the technical field of NOx emission from coal-fired boilers, and in particular relates to a method for predicting full-load NOx emission from boilers using a channel selection convolutional neural network. Background Art

[0002] Traditional NOx emission prediction models use dimensionality reduction to prevent redundant input information, and cannot establish a mapping relationship between the model output and the complete controlled variables of the boiler burner. Improving NOx prediction accuracy is not the ultimate goal. Obtaining the mapping relationship between the controlled variables involved in furnace combustion and NOx emissions under corresponding working conditions is the key to optimizing combustion and reducing NOx emissions.

[0003] The CS-CNN network simulates the way the human brain learns through the hippocampus, which is very consistent with the model's requirements for adjusting the parameters involved in combustion control according to the different operating conditions of the boiler's coal mill.

[0004] The prior art with publication number CN106021916A discloses a calculation method suitable for analyzing the NOx emissions of ultra-supercritical boilers, and relates to a supercritical boiler operation control method. 1) Divide the furnace from bottom to top into several chambers according to the furnace structure and calculation accuracy; 2) Input feed parameters; 3) Calculate the reaction rate of each substance in each chamber; 4) Calculate the radiation heat transfer of the flame and flue gas in each chamber to the water-cooled wall and the radiation heat transfer of the adjacent chamber; 5) Establish the mass and energy balance equations of each chamber; 6) Solve the conservation equations and output the data of the NOx generation of each chamber under the current operating conditions; 7) Change the feed parameters and repeat steps 2) to 6) to obtain the data of the NOx generation of each chamber under other operating conditions. It can analyze and calculate the NOx emissions of ultra-supercritical boilers, and can intuitively reflect the influence of feed parameters on the NOx emissions in the furnace. It can only be calculated based on data obtained at a certain moment. Since the types of control variables involved in combustion change in real time with the changes in operating conditions during the actual operation of the boiler, it is related to NO x The types of emission-related variables also change. Existing technologies cannot make continuous predictions for the entire mine and lack dynamic prediction capabilities. Summary of the invention

[0005] In view of the shortcomings of the existing technology, a method for predicting boiler full-load NOx emissions by using a channel selection convolutional neural network is provided. The method is simple, easy to use, and can quickly and efficiently predict boiler full-load NOx emissions.

[0006] To achieve the above purpose, a method for predicting boiler full-load NOx emissions using a channel selection convolutional neural network is provided. Since the actual operation of the boiler changes with the operating conditions, the types of control variables involved in combustion change in real time, so the types of variables related to NOx emissions also change immediately. The CS-CNN model is used to select the importance of the input channel, and dynamic prediction of NOx emissions under different operating conditions is carried out.

[0007] Here are the steps:

[0008] The distributed control system DCS of the power generation boiler continuously uses sensors to extract variable data during boiler operation, including boiler load, total fuel quantity, coal mill capacity air flow, total air volume, secondary air door opening, steam pressure, flow, and temperature;

[0009] The collected variable data of the boiler during operation is stored;

[0010] Select the stored historical time period to extract the historical data of boiler combustion operation as needed, and then pre-process the boiler operation data in the historical data;

[0011] The preprocessed boiler historical data is divided into 4-D tensors by the sliding window method, and then the boiler historical data is input into the channel selective convolutional neural network CS-CNN model in the form of tensors for training. The CS-CNN model will automatically identify the channels ranked in the top 80% of importance according to the operating condition changes of the input boiler, release the hyperparameters of the channels with low importance, and maximize the use of the limited parameter resources of the CS-CNN model to achieve the purpose of improving the prediction accuracy of the CS-CNN model. The operating conditions include the operation mode of the boiler burner damper caused by the changes in the coal mill operation combination;

[0012] Before use, the CS-CNN model needs to be trained and verified using preprocessed historical boiler data. After training, the filtered operating conditions can be predicted.

[0013] Furthermore, the data preprocessing method is to use the 3σ-rule to detect missing and abnormal data that deviates from the normal value, and replace the abnormal data using the following formula:

[0014] Δx i (t) = x i (t)-x i (t-1)

[0015]

[0016]

[0017]

[0018] Where Δx i (t) represents the auxiliary variable x i The rate of change between two adjacent moments is Denotes Δx i (t) the average rate of change before time t, σ i (t) represents the standard deviation, i=1,2,…,m represents the number of variables, and t=2,…,n represents the sampling time.

[0019] Furthermore, the CS-CNN model structure is: sequentially connected input layer, convolution layer, fully connected layer and output layer; the convolution layer contains sequentially connected convolution filters, batch normalization and nonlinear activation functions, and extracts the spatial characteristics of NOx emissions as the target variable from the input tensor by setting multiple convolution filters; the fully connected layer is one or more to perform high-level thrust, connect all neurons of the previous layer to each neuron of the current layer, and then generate global semantic information; the output layer includes a loss function and an activation function, which are different according to the classification and regression tasks. For the classification task, the cross entropy loss function and the softmax nonlinear activation function are used;

[0020] The input layer imitates the single-channel input form of gray images. The convolution layer is used for automatic feature extraction. In CS-CNN, the conventional convolution layer is replaced by the same number of channel selection convolution layers. The first convolution layer still uses the conventional convolution layer. After the convolution layer, a downsampling layer is set to reduce the size of the feature map, retain significant features, increase the receptive field, reduce the impact of noise, and make each feature more robust. After the feature map in the last convolution layer is expanded in one-dimensional form, it will be connected to all neurons in the fully connected layer. Finally, the output layer realizes the continuous prediction of NOx emissions.

[0021] Furthermore, the basic architecture of CS-CNN is a conventional convolutional layer, which includes convolutional filters, batch normalization, and nonlinear activation functions. By setting multiple convolutional filters, NOx emissions are extracted from the input tensor as the spatial features of the target variable. The convolution operation on the kth convolutional filter in the lth layer is as follows:

[0022]

[0023]

[0024] Among them, formula (1) is a matrix operation, formula (2) is an element operation, * represents the convolution operator, They represent the kth feature map and feature map element of the lth layer before batch normalization after the convolution operation, respectively. It is worth noting here that the number of convolution filters is equal to the number of output feature maps. represents the weight matrix of the kth convolution filter in the lth layer, X l-1represents the input matrix of the lth layer, It represents the deviation parameter corresponding to the kth convolution filter of the lth layer, and H' and W' represent the height and width of the convolution filter respectively.

[0025] Furthermore, batch normalization is used to improve the convergence efficiency of the CS-CNN model:

[0026] After batch normalization BN, the input data is distributed in a standardized way, which can adapt to data characteristics of different distributions without further learning, and improve the generalization ability of the network. Batch normalization is performed on the convolution layer between batch normalization and nonlinear activation function: batch normalization of the convolution layer occurs after convolution calculation and before nonlinear activation function;

[0027] Assume that there are m samples in the batch input data. On the kth feature channel of the lth layer, batch normalization is performed on m×H×W elements in the feature channel at the same time. H and W represent the height and width of the feature map respectively. The input of batch normalization is expressed as For any sample in batch Ξ And 1≤i≤m, the batch normalization calculation process is as follows:

[0028]

[0029]

[0030]

[0031] Among them, μ Ξ , represent the mean and variance of Ξ respectively, represents the output of batch normalization, ζ k and β k Represent the stretching parameter and offset parameter of the kth output channel respectively, and each channel has independent stretching and offset parameters;

[0032] After batch normalization, the nonlinear activation function will be applied to all feature maps, and the ReLU function is used as the nonlinear activation function, which is described as follows:

[0033]

[0034] Furthermore, in the CS-CNN model, the channel-selective convolution operation includes De-alloc channel deallocation, Re-alloc channel reallocation and spatial shift, and spatial shift is used to offset the diversity loss caused by too many similar channels. The CS-CNN model completes the training task and realizes the purpose of reallocating the input channels:

[0035] 1). De-alloc channel: for input conditions, prevent boiler variables with low importance from being used in subsequent calculations and release the parameters of these channels;

[0036] 2). Re-alloc channel: Use the Top-C important channels to replace the blocked channels;

[0037] The CS-CNN model retains important channels after multiple trainings; spatial shift is used to offset the diversity loss caused by too many similar channels.

[0038] Furthermore, the expected channel damage matrix ECDM is used to identify the channels that contribute less to the output and the importance of the input channels;

[0039] Since the boiler historical data are arranged in chronological order, setting the convolution kernel size needs to consider the feature extraction at the same time. Therefore, let Conv(X,ω) represent the conventional convolution operation, X represents the input variable of the convolution layer, and ω represents the weight matrix of the convolutional layer, and IN and O represent the number of input and output channels, respectively. H and W represent the height and width of the input feature map, the feature map equivalent channel, and the product kernel size is K. l ×1, and in=1,2,...,IN, h=1,2,...,H, w=1,2,...,W;

[0040] Define ECDM(X;ω) to calculate the expected mean in the time dimension, as shown in formula (7):

[0041]

[0042] When in =0, it indicates that the in-th channel has been set as a low-importance channel and needs to be blocked, and the expected channel damage matrix ECDM thereby identifies the importance of the input channel.

[0043] Furthermore, channel selection and spatial shift are defined: the channel selection convolution layer is defined as SelectConv(X,ω), which implements the selection function of the input channel X based on the conventional convolution layer. The selective convolution layer is expressed as:

[0044] SelectConv(X;ω)=Conv(SelectChannel(X);ω)

[0045] The channel SelectChannel needs to perform channel blocking and re-indexing for De-alloc and Re-alloc respectively; the gate variable g is introduced in ∈{0,1} and index π in∈{1,2,…,IN}, the input channel X is represented as if Does not contribute to the output, the corresponding g in =0; if multiple π in If the channels are identical, it means that a channel has been replicated multiple times. To prevent the serious consequence of weight degradation caused by identical channels, a spatial offset is applied to the reallocated channels. At this time, the description of SelectChannel is as follows:

[0046]

[0047] in={1,2,…,IN}

[0048] π in ={1,2,…,IN}

[0049] g in ∈{0,1}

[0050]

[0051] Among them, shift(X,b) represents the spatial shift operation of the input channel X. For each element (x,y) on the channel, shift(X,b) x,y Defined as:

[0052]

[0053]

[0054] Among them, the parameter b is a continuous value, which is jointly learned with other hyperparameters through stochastic gradient descent SGD, and through a similar spatial shift operation, efficient use of parameters is achieved. By recycling its parameters to selectively expand the kernel size, the number of parameters used is reduced while deepening the model network.

[0055] Furthermore, the channel-selective convolutional layer training scheme is:

[0056] In order to use ECDM to identify the canceled / reallocated channels, the parameters S = (X; g, π, b) are designed in the channel-selective convolutional layer, and then S is iteratively updated during the de-alloc and re-alloc training process; by setting g in = 0 to complete the de-alloc channel. Given the expected damage level γ>0, the de-allocating objective is rewritten as the following optimization problem:

[0057]

[0058]

[0059] g in =0 or 1; in = 1, 2, ..., IN

[0060] If the above formula is used in the complete enumeration method, we will get 2 IN combinations; therefore, when IN is large, a greedy algorithm is used to calculate: ECDM is normalized according to the output dimension nECDM:

[0061]

[0062] Where, let j = 1, 2, ..., O. After nECDM is calculated, the minimum channel ||nECDM(ω; X) in || ∞ Iterate to determine the channel to be released, but ensure that nECDM(ω;X) in Vector sum of l ∞ The norm is less than γ;

[0063] Re-alloc, according to nECDM's l 2 The norm selects the top-C largest channel: ||nECDM(ω;X) in ||2; The first C channels selected randomly occupy the unassigned channels (g in =0); when the in-th channel is de-assigned, the corresponding parameter in ω is set to zero so that the operation does not damage the training; set the maximum reallocation count N max , to limit the redundant allocation of the same feature map; dynamic training is performed using the general stochastic gradient descent method SGD. When training ω through SGD, unassignment and reallocation will update other parameters of S: Unassignment is used to update the entry variable g in , and reallocation Re-alloc is used to update the deviation parameter b and the entry variable g in , π in and the weight matrix ω.

[0064] Further, the fully connected layer and output layer of the CNN model are as follows: one or more fully connected layers are used after the convolutional layer to perform high-level reasoning, connect all neurons in the previous layer to each neuron in the current layer, and then generate global semantic information, and set the fully connected layer. The mapping relationship is as follows:

[0065]

[0066] Where f(·) represents the ReLU nonlinear activation function, represents the output value of the jth node in the lth layer, represents the weight between the jth node in the lth layer and the ith node in the previous layer, represents the input value of the i-th node in the l-th layer, represents the deviation of the lth layer;

[0067] The loss function and activation function in the output layer vary depending on the classification and regression tasks. For classification tasks, the cross entropy loss function and the softmax nonlinear activation function are used. For a model with m batch samples and t output categories, the loss function and activation function are as follows:

[0068]

[0069]

[0070] In the regression task, the mean square error loss MSE function and the linear activation function are used. Taking the single output prediction model as an example, the loss function is as follows:

[0071]

[0072] Among them, y i,j and i is the true value of the i-th sample, y i,j and i is the predicted value of the sample; the NOx emission prediction regression task adopts the mean square error loss function and linear activation function.

[0073] Compared with the prior art, the advantages of the present invention are:

[0074] The present invention provides a channel-selective convolutional layer for replacing a conventional convolutional layer. The channel-selective convolutional layer performs channel importance evaluation on input variables, selects TOP-C important channels and releases hyperparameters of remaining low-importance channels, thereby maximizing the utilization of limited parameter resources of the model. The model retains complete input parameters while preventing the model overfitting problem caused by input redundancy.

[0075] Based on the sliding window method, the present invention designs a preprocessing method for boiler time series historical data, including window length and label selection, so as to divide the boiler historical data into three types of four-dimensional tensors, and select the most suitable data set type based on CS-CNN. In order to improve the model prediction performance, a comparative test was conducted on LSTM, baseline CNN and CS-CNN using the historical operation data of a 670MW coal-fired boiler. The test results show that CS-CNN has higher prediction performance than traditional deep learning models.

[0076] The average time to predict a boiler NOx emission is about 2.35×10 -2ms, which fully meets the requirements of actual industrial scenarios. Through ablation experiments and model parameter visualization, it can be seen that CNN has the function of automatically extracting features, which is more suitable for complex control systems with variable working conditions and multiple variables such as coal-fired boilers in power plants.

[0077] The types of controlled variables involved in combustion are different under different working conditions. By utilizing the channel selectivity function of CS-CNN, the input variables of the model can retain most of the controlled variables related to burner control, and overfitting caused by input redundancy can be prevented even without using dimensionality reduction methods.

[0078] The data in the tensor is time-continuous, so the model shows high convergence for both steady-state and transient (load lifting) conditions. Compared with traditional deep learning models such as LSTM and baseline CNN, CS-CNN has higher prediction accuracy and meets the requirements of actual industrial scenarios.

[0079] The complete selective convolution layer is the guarantee of CS-CNN prediction performance. Through the visualization analysis of the spatial shift of the input channel, it is found that the closer to the output layer, the more dispersed the spatial shift on the convolution layer, and the more obvious the effect of expanding the convolution kernel. With the help of CNN input channel visualization, it is found that under different coal mill operation combinations, the burner damper related to NOx emissions can realize channel importance identification under the action of CS-CNN. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] Figure 1 Schematic diagram of a NOx prediction model based on CS-CNN according to an embodiment of the present invention, in which P and T represent sensors related to pressure and temperature, and CV represents a control-related valve;

[0081] Figure 2 It is a schematic diagram of the structure of the selective convolutional neural network CS-CNN of the present invention. DETAILED DESCRIPTION

[0082] The embodiments of the present invention are further described below in conjunction with the accompanying drawings:

[0083] like Figure 1 As shown, a method of predicting boiler full-load NOx emissions by using a channel selection convolutional neural network of the present invention obtains the mapping relationship between the controlled variables involved in the combustion in the furnace under the corresponding working conditions and the NOx emissions, which is the key to optimizing combustion and reducing NOx emissions.

[0084] The research was conducted on a coal-fired boiler with a rated load of 670MW. The boiler is a supercritical parameter variable pressure DC furnace with four low NOx emission burner assemblies and a four-corner tangential combustion method. The burner adopts a double-inlet and double-outlet coal mill design. The six coal mills transport coal powder with six layers of coarse coal powder nozzles and six layers of fine coal powder nozzles. In the actual operation of the boiler, the types of control variables involved in combustion change in real time with the change of operating conditions, so the types of variables related to NOx emissions also change immediately. CS-CNN is used to select the importance of input channels and dynamically predict NOx emissions under different operating conditions. The original sensor signals such as pressure and temperature related to boiler operation are collected and stored in DCS or SIS. After extracting the boiler historical data from DCS or SIS, the data is preprocessed and divided into 4-D tensors by the sliding window method. Then the data is input into CS-CNN in the form of tensors for training and output of NOx emission prediction results. Among them, CS-CNN will automatically identify the Top-C channels according to the changes in input operating conditions (including the types of burner dampers caused by changes in the coal mill operation combination), release the hyperparameters of low-importance channels, maximize the utilization of the limited parameter resources of the model, and achieve the purpose of improving the prediction accuracy of the model.

[0085] The present invention provides a method for predicting boiler full-load NOx emissions by using a channel selection convolutional neural network. Since the actual operation of the boiler changes with the working conditions, the types of control variables involved in combustion change in real time, so the types of variables related to NOx emissions also change immediately. The CS-CNN model is used to select the importance of the input channel, and dynamic prediction of NOx emissions under different working conditions is carried out.

[0086] The distributed control system DCS of the power generation boiler continuously uses sensors to extract variable data during boiler operation, including boiler load, total fuel quantity, coal mill capacity air flow, total air volume, secondary air door opening, steam pressure, flow, and temperature;

[0087] Store the collected variable data of boiler operation;

[0088] Select the stored historical time period to extract the historical data of boiler combustion operation as needed, and then pre-process the boiler operation data in the historical data;

[0089] The preprocessed boiler historical data is divided into 4-D tensors by the sliding window method, and then the boiler historical data is input into the channel selective convolutional neural network CS-CNN model in the form of tensors for training. The CS-CNN model will automatically identify the channels ranked in the top 80% of importance according to the operating condition changes of the input boiler, release the hyperparameters of the channels with low importance, and maximize the use of the limited parameter resources of the CS-CNN model to achieve the purpose of improving the prediction accuracy of the CS-CNN model. The operating conditions of the boiler include the operation mode of the boiler burner damper caused by the change of the coal mill operation combination. The CS-CNN model needs to be trained and verified with the preprocessed boiler historical data before use. After the training is completed, the filtered operating conditions can be predicted.

[0090] Among them, the sliding window method is used to divide the data set into several 28×28 two-dimensional matrices, and each two-dimensional matrix is ​​used as the input of CNN for refrigerant charging fault detection (RCFD) prediction. This paper constructs a two-dimensional sliding window with the boiler historical operation data in the form of feature × time series. For the study of setting the time length of the sliding window, the method of processing data sets in the HAR field is used to set the step length of the window sliding according to the 50% time overlap of the data. For the processing of labels, the NOx emissions corresponding to the last time series in the sliding window are used as the label of the single window data. Schematic diagram of data segmentation in the sliding window, the number of features is W, the time length is H, the size of the sliding window: width × height = W × H, the window sliding step is H / 2, that is, the data maintains a 50% repetition rate.

[0091] The collected historical data of combustion boilers are analyzed, and the missing data are replaced by the median of the non-missing values ​​of the variable. The 3σ-rule is used to detect missing data and abnormal data that deviates from the normal value, and the abnormal data are replaced using the following formula:

[0092] Δx i (t) = x i (t)-x i (t-1)

[0093]

[0094]

[0095]

[0096] Where Δx i (t) represents the auxiliary variable x i The rate of change between two adjacent moments is Denotes Δx i (t) the average rate of change before time t, σ i(t) represents the standard deviation, i=1,2,…,m represents the number of variables, and t=2,…,n represents the sampling time.

[0097] The structure diagram of the selective convolutional neural network, CS-CNN, is shown in Figure 2 As shown in the figure, the CS-CNN model includes an input layer, a convolutional layer, a fully connected layer, and an output layer. The input layer imitates the single-channel input form of a gray image. The convolutional layer is used for automatic feature extraction. In CS-CNN, the conventional convolutional layer is replaced by the same number of channel selection convolutional layers. Since the input layer is a single channel, the first convolutional layer still uses a conventional convolutional layer. Usually, a downsampling layer is set after the convolutional layer to reduce the size of the feature map, retain significant features, increase the receptive field, reduce the impact of noise, and make each feature more robust. After the feature map in the last convolutional layer is expanded in one-dimensional form, it will be connected to all neurons in the fully connected layer. Finally, the output layer realizes the continuous prediction of NOx emissions.

[0098] The conventional convolution layer is the basic architecture of CS-CNN, which contains convolution filters, batch normalization and nonlinear activation functions. It extracts the spatial features of the target variable (NOx emissions) from the input tensor by setting multiple convolution filters. The convolution operation on the kth convolution filter in the lth layer is described as follows:

[0099]

[0100]

[0101] Among them, * represents the convolution operator, They represent the kth feature map and feature map element of the lth layer before batch normalization after the convolution operation, respectively. It is worth noting here that the number of convolution filters is equal to the number of output feature maps. represents the weight matrix of the kth convolution filter in the lth layer, X l-1 represents the input matrix of the lth layer, Represents the deviation parameter corresponding to the kth convolution filter in the lth layer. H', W' represent the height and width of the convolution filter respectively.

[0102] Training deep neural networks is very difficult, especially making them converge in a short time. Batch normalization can make training easier. After the input data passes through BN, the distribution of the data is standardized, and the network can adapt to the characteristics of data with different distributions without further learning, thereby improving the generalization ability of the network. Batch normalization of the convolutional layer occurs after the convolution calculation and before the nonlinear activation function. Suppose there are m samples in the mini-batch. On the k-th feature channel of the l-th layer, batch normalization needs to be performed on m×H×W elements in the channel at the same time. H and W represent the height and width of the feature map respectively. The input of batch normalization can be expressed as For any sample in a small batch And 1≤i≤m, the batch normalization calculation process is as follows:

[0103]

[0104]

[0105]

[0106] Among them, μ Ξ , represent the mean and variance of Ξ respectively, represents the output of batch normalization, ζ k and β k They represent the stretching parameters and offset parameters of the kth output channel feature map, and each channel has independent stretching and offset parameters.

[0107] After batch normalization, the nonlinear activation function is applied to all feature maps. Existing experimental results show that when the input is greater than 0, the ReLU function converges 6-7 times faster than the sigmoid function and the tanh hyperbolic tangent function due to gradient preservation. Therefore, this paper uses the ReLU function as the nonlinear activation function, which is described as follows:

[0108]

[0109] Selective Convolutional Layer

[0110] In CS-CNN, channel-selective convolution operations include De-alloc channel deallocation, Re-alloc channel reallocation, and spatial shift. Figure 2 As shown in the figure, the channel selective convolution layer De-alloc and Re-alloc process is demonstrated at the bottom of the figure. CS-CNN completes the training task through the following two steps to achieve the purpose of reallocating the input channels:

[0111] (1).De-alloc: Based on the input conditions, prevent boiler variables with low importance from being used in subsequent calculations and release the parameters of these channels.

[0112] (2).Re-alloc: Use the Top-C important channels to replace the blocked channels.

[0113] After the model is trained multiple times in the above steps, important channels are retained. However, since convolution is a linear operation, multiple trainings do not bring any benefits to the channels: if two input channels are exactly the same, it will cause weight degradation. Therefore, we designed spatial shift to offset the diversity loss caused by too many similar channels.

[0114] A. Channel importance identification indicators

[0115] The expected channel corruption matrix (ECDM) provides an efficient and safe way to identify channels that contribute less to the output. Assume that Conv(X,ω) represents a regular convolution operation, X represents the input variable of the convolution layer, and ω represents the weight matrix of the convolutional layer, and Among them, I and O represent the number of input and output channels, respectively, and H and W represent the height and width of the input channel, respectively. Since the boiler historical data are arranged in chronological order, setting the convolution kernel size needs to consider the feature extraction at the same time, so the convolution kernel size is set to K l ×1. And i=1,2,...,I,h=1,2,...,H,w=1,2,...,W. Define ECDM(X;ω) to calculate the expected mean in the time dimension, as shown in formula (7):

[0116]

[0117] When i = 0, it means that the ith channel has been set as a low-importance channel and needs to be blocked. The expected channel impairment matrix (ECDM) thus identifies the importance of the input channel.

[0118] B. Defining channel selection and spatial shift

[0119] CS-CL is defined as SelectConv(X,ω), which implements the selection function of the input channel X based on the conventional convolution layer, so the selective convolution layer can be expressed as:

[0120] SelectConv(X;ω)=Conv(SelectChannel(X);ω)

[0121] SelectChannel needs to perform channel blocking and re-indexing for De-alloc and Re-alloc respectively. Therefore, the gate variable g is introduced. i ∈{0,1} and index π i ∈{1,2,…,I}. The input channel X can be expressed as Here if Does not contribute to the output, the corresponding g i =0; if multiple π i If the channels are identical, it means that a channel is replicated multiple times. To prevent the serious consequence of weight degradation caused by identical channels, inspired by the literature [35,36], a spatial offset is imposed on the reallocated channels. The SelectChannel description is as follows:

[0122]

[0123] i={1,2,…,I}

[0124] π i ={1,2,…,I}

[0125] g i ∈{0,1}

[0126]

[0127] Among them, shift(X,b) represents the spatial shift operation of the input channel X. For each element (x,y) on the channel, shift(X,b) x,y Defined as:

[0128]

[0129]

[0130] Among them, the parameter b is a continuous value and is learned jointly with other hyperparameters through SGD.

[0131] C. Channel-Selective Convolutional Layer Training Scheme

[0132] In order to use ECDM to identify the de-allocated / reallocated channels, the parameters S = (X; g, π, b) are cleverly designed in the channel-selective convolutional layer, and then S is iteratively updated during the de-alloc and re-alloc training process. i = 0. Given the expected damage level γ>0, the unlocking goal can be written as the following optimization problem:

[0133]

[0134]

[0135] g i =0 or 1; i=1,2,…,I

[0136] Calculate using greedy algorithm: Normalize ECDM according to the output dimension (nECDM):

[0137]

[0138] Assuming j = 1, 2, ..., O, when nECDM is calculated, the minimum channel ||nECDM(ω; X) i || ∞Iterate to determine the channel to be released, but ensure that nECDM(ω;X) i Vector sum of l ∞ norm is less than γ.

[0139] For Re-alloc, according to nECDM's l 2 The norm selects the top-C maximum channel, i.e. ||nECDM(ω;X) i ||2. The first C channels selected randomly occupy the unassigned channels (g i =0). When the i-th channel is de-assigned, the corresponding parameter in ω is set to zero so that the operation does not damage the training. In addition, a maximum reallocation count N is set max , to limit redundant allocations to the same feature map. Finally, the dynamic training scheme is based on the general stochastic gradient descent (SGD). When training ω via SGD, deallocation and reallocation will update other parameters of S: deallocation is used to update g, while reallocation is used to update b, g, π, and ω.

[0140] Fully connected layer and output layer

[0141] One or more fully connected layers are used after the convolutional layer. The purpose is to perform high-level reasoning; connect all neurons in the previous layer to each neuron in the current layer, and then generate global semantic information. A fully connected layer is set after the 6th convolutional layer, and the mapping relationship is as follows:

[0142]

[0143] Where f(·) represents the ReLU nonlinear activation function, represents the output value of the jth node in the lth layer, represents the weight between the jth node in the lth layer and the ith node in the previous layer, represents the input value of the i-th node in the l-th layer, represents the bias of the lth layer.

[0144] The loss function and activation function in the output layer vary depending on the classification and regression tasks. For classification tasks, the cross entropy loss function and softmax nonlinear activation function are used. For a model with m batch samples and t output categories, the loss function and activation function are as follows:

[0145]

[0146]

[0147] In regression tasks, the mean square error loss (MSE) function and the linear activation function are generally used. Taking the single output prediction model as an example, the loss function is as follows:

[0148]

[0149] Among them, y i,j and i is the true value of the i-th sample, y i,j and i is the predicted value of the sample. x For regression tasks such as emission prediction, this paper adopts the mean square error loss function and linear activation function.

Claims

1. A method for predicting boiler full load NOx emissions using a channel selection convolutional neural network, characterized in that: Since the actual operation of the boiler changes with the working conditions, the types of control variables involved in combustion change in real time, so the types of variables related to NOx emissions also change immediately. The CS-CNN model is used to select the importance of the input channel and dynamically predict NOx emissions under different working conditions. Here are the steps: The distributed control system of the power generation boiler continuously uses sensors to extract variable data during boiler operation, including boiler load, total fuel quantity, coal mill capacity air flow, total air volume, secondary air door opening, steam pressure, flow, and temperature; The collected variable data of the boiler during operation is stored; Select the stored historical time period to extract the historical data of boiler combustion operation as needed, and then pre-process the boiler operation data in the historical data; The preprocessed boiler historical data is divided into 4-D tensors by the sliding window method, and then the boiler historical data is input into the channel selective convolutional neural network CS-CNN model in the form of tensors for training. The CS-CNN model will automatically identify the channels ranked in the top 80% of importance according to the operating condition changes of the input boiler, release the hyperparameters of the channels with low importance, and maximize the use of the limited parameter resources of the CS-CNN model to achieve the purpose of improving the prediction accuracy of the CS-CNN model. The operating conditions include the operation mode of the boiler burner damper caused by the changes in the coal mill operation combination; Before using the CS-CNN model, it is necessary to use the pre-processed historical data of the boiler for training and verification. After the training is completed, the operating conditions of the boiler can be predicted. The CS-CNN model structure is: sequentially connected input layer, convolution layer, fully connected layer and output layer; the convolution layer contains sequentially connected convolution filters, batch normalization and nonlinear activation functions, and extracts the spatial features of NOx emissions as the target variable from the input tensor by setting multiple convolution filters; there are one or more fully connected layers to perform high-level thrust, connect all neurons in the previous layer to each neuron in the current layer, and then generate global semantic information; the output layer includes loss function and activation function, which are different according to classification and regression tasks. For classification tasks, cross entropy loss function and softmax nonlinear activation function are used; Among them, the input layer imitates the single-channel input form of the gray image, and the convolution layer is used for automatic feature extraction. In CS-CNN, the conventional convolution layer is replaced by the same number of channel selection convolution layers; the first convolution layer still uses the conventional convolution layer, and a downsampling layer is set after the convolution layer to reduce the size of the feature map, retain significant features, increase the receptive field, reduce the impact of noise, and make each feature more robust. After the feature map in the last convolution layer is expanded in one-dimensional form, it will be connected to all neurons in the fully connected layer. Finally, the output layer realizes the continuous prediction of NOx emissions; The basic architecture of CS-CNN is a conventional convolutional layer, which includes convolutional filters, batch normalization, and nonlinear activation functions. Multiple convolutional filters are set to extract NOx emissions from the input tensor as the spatial features of the target variable. The convolution operation on the kth convolutional filter in the lth layer is as follows: Among them, * represents the convolution operator, They represent the kth feature map and feature map element of the lth layer before batch normalization after the convolution operation. It is worth noting that the number of convolution filters is equal to the number of output feature maps. represents the weight matrix of the kth convolution filter in the lth layer, X l-1 represents the input matrix of the lth layer, It represents the deviation parameter corresponding to the k-th convolution filter of the l-th layer, H' and W' represent the height and width of the convolution filter respectively; Improving the convergence efficiency of the CS-CNN model through batch normalization: After batch normalization BN, the input data is distributed in a standardized way, which can adapt to data characteristics of different distributions without further learning, and improve the generalization ability of the network. Batch normalization is performed on the convolution layer between batch normalization and nonlinear activation function: batch normalization of the convolution layer occurs after convolution calculation and before nonlinear activation function; Assume that there are m samples in the batch input data. On the kth feature channel of the lth layer, batch normalization is performed on m×H×W elements in the feature channel at the same time. H and W represent the height and width of the feature map respectively. The input of batch normalization is expressed as For any sample in batch Ξ And 1≤i≤m, the batch normalization calculation process is as follows: in, represent the mean and variance of Ξ respectively, represents the output of batch normalization, ζ k and β k Represent the stretching parameter and offset parameter of the kth output channel respectively, and each channel has independent stretching and offset parameters; After batch normalization, the nonlinear activation function will be applied to all feature maps, and the ReLU function is used as the nonlinear activation function, which is described as follows:

2. The method for predicting boiler full load NOx emissions using a channel selection convolutional neural network according to claim 1, characterized in that: The data preprocessing method is to use the 3σ rule to detect missing data and abnormal data that deviates from normal values, and replace the abnormal data using the following formula: Δx i (t)=x i (t)-x i (t-1) Where Δx i (t) represents the auxiliary variable x i The rate of change between two adjacent moments is Denotes Δx i (t) the average rate of change before time t, σ i (t) represents the standard deviation, i=1,2,…,m represents the number of variables, and t=2,…,n represents the sampling time.

3. The method for predicting boiler full load NOx emissions using a channel selection convolutional neural network according to claim 1, characterized in that: In the CS-CNN model, channel-selective convolution operations include De-alloc channel deallocation, Re-alloc channel reallocation, and spatial shifting. Spatial shifting is used to offset the diversity loss caused by too many similar channels. The CS-CNN model completes the training task and realizes the purpose of reallocating the input channels: 1). De-alloc channel: for input conditions, prevent boiler variables with low importance from being used in subsequent calculations and release the parameters of these channels; 2). Re-alloc channel: Use the Top-C important channels to replace the blocked channels; The important channels of the CS-CNN model are retained after multiple trainings; Spatial shifting is used to offset the loss of diversity caused by too many similar channels.

4. The method for predicting boiler full load NOx emissions using a channel selection convolutional neural network according to claim 3 is characterized in that: The expected channel damage matrix ECDM is used to identify the channels that contribute less to the output and the importance of the input channels. Since the boiler historical data are arranged in chronological order, setting the convolution kernel size needs to consider the feature extraction at the same time. Therefore, let Conv(X,ω) represent the conventional convolution operation, X represents the input variable of the convolution layer, and ω represents the weight matrix of the convolutional layer, and IN and O represent the number of input and output channels, respectively. H and W represent the height and width of the input feature map, the feature map equivalent channel, and the product kernel size is K. l ×1, and in=1,2,...,IN, h=1,2,...,H, w=1,2,...,W; Define ECDM(X;ω) to calculate the expected mean in the time dimension, as shown below: When in =0, it indicates that the in-th channel has been set as a low-importance channel and needs to be blocked, and the expected channel damage matrix ECDM thereby identifies the importance of the input channel.

5. The method for predicting boiler full load NOx emissions using a channel selection convolutional neural network according to claim 4 is characterized in that: Define channel selection and spatial shift: The channel selection convolution layer is defined as SelectConv(X,ω), which implements the selection function of the input channel X based on the conventional convolution layer. The selective convolution layer is expressed as: SelectConv(X;ω)=Conv(SelectChannel(X);ω) The channel SelectChannel needs to perform channel blocking and re-indexing for De-alloc and Re-alloc respectively; the gate variable g is introduced in ∈{0,1} and index π in ∈{1,2,…,IN}, the input channel X is represented as if Does not contribute to the output, the corresponding g in =0; if multiple π in If the channels are identical, it means that a channel has been replicated multiple times. To prevent the serious consequence of weight degradation caused by identical channels, a spatial offset is applied to the reallocated channels. At this time, the description of SelectChannel is as follows: in={1,2,…,IN} π in ={1,2,…,IN} g in ∈{0,1} Among them, shift(X,b) represents the spatial shift operation of the input channel X. For each element (x,y) on the channel, shift(X,b) x,y Defined as: Among them, parameter b is the bias parameter, which is jointly learned with other hyperparameters through stochastic gradient descent SGD, and through similar spatial shift operation, efficient parameter utilization is achieved. By recycling its parameters to selectively expand the kernel size, the number of parameters used is reduced while deepening the model network.

6. The method for predicting boiler full load NOx emissions using a channel selection convolutional neural network according to claim 5, characterized in that: The channel-selective convolutional layer training scheme is: In order to use ECDM to identify the canceled / reallocated channels, the parameters S = (X; g, π, b) are designed in the channel-selective convolutional layer, and then S is iteratively updated during the de-alloc and re-alloc training process; by setting g in = 0 to complete the de-alloc channel. Given the expected damage level γ>0, the de-allocating objective is rewritten as the following optimization problem: g in =0or 1;in=1,2,...,IN If the above formula is used in the complete enumeration method, we will get 2 IN combinations; therefore, when IN is large, a greedy algorithm is used to calculate: ECDM is normalized according to the output dimension nECDM: Where, let j = 1, 2, ..., O. After nECDM is calculated, the minimum channel ||nECDM(ω; X) in || ∞ Iterate to determine the channel to be released, but ensure that nECDM(ω;X) in Vector sum of l ∞ The norm is less than γ; Re-alloc, according to nECDM's l 2 The norm selects the top-C largest channel: ||nECDM(ω;X) in ||2; The first C channels selected randomly occupy the unassigned channel g in = 0; when the in-th channel is de-assigned, the corresponding parameter in ω is set to zero so that the operation does not damage the training; set the maximum reallocation count N max , to limit the redundant allocation of the same feature map; dynamic training is performed using the general stochastic gradient descent method SGD. When training ω through SGD, unassignment and reallocation will update other parameters of S: Unassignment is used to update the entry variable g in , and reallocation Re-alloc is used to update the deviation parameter b and the entry variable g in , π in and the weight matrix ω.

7. The method for predicting boiler full load NOx emissions using a channel selection convolutional neural network according to claim 6, characterized in that: The fully connected layer and output layer of the CNN model are as follows: one or more fully connected layers are used after the convolutional layer to perform high-level reasoning, connect all neurons in the previous layer to each neuron in the current layer, and then generate global semantic information, and set up the fully connected layer. The mapping relationship is as follows: Where f(·) represents the ReLU nonlinear activation function, represents the output value of the jth node in the lth layer, represents the weight between the jth node in the lth layer and the ith node in the previous layer, represents the input value of the i-th node in the l-th layer, represents the deviation of the lth layer; The loss function and activation function in the output layer vary depending on the classification and regression tasks. For classification tasks, the cross entropy loss function and the softmax nonlinear activation function are used. For a model with m batch samples and t output categories, the loss function and activation function are as follows: In the regression task, the mean square error loss MSE function and the linear activation function are used. The loss function of the single output prediction model is as follows: Among them, y i,j and i is the true value of the i-th sample, and is the predicted value of the sample; the NOx emission prediction regression task adopts the mean square error loss function and linear activation function.

Images