NOx Emission Prediction Method for Coal-Fired Boilers Based on Modulated Deformable Neural Network

Abstract

The present invention discloses a method for predicting NOx emissions of coal-fired boilers based on a modulated deformable neural network, belonging to the technical field of boiler operation optimization control. The method includes the following steps: extracting historical operation data from the coal-fired boiler monitoring system, processing abnormal data using the 3σ rule, and performing data standardization through the min-max normalization method; evaluating variable importance based on the Gini coefficient and screening 27 key input variables; constructing a modulated deformable convolutional neural network model, and by introducing learnable offsets and modulation factors, adaptively adjusting the sampling positions of convolutional kernels and feature weights to enhance the model's ability to extract complex working condition features; the model includes 3 convolutional layers and fully connected layers, and is trained using the mean square error loss function and the Adam optimizer. The present invention provides an efficient solution for the accurate prediction of NOx emissions of coal-fired boilers.

Description

Technical Field

[0001] The present invention belongs to the technical field of boiler operation optimization control, and particularly relates to a method for predicting NOx emissions of a coal-fired boiler based on a modulated deformable neural network, which is applicable to the prediction of NOx emissions of a coal-fired boiler. Background Technique

[0002] Coal-fired power stations play a key role in industrial production, heating, power generation and other fields. With the improvement of environmental awareness and the gradual tightening of emission policies, coal-fired boilers are facing the pressure of transformation and upgrading. Nitrogen oxides (NOx) are generated during the operation of coal-fired boilers, so the clean treatment of flue gas has become the key to controlling pollutant over-standard. When the SCR flue gas denitrification technology operates, it is necessary to accurately control the injection amount of NH3, so it is extremely crucial to accurately measure the NOx content at the SCR inlet.

[0003] Data-driven methods provide better opportunities for establishing boiler combustion models. By using a large amount of data collected by a distributed control system (DCS) and combining various algorithms to analyze and process the data, a digital boiler can be established. For example, using the differential evolution (DE) algorithm to construct DELSSVM has achieved good results in NOx emission prediction; the GA-LSSVM model combined with the adaptive GA and the least squares support vector machine (LSSVM) has also been used to accurately and quickly predict the NOx concentration; using a comparative linear transfer function-noise model and a neural network to simulate the boiler fouling condition, etc. However, the above models are only applicable to the steady-state conditions of boilers, and the prediction accuracy for variable load conditions is relatively low. Deep learning models have a high representation ability for complex processes and have shown good effects in dealing with multi-variable coupling and non-linear data. For example, using the global enhanced general regression neural network (GE-GRNN) to solve the problem of multiple emission predictions of an on-line boiler, using a generative adversarial network (GAN) to repair industrial boiler data, using a recurrent neural network (RNN) and a gated recurrent unit (GRU) to simulate the relationship between operating parameters and NOx emissions; the long short-term memory neutral network (IPSO-LSTM) realizes the optimization of on-line combustion of boilers.

[0004] Compared with other neural networks, CNN adds the element of time series, enabling it to effectively improve the simulation deviation of the model caused by time factors, automatically extract stable features of data, and greatly reduce the model parameters and computational complexity by using parameter sharing and sparse connection methods. It can handle large-scale data well while avoiding the problem of overfitting, reduce the workload of manual feature extraction, and adapt to different application scenarios. However, conventional CNN processes information through local perception, thus being unable to capture global features well. Its input data has fixed-size format requirements, and excessive data preprocessing may lead to information loss or distortion. In the operation of a boiler, different data parameters have different contributions to the normal operation of the boiler. If the differences are not distinguished, a large amount of noise will inevitably affect the normal prediction accuracy. Secondly, the scales of different data parameters also vary, which results in the inability of conventional CNN to highly adapt to the work of boiler prediction. In conventional CNN, the same convolutional layer uses the same-sized receptive field, assuming that the input features are spatially invariant, and unable to consider the important weight dependencies between elements. As long as some key features of an object exist in the test data, the test data will be classified as that object. Additionally, the structure of conventional CNN is relatively fixed, which makes it difficult to process complex and irregular input data, greatly limiting the modeling ability of CNN. Summary of the Invention

[0005] The object of the present invention is a method for improving the boiler combustion process, which improves the model based on conventional CNN. The improved model adaptively changes according to the specific situation of the data, adapts to the changes in the input data by changing the size of the receptive field and changing the modulation factor. Compared with traditional CNN, the present invention adds a variable of offset to the conventional convolutional neural network and enables it to be adaptively adjusted during the training process to achieve the expansion of the receptive field. At the same time, based on the characteristics of the input data, the modulation factor is changed to make the prediction result as close as possible to the actual operation of the boiler.

[0006] The present invention proposes a method for predicting NOx emissions from a coal-fired boiler based on a modulated deformable neural network, including the following steps:

[0007] Step 1, screening key input variables related to NOx emissions;

[0008] Step 2, constructing a modulated deformable convolutional neural network DFC-CNN model for predicting NOx emissions;

[0009] Step 3, constructing a data set for training the modulated deformable convolutional neural network DFC-CNN model to obtain a trained DFC-CNN model;

[0010] The data set includes key input variables and corresponding NOx emission amounts;

[0011] Step 4: Input the key input variables to be predicted into the trained DFC-CNN model for NOx emission prediction.

[0012] Further, in step 1, screening the key input variables related to NOx emissions specifically involves: evaluating the feature importance of variables for boiler NOx emissions based on the Gini index of the random forest RF (Random Forest).

[0013] Further, let the number of decision trees be n, and the importance of the j var th feature in the i tree th decision tree is:

[0014] (3)

[0015] where represents the importance contribution of the j var th feature calculated through the Gini index at index m, and M represents the set of indices;

[0016] The importance of the j var th feature on all decision trees is:

[0017] (4)

[0018] Normalize the importance of the j var th feature on all decision trees to obtain the normalized importance of the j var th feature on all decision trees , expressed as:

[0019] (5)

[0020]

[0021] where J represents the total number of all NOx-related feature variables;

[0022] If the normalized importance of the j var th feature on all decision trees is greater than or equal to the set value, then the j var th feature is a key input variable.

[0023] Further, the key input variables include feed water flow rate, SOFA-D#1 angle, main steam flow rate, unit load, FR#1 angle, E mill E1 side volumetric air flow rate, EL#1 angle, DE#1 angle, main steam temperature, DL#1 angle, flue gas oxygen content A, exhaust gas temperature A, CL#1 angle, SOFA-E#1 angle, DEE#1 angle, B mill B1 side volumetric air flow rate, A mill A2 side volumetric air flow rate, inlet air temperature of B forced draft fan, SOFA-A#1 angle, SOFA-F#1 angle, total coal quantity, EFF#1 angle, SOFA-B#1 angle, AAA layer#1 angle, AA#1 angle, B mill B2 side volumetric air flow rate, ER#1 angle. A total of 27 variables are used as the input variables for modulating the deformable convolutional neural network DFC-CNN model.

[0024] Further, in the deformable convolutional neural network DFC-CNN model, the features input to the DFC-CNN model Figure X are subjected to offset and modulation processing to obtain the modulated output feature map , and then the eigenvalues in the feature map are successively processed through multiple convolutional layers to obtain the eigenvalue Xfcl output by the convolutional layer. The eigenvalue Xfcl is processed through a fully connected layer to obtain the predicted value of NOx, which is the output value of the deformable convolutional neural network DFC-CNN model;

[0025] The eigenvalue at each position (i, j) in the modulated output feature map is expressed as:

[0026] (10)

[0027] where u and v are the indices of the convolutional kernel, is the weight of the convolutional kernel, is the modulation factor, is the offset;

[0028] X fcl is a four-dimensional tensor. In the fully connected layer, first flatten X fcl . The flattened vector is shown in Equation (11), and then weighted summation is performed as shown in Equation (12):

[0029] (11)

[0030] (12)

[0031] where vec(.) is the operation of converting a matrix or tensor into a vector, represents the one-dimensional tensor obtained after flattening the four-dimensional tensor, Represents the weight matrix of the fully connected layer, represents the bias term of the fully connected layer, represents the result after linear transformation, and finally passes through the activation function to obtain the output target , that is, the predicted value of NOx emissions from coal-fired boilers , which is expressed as:

[0032] (13)

[0033] Furthermore, when training the modulated deformable convolutional neural network DFC-CNN model, the key input variables and NOx emissions are reconstructed into a two-dimensional time series tensor, and the sliding window is used to segment the two-dimensional time series tensor in the time dimension, and the sliding step is set to 1.

[0034] Furthermore, the loss function of the DFC-CNN model is the mean square error MSE, and the activation function is the Sigmoid function.

[0035] Furthermore, the construction of the data set in step 3 is obtained by extracting historical operation data in the coal-fired boiler monitoring system and through preprocessing and normalization.

[0036] Furthermore, the preprocessing refers to processing the abnormal data in the historical operation data of the boiler using the 3σ rule. Since the monitoring system will be interfered by signal noise during the operation of the boiler, resulting in abnormal original data, processing the abnormal data using the 3σ rule can obtain the historical data after processing the abnormal data;

[0037] To improve the accuracy of the model and avoid large errors in the prediction results, the minimum-maximum scaling method is used to normalize the historical data after processing the abnormal data to obtain the normalized historical data.

[0038] Furthermore, the Adam optimizer is used to train the DFC-CNN model.

[0039] Compared with the prior art, the present invention focuses on improving the conventional CNN on the basis of simulation experiments, and uses the Gini index and the modulated deformable convolutional neural network DFC-CNN to establish a prediction model. The Gini index is used to evaluate the importance of candidate variables and provides key input features for the model; the modulated deformable convolutional neural network DFC-CNN improves the ability of the model to extract implicit features in the input data through the modulated deformable receptive field. The present invention can meet the requirements of actual industrial production, accurately predict NOx emissions, and contribute to the clean combustion of coal-fired boilers. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1It is the structural diagram of the modulated deformable convolution model;

[0041] Figure 2 It is the variable importance distribution diagram;

[0042] Figure 3 It is the schematic diagram of data segmentation by the sliding window method;

[0043] Figure 4 It is the training loss diagram at different learning rates in the embodiment;

[0044] Figure 5 It is the scatter diagram of BP, LSTM, and DFC-CNN processing the validation set in the embodiment;

[0045] Figure 6 It is the performance parameter comparison diagram of BP, LSTM, and DFC-CNN in the embodiment;

[0046] Figure 7 It is the load increase comparison curve and scatter diagram of Baseline-CNN and DFC-CNN in the embodiment;

[0047] Figure 8 It is the load decrease comparison curve and scatter diagram of Baseline-CNN and DFC-CNN in the embodiment;

[0048] Figure 9 It is the variable load comparison curve and scatter diagram of Baseline-CNN and DFC-CNN in the embodiment;

[0049] Figure 10 It is the stable load comparison curve and scatter diagram of Baseline-CNN and DFC-CNN in the embodiment. Detailed implementation manners

[0050] To deepen the understanding of the present invention, the present invention will be further described in detail below in conjunction with embodiments. These embodiments are only used to explain the present invention and do not limit the protection scope of the present invention.

[0051] A method for predicting NOx emissions of a coal-fired boiler based on a modulated deformable neural network uses a modulated deformable convolutional neural network DFC-CNN to establish a prediction model. The Gini index is used to evaluate the importance of candidate variables and provide key input features for the prediction model. The modulated deformable convolutional neural network DFC-CNN improves the model's ability to extract implicit features in input data through a modulated deformable receptive field. To verify the prediction performance of the modulated deformable convolutional neural network DFC-CNN, simulation experiments were conducted on the real historical data of a 600MW coal-fired boiler. Comparative experiments between the modulated deformable convolutional neural network DFC-CNN and BP, LSTM show that the modulated deformable convolutional neural network DFC-CNN has a higher prediction accuracy than traditional models. The comparison results with the baseline CNN show that the expansion of the receptive field and the introduction of the modulation factor significantly improve the prediction performance of the modulated deformable convolutional neural network DFC-CNN. The method based on GINI-DFC-CNN can meet the needs of actual industrial production and contribute to the clean combustion of coal-fired boilers. The method of the present invention includes the following steps:

[0052] Step 1: Extract part of the historical data of the target boiler from the boiler monitoring system. The sampling frequency is 1 minute. The target boiler uses a Selective Catalytic Reduction (SCR) denitration system. The input amount of ammonia greatly affects the treatment efficiency of nitrogen oxides. Excessive NH3 will react with SO3 in the flue gas to form liquid NH4HSO4, poisoning the catalyst, clogging the air preheater, and affecting the safe operation of the boiler. Too little NH3 cannot eliminate NOx. The burners in the boiler adopt a multi-compartment auxiliary air design. Part of the secondary air flow is staged horizontally, wrapping the primary air pulverized coal flow in the center of the furnace to form a fuel-rich area. The lowest part of the main burner uses an Under-fire Air (UFA) nozzle, and Closed-Coupled Over-Fire Air (CCOFA) and Separated Over-Fire Air (SOFA) are used to achieve multi-stage control of the combustion area. Each layer of air chamber of the burner is equipped with corresponding secondary air damper plates. Generally, the damper plates of the same layer act synchronously. The R layer is the rich powder fuel air, and the L layer is the lean powder fuel air. The opening degrees of these damper plates are functions of the primary air volume of the coal mill, the primary air volume of the coal mill is a function of the coal mill load, and the secondary air damper plates of SOFA and CCOFA are functions of the total boiler air flow, mainly used to control the NOx emissions. The input amount of NH3 depends on the concentration of NOx. At the same time, there is a lag in the monitoring process, and the monitoring system will be interfered by signal noise during the operation of the boiler, resulting in anomalies in the original data. The 3σ rule is used to process the abnormal data to obtain the historical data after processing the abnormal data, so that it will not have a negative impact on the prediction results of the prediction model of the present invention, and the processed data is used to replace the original data.

[0053] Step 2: To improve the accuracy of the prediction model and avoid large errors in the prediction results, the min-max scaling method is used to normalize the historical data after processing the abnormal data, in order to obtain a more stable and accurate prediction result.

[0054] Step 3: Evaluate the feature importance of variables for boiler NOx emissions based on the Gini index of the Random Forest (RF). The importance of variables is represented by the Variable Importance Measure (VIM), and the Gini index is represented by GI, as shown in the following formula:

[0055] (1)

[0056] (2)

[0057] where k c represents the number of types contained in the dataset of a certain variable after normalization. p mkThe probability that the variable feature with index m is classified into category k c GI m represents the Gini index of the variable feature with index m, represents the Gini index of the left child node after splitting, GI r represents the Gini index of the right child node after splitting, represents the importance contribution of the j var -th feature calculated by the Gini index at index m.

[0058] Let the number of decision trees be n, and the importance of the j var -th feature in the i tree -th decision tree is as follows:

[0059] (3)

[0060] where, represents the importance contribution of the j var -th feature calculated by the Gini index at index m, and M represents the set of indices;

[0061] The importance of the j var -th feature on all decision trees is as follows:

[0062] (4)

[0063] Normalize the importance of the j var -th feature on all decision trees to obtain the normalized importance of the j var -th feature on all decision trees , denoted as:

[0064] (5)

[0065] ;

[0066] where J represents the total number of all NOx-related feature variables;

[0067] If the importance of the j var -th feature on all decision trees after normalization is greater than or equal to the set value, then the j var -th feature is a key input variable.

[0068] The feature importance obtained according to the above method , with 0.9 as the set value, a total of key input variables were selected from multiple initial variables as the input variables of the prediction model, including feed water flow rate, SOFA-D #1 corner, main steam flow rate, unit load, FR #1 corner, E mill E1 side capacity air flow rate, EL #1 corner, DE #1 corner, main steam temperature, DL #1 corner, flue gas oxygen content A, exhaust gas temperature A, CL #1 corner, SOFA-E #1 corner, DEE #1 corner, B mill B1 side capacity air flow rate, A mill A2 side capacity air flow rate, B forced draft fan inlet air temperature, SOFA-A #1 corner, SOFA-F #1 corner, total coal quantity, EFF #1 corner, SOFA-B #1 corner, AAA layer #1 corner, AA #1 corner, B mill B2 side capacity air flow rate, ER #1 corner 27 variables, as Figure 2 shown. According to the characteristics related to NOx emissions, the 27 selected input variables and NOx emissions were reconstructed into a two-dimensional tensor. The input variables and NOx emissions were segmented in the time dimension using a sliding window. The time dimension of the sliding window was 20, the variable dimension was 27, and the sliding step was set to 1. The prediction target was the boiler nitrogen oxide emissions at the 21st moment. The schematic diagram of data reconstruction is as Figure 3 shown. 70% of the entire dataset was divided as the training set, and the remaining 30% was used as the test set to verify the simulation results with the test dataset.

[0069] Step 4: Write each input data after being processed in Step 1, Step 2, and Step 3 into the form of a feature matrix to obtain the input features Figure X , expressed as:

[0070] (6)

[0071] Perform data processing on the said input features Figure X to obtain the modulated input feature map;

[0072] The said data processing refers to obtaining the offset data matrix by sampling point offset, and processing the offset data matrix through the convolution kernel and modulation eigenvalue to obtain the modulated input feature map , and passing the modulated input feature map through multiple convolutional layers for processing and combining by the fully connected layer to predict the NOx emissions; the specific steps are as follows:

[0073] Step 4.1, in the modulated deformable convolution, the modulated deformable convolution performs the process of offset and modulation eigenvalue processing on the basis of the conventional deformable convolution; an offset (offset) is introduced into the sampling position of the convolution kernel, enabling the convolution kernel to sample more flexibly on the input feature matrix, and the extracted features become more obvious.

[0074] For the modulated deformable convolution in the prior art, the convolution kernel slides on the input feature map at fixed intervals and in a fixed direction. The operation formula of the convolution kernel when processing the feature matrix is as follows:

[0075] (7)

[0076] Wherein, is the element at position (i, j) in the output feature map Y after being processed by the convolution kernel, b is the bias, is the element of the c-th convolution kernel at position (m, n), is an element in the data set X and is the element of the input data I at the c-th channel at position (i + m, j + n), represents the height of the convolution kernel, represents the width of the convolution kernel.

[0077] In this specific embodiment, the modulated deformable convolution adds offsets on the basis of the prior art. These offsets are learnable, and they enable the convolution kernel to adaptively adjust the sampling position of the convolution kernel according to the position of the input features.

[0078] Let X be the input feature map input to the modulated deformable convolution, Y be the output feature map of the modulated deformable convolution, W be the weight of the convolution kernel, be the offset, and M be the modulation factor map. Then:

[0079] (8)

[0080] Wherein, is a convolution operation.

[0081] (9)

[0082] Wherein, is also a convolution operation. Then for each position (i, j) of the output feature map Y, the eigenvalue is:

[0083] (10)

[0084] Wherein, u and v are the indices of the convolution kernel, is the weight of the convolution kernel, is the modulation factor, is the offset, is calculated using bilinear interpolation because may produce non-integer coordinates.

[0085] Step 4.2, as Figure 1As shown, there are 3 layers in the multi-layer convolutional layer in this specific embodiment, including the first convolutional layer, the second convolutional layer, and the third convolutional layer; all use conventional convolutional layers, including convolutional kernels, batch normalization (BN), and non-linear activation functions, and perform feature extraction on the input feature map through the convolutional kernels to improve the classification and recognition accuracy of the prediction model of the present invention.

[0086] The non-linear activation function after the BN layer is a Sigmoid function, which is used to compress any real number input into the output range of (0, 1). This non-linear transformation is the key for the neural network to fit complex functional relationships. By stacking multiple neuron layers with non-linear activation functions, the neural network can approximate complex non-linear functions.

[0087] The eigenvalue After being processed by the first convolutional layer, the second convolutional layer, and the third convolutional layer in sequence, all the eigenvalues X output by the convolutional layer are obtained fcl ; in the convolutional layer, the input feature matrix performs local connection and convolutional operations by the convolutional kernel to generate a feature map.

[0088] Step 4.3, the fully connected layer receives all the eigenvalues from the previous convolutional layer as the input X fcl and, through weighted summation and activation function processing, comprehensively maps the features to the target output space to extract global features. The specific steps are as follows:

[0089] For the input sample X of the fully connected layer fcl , X fcl is a four-dimensional tensor. First, it is flattened. The flattened vector is shown in Equation (11), and then weighted summation is performed, as shown in Equation (12):

[0090] (11)

[0091] (12)

[0092] where, vec(.) is an operation to convert a matrix or tensor into a vector, represents the one-dimensional tensor obtained after flattening the four-dimensional tensor, represents the weight matrix of the fully connected layer, represents the bias term of the fully connected layer, represents the result after linear transformation, and finally the output target is obtained through the activation function , that is, the predicted value of the NOx emission of the coal-fired boiler , which is expressed as:

[0093] (13)

[0094] In the regression task, the mean squared error (MSE) loss function and linear activation function are used, and the calculation formula is as follows:

[0095] (14)

[0096] where is the mean squared error loss, N is the total number of samples, is the predicted value of NOx emissions for the i-th sample, represents the true value of NOx emissions corresponding to the i-th sample.

[0097] Training a deep neural network is extremely challenging, especially in the pursuit of fast convergence. Therefore, batch normalization (BN) is adopted during the training of the modulated deformable convolutional neural network prediction model. Batch normalization (BN) can not only make the input distribution of each layer of the modulated deformable convolutional neural network more stable by normalizing the input of each mini-batch, accelerating the training speed, but also stabilize the input distribution and gradient propagation, enabling the model to withstand a higher learning rate.

[0098] All the experiments in this example were conducted under the Pytorch DL framework. All experiments used a server equipped with a 24GB NVIDIA Tesla GPU, 192GB of memory, and an Intel Xeon Gold 6226R CPU. To evaluate the prediction performance of the modulated deformable convolutional neural network DFC-CNN, two sets of comparative experiments were carried out. One was to compare the modulated deformable convolutional neural network DFC-CNN with the backpropagation neural network BP and the long short-term memory network LSTM; the other was to compare it with the baseline CNN with the same architecture.

[0099] Hyperparameters are crucial for the prediction performance of the model. In this study, the Adam optimizer was selected. Among a series of hyperparameters, taking the learning rate lr as an example, the iterative curves of the model loss values at different learning rates are shown. As Figure 4Iteration curves of the loss function of the prediction model of the present invention when the learning rates shown are lr = 0.01, 0.001, and 0.0001. When lr = 0.01, in the initial stage of training the prediction model of the present invention, the reduction rate of the training loss is more significant than when lr = 0.001 and lr = 0.0001. However, when the training reaches the 5th epoch, due to the too high learning rate setting, the training loss fluctuates greatly. Further observing the subsequent iteration process, it is found that when lr = 0.001, the training loss can continuously decrease smoothly, and throughout the training process, compared with lr = 0.01 and lr = 0.0001, a better training effect is achieved. On the other hand, when lr = 0.0001, although the training loss always shows a downward trend throughout the iteration process and the fluctuation of the iteration curve is the smallest, the low learning rate leads to a reduction in training efficiency. Until the iteration ends, the training loss of the model at this learning rate is still higher than that when lr = 0.001, meaning that in order to obtain a lower training loss, the prediction model of the present invention needs to go through more iteration times and a longer training time. At the 150th epoch of iteration, when lr = 0.01, the training loss of the prediction model of the present invention is 0.059991; when lr = 0.001, the training loss of the model is 0.058161; when lr = 0.0001, the training loss of the prediction model of the present invention is 0.058257. Therefore, lr = 0.001 is selected.

[0100] To verify the prediction performance of the modulation deformable convolutional neural network DFC-CNN of the present invention and traditional models, DFC-CNN was compared with BP and LSTM on the validation set. Figure 5 The scatter plot of the predicted values and the true values is shown. It can be seen from the distribution fitting curve of the predicted values and the measured values that Figure 5 in (a), the fitting slope of the BP scatter points is 0.91886, with the largest deviation from the perfect line, and the determination coefficient R 2 is 0.91222; Figure 5 in (b), the fitting slope of the LSTM scatter points is 0.95459, and the determination coefficient R 2 is 0.923; Figure 5 in (c), the fitting slope of the DFC-CNN scatter points is 0.97583, and the determination coefficient R 2 is 0.973, which is the closest to the perfect line, indicating that the overall predicted values of DFC-CNN are closer to the measured values. In addition, among the prediction results of BP and LSTM, the number of samples scattered outside the 95% confidence interval exceeds 40 sample points. Taking the sample point with a measured value of about 420 mg / m 3 as an example, the predicted values generally exceed 480 mg / m 3, indicating that the prediction accuracy of these two models for low-emission conditions is relatively low. In the prediction results of DFC-CNN, the vast majority of sample points are concentrated within the 95% confidence interval, the number of samples scattered outside the 95% confidence interval is the least, and most of them are concentrated on both sides of the fitting line. This benefits from the weight adjustment of channels during the training process of the model. The prediction performances of the three models, DFC-CNN, BP, and LSTM, are as Figure 6 shown. The RMSE of DFC-CNN is 20.006 mg / m 3 , the MAE is 14.509 mg / m 3 , and the MAPE is 3.04%; the RMSE of BP is 29.079, the MAE is 22.361, and the MAPE is 4.66%; the RMSE of LSTM is 24.24376723, the MAE is 17.2585, and the MAPE is 3.64%. Compared with BP and LSTM, the prediction accuracy of DFC-CNN is improved by 31.20%, 17.48% and 35.11%, 15.93% respectively in terms of root mean square error (RMSE) and mean absolute error (MAE). The experimental results show that the DFC-CNN model is significantly superior to the traditional models in the prediction accuracy of NOx emissions.

[0101] Figure 7 In (a) of Figure 7 is the comparison curve of load increase between Baseline-CNN and DFC-CNN. 3 In (b) of 3 is the scatter plot of load increase error comparison between Baseline-CNN and DFC-CNN, showing that when the boiler load is in the rising stage, the DFC-CNN model exhibits good prediction performance. The change trend of the predicted value is the same as that of the true value, and the error shows a normal distribution with the mean close to 0. The error of 96 sampling points is within ±20 mg / m 3 , and the maximum error is -27.02 mg / m 3 . Although the change trend of the predicted value of the baseline CNN model (Baseline) is the same as that of the true value, the mean distribution of the error deviates from 0. This is because as the boiler load increases, the coal combustion amount increases, the combustion temperature will rise, and the NOx emissions change with the temperature.

[0102] Figure 8 In (a) of Figure 8Among them, (b) is the scatter plot of the load reduction error comparison between Baseline-CNN and DFC-CNN, showing that when the boiler load is in the decreasing stage, the prediction error of the DFC-CNN model for NOx emissions is evenly distributed on both sides of the baseline. This is because at low load conditions, the efficiency of coal combustion decreases, resulting in uneven furnace temperature distribution and reduced combustion stability.

[0103] Figure 9 Among them, (a) is the variable load comparison curve between Baseline-CNN and DFC-CNN, Figure 9 Among them, (b) is the scatter plot of the variable load error comparison between Baseline-CNN and DFC-CNN, showing that the boiler load changes frequently and the NOx emissions fluctuate greatly. This is because when the unit is in a load fluctuation state, the balance of the air-coal ratio is disrupted, which will have a significant impact on the NOx concentration. The performance of the baseline CNN model deteriorates severely under such working conditions, and the predicted NOx emissions are generally lower than the actual values. The absolute error of 45% of the sampling points is far greater than 20 mg / m 3 ; The error of the NOx emissions predicted by the DFC-CNN model is evenly distributed on both sides of 0 mg / m 3 and only 10% of the sampling points have an absolute error greater than 20 mg / m 3 .

[0104] Figure 10 Among them, (a) is the stable load comparison curve between Baseline-CNN and DFC-CNN, Figure 10 Among them, (b) is the scatter plot of the stable load error comparison between Baseline-CNN and DFC-CNN, showing that when the boiler is in a stable working condition, the load change tends to be stable and the combustion environment in the furnace is relatively stable. Both the DFC-CNN model and the baseline CNN model can better predict the NOx emissions, and the error value is smaller than the above three working conditions, but the prediction accuracy of the DFC-CNN model is higher.

[0105] In summary, in comparison with the traditional BP neural network and LSTM neural network models, the modulated deformable convolutional neural network DFC-CNN exhibits higher prediction accuracy. When the three models simultaneously process the test set data, the fitting slope between the predicted values and the true values of the modulated deformable convolutional neural network DFC-CNN is closer to the perfect line, significantly outperforming the BP and LSTM models. In addition, among the prediction results of the modulated deformable convolutional neural network DFC-CNN, the number of samples scattered outside the 95% confidence interval is significantly less than that of the BP and LSTM. Compared with the Baseline-CNN model, the modulated deformable convolutional neural network DFC-CNN maintains good prediction performance under complex and variable working conditions. Under the condition of load reduction, there are 156 working condition points where the error values of the modulated deformable convolutional neural network DFC-CNN are distributed within ±20 mg / m 3 while there are 38 working condition points where the error values of the Baseline-CNN are distributed within ±20 mg / m 3 ; under the condition of variable load, the error of all working condition points of the modulated deformable convolutional neural network DFC-CNN is within ±20 mg / m 3 while there are 123 working condition points where the error values of the Baseline-CNN are distributed within ±20 mg / m 3 .

[0106] In actual operation, the operation economy or nitrogen oxide emissions are not simply pursued. Usually, the economy and pollutant generation are comprehensively optimized. The present invention is a NOx concentration prediction model for the denitration process of coal-fired boilers based on the modulated deformable convolutional neural network DFC-CNN, which has the advantages of high accuracy, stability, and wide applicability. The successful establishment and application of this model provide strong technical support for optimizing the operation of coal-fired boilers and reducing NOx emissions, and are of great significance for promoting the environmental protection transformation and sustainable development of coal-fired boilers.

[0107] The above are only the embodiments of the present invention, and do not limit the patent scope of the present invention accordingly. Any equivalent structure or equivalent process transformation made by using the content of the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, shall be similarly included in the patent protection scope of the present invention.

Claims

1. A method for predicting NOx emissions from coal-fired boilers based on modulated deformable neural networks, characterized in that: The steps include: Step 1, screening key input variables related to NOx emissions; Step 2, construct a modulated deformable convolutional neural network DFC-CNN model to predict NOx emissions; In the modulated deformable convolutional neural network DFC-CNN model, the input feature map X is offset and modulated to obtain a modulated output feature map , and then the feature map The eigenvalues ​​in After being processed in sequence by multiple convolutional layers, the eigenvalue Xfcl output by the convolutional layer is obtained. After being processed by the fully connected layer, the predicted value of NOx is obtained, that is, the output value of the modulated deformable convolutional neural network DFC-CNN model; Output feature map after modulation The eigenvalue of each position (i, j) in , expressed as: （10） Among them, u, v are the indices of the convolution kernel, is the weight of the convolution kernel, is the modulation factor, is the offset; X fcl Is a four-dimensional tensor. In the fully connected layer, first X fcl Flatten, the flattened vector is shown in formula (11), and then weighted summation is shown in formula (12): (11) (12) Among them, vec(.) is the operation of converting a matrix or tensor into a vector. represents the one-dimensional tensor obtained by flattening the four-dimensional tensor, represents the weight matrix of the fully connected layer, represents the bias term of the fully connected layer, Represents the result after linear change, and finally through the activation function Find the output target , which is the predicted value of NOx emissions from coal-fired boilers , expressed as: (13) Step 3, constructing a data set for training a modulated deformable convolutional neural network DFC-CNN model to obtain a trained DFC-CNN model; The data set includes key input variables and corresponding NOx emissions; Step 4: Input the key input variables to be predicted into the trained DFC-CNN model to predict NOx emissions.

2. The method for predicting NOx emissions from coal-fired boilers based on modulated deformable neural networks according to claim 1, characterized in that: In step 1, key input variables related to NOx emissions are screened, specifically: the Gini index based on random forest RF is used to evaluate the characteristic importance of variables for boiler NOx emissions.

3. The method for predicting NOx emissions from coal-fired boilers based on modulated deformable neural networks according to claim 2 is characterized in that: The number of decision trees is n, and the jth var The feature in the i tree Importance in decision trees for: （3） in, Indicates the jth var The importance contribution of a feature at index m is calculated by the Gini index, where M represents the set of indices; No. var The importance of features in all decision trees for: （4） The j var The importance of features in all decision trees Perform normalization to obtain the normalized jth var The importance of features in all decision trees , expressed as: （5） ； Wherein, J represents the total number of all characteristic variables related to NOx; If the normalized jth var The importance of features in all decision trees is greater than or equal to the set value, then the jth var The features are key input variables.

4. The method for predicting NOx emissions from coal-fired boilers based on modulated deformable neural networks according to claim 1, characterized in that: The key input variables include feed water flow, SOFA-D#1 angle, main steam flow, unit load, FR#1 angle, E mill E1 side capacity wind flow, EL#1 angle, DE#1 angle, main steam temperature, DL#1 angle, flue gas oxygen content A, exhaust temperature A, CL#1 angle, SOFA-E#1 angle, DEE#1 angle, B mill B1 side capacity wind flow, A mill A2 side capacity wind flow, B blower inlet air temperature, SOFA-A#1 angle, SOFA-F#1 angle, total coal quantity, EFF#1 angle, SOFA-B#1 angle, AAA layer #1 angle, AA#1 angle, B mill B2 side capacity wind flow, ER#1 angle, a total of 27 variables are used as input variables for modulating the deformable convolutional neural network DFC-CNN model.

5. The method for predicting NOx emissions from coal-fired boilers based on modulated deformable neural networks according to claim 1, characterized in that: When training the modulated deformable convolutional neural network DFC-CNN model, the key input variables and NOx emissions are reconstructed into a two-dimensional time series tensor, and the two-dimensional time series tensor is segmented by the time dimension using a sliding window, and the sliding step is set to 1.

6. The method for predicting NOx emissions from coal-fired boilers based on modulated deformable neural networks according to claim 1, characterized in that: The loss function of the DFC-CNN model is the mean square error MSE, and the activation function is the Sigmoid function.

7. The method for predicting NOx emissions from coal-fired boilers based on modulated deformable neural networks according to claim 1, characterized in that: The data set constructed in step 3 is obtained by extracting historical operation data from the coal-fired boiler monitoring system and performing preprocessing and normalization.

8. The method for predicting NOx emissions from coal-fired boilers based on modulated deformable neural networks according to claim 7, characterized in that: The preprocessing refers to processing abnormal data in historical operation data using the 3σ rule.

Images